<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0" xmlns:cc="http://cyber.law.harvard.edu/rss/creativeCommonsRssModule.html">
    <channel>
        <title><![CDATA[Stories by Luigi Pascal Rondanini on Medium]]></title>
        <description><![CDATA[Stories by Luigi Pascal Rondanini on Medium]]></description>
        <link>https://medium.com/@info_38412?source=rss-1a66c626d0c4------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*UJQ9_xDWdiSY2PNijldGag.png</url>
            <title>Stories by Luigi Pascal Rondanini on Medium</title>
            <link>https://medium.com/@info_38412?source=rss-1a66c626d0c4------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Sat, 21 Nov 2026 23:48:36 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@info_38412/feed" rel="self" type="application/rss+xml"/>
        <webMaster><![CDATA[yourfriends@medium.com]]></webMaster>
        <atom:link href="http://medium.superfeedr.com/" rel="hub"/>
        <item>
            <title><![CDATA[When Authors Stop Believing Publishers]]></title>
            <link>https://medium.com/rondanini-publishing/why-authors-no-longer-trust-publishers-afa5154cfbf0?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/afa5154cfbf0</guid>
            <category><![CDATA[publishing]]></category>
            <category><![CDATA[book-publishing]]></category>
            <category><![CDATA[independent-publishing]]></category>
            <category><![CDATA[authors]]></category>
            <category><![CDATA[writing]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Sat, 21 Nov 2026 10:26:36 GMT</pubDate>
            <atom:updated>2026-11-21T10:26:36.931Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*eYpQx0IDTBz7VVrE" /></figure><p><em>Why we founded Rondanini Publishing</em></p><p>A recent exchange with a prospective author left me thinking about something rather troubling.</p><p>The author had submitted an interesting nonfiction proposal. We reviewed the initial material, offered constructive editorial feedback and explained how our publisher-funded programme works.</p><p>Everything was going well. The author was interested in proceeding.</p><p>But then came the questions.</p><p>Would the manuscript assessment be free? Would the editorial report cost anything? Was there an obligation to buy copies? Would the submission remain non-exclusive?</p><p>Perfectly reasonable questions. Except that we had already answered them. More than once. And the same information is clearly stated on our website.</p><p>At first, I wondered whether we had failed to communicate properly. Then I realised that the problem might be much bigger than our communication.</p><p><strong>Authors have become so accustomed to hidden costs, unexpected conditions and disappointing experiences that straightforward promises are no longer enough.</strong></p><p>And that says a great deal about the state of the publishing industry.</p><p>Rondanini Publishing offers editorial services as well as traditional publishing. Authors can approach us independently for professional editorial assistance, and those services have a price.</p><p>But there is a distinction I consider fundamental.</p><p><strong>We will never sell our editorial or production services to authors whose manuscripts we accept for publication.</strong></p><p>If we decide to publish your book, the investment is ours. From manuscript assessment and editing to proofreading, typesetting, cover design and production, we take responsibility for the work and its costs.</p><p>We do not ask authors to finance their own acceptance. We do not require them to purchase copies. We do not turn an offer of publication into an invoice.</p><p>Of course, we cannot publish every manuscript we receive. And acceptance may involve considerable editorial work, sometimes including substantial revisions. Publishing should be a partnership, with responsibilities on both sides.</p><p>But the financial risk of publishing a book we have chosen to acquire belongs to the publisher.</p><p>What struck me about this recent conversation was not the author’s caution. I respect it. What struck me was how difficult it has apparently become for authors to believe that a publisher might simply mean what it says.</p><p>I have discussed these problems in our Unseen Author Journeys podcast series. Every conversation reinforces my conviction that the relationship between authors and publishers needs rebuilding.</p><p>Not through grand promises or clever marketing, but through ordinary things: transparency, fair agreements, honest criticism and keeping our word.</p><p>Perhaps the saddest part is that none of this should be remarkable.</p><p><strong>It should be normal.</strong></p><p>And the fact that it no longer feels normal is precisely why Rondanini Publishing exists.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/when-authors-stop-believing-publishers-rondanini-publishing-ltd-8nffe/"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=afa5154cfbf0" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/why-authors-no-longer-trust-publishers-afa5154cfbf0">When Authors Stop Believing Publishers</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Invisible Language of Books]]></title>
            <link>https://medium.com/rondanini-publishing/typography-across-countries-languages-books-3b25ee8e0492?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b25ee8e0492</guid>
            <category><![CDATA[language]]></category>
            <category><![CDATA[typography]]></category>
            <category><![CDATA[book-design]]></category>
            <category><![CDATA[literature]]></category>
            <category><![CDATA[book-publishing]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Thu, 19 Nov 2026 22:51:16 GMT</pubDate>
            <atom:updated>2026-11-19T22:51:16.800Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Editorial cover for The Invisible Language of Books, featuring a genuine photograph of books arranged in a circular pattern. An ivory panel displays the article title alongside examples of English, Italian, French and German quotation marks, illustrating how typography differs across languages and publishing traditions." src="https://cdn-images-1.medium.com/max/1024/1*CzucmrcsIeAyzUamggHu5w.png" /><figcaption>The same words, different traditions. From quotation marks to page layouts, every publishing culture has its own visual language. Photography by Nataly Q. / Pexels. Design by Rondanini Publishing.</figcaption></figure><h3>Why the same words look different depending on where a book is published</h3><p><em>By Luigi Pascal Rondanini</em></p><p>There is a peculiar thing about books that most readers never notice.</p><p>You can take exactly the same manuscript, give it to five publishers in five different countries, and receive five noticeably different books in return.</p><p>Not different translations. Not different stories. The same words, sometimes even in the same language.</p><p>Yet the pages will look different. The dialogue will be punctuated differently. Paragraphs will begin differently. The margins, chapter openings, quotation marks, page numbers and even the amount of white space will reflect the country in which the book was produced.</p><p>And here’s the interesting part: none of those publishers necessarily did anything wrong.</p><p>They simply followed different traditions.</p><p>As a publisher working with books in several languages, I have become increasingly fascinated by the small decisions that turn a manuscript into a physical book. Some are technical. Others are historical. Quite a few are cultural.</p><p>Most remain invisible until you cross a border.</p><h3>A Book Has More Than One Language</h3><p>When we talk about the language of a book, we generally mean English, Italian, French, Spanish or whatever language the author happens to write in.</p><p>But books have another language: typography.</p><p>Typography tells the reader where to pause, where a conversation begins, how important a heading is, whether a passage belongs to the narrator or a character, and how the eye should travel across a page.</p><p>It also establishes a relationship between the reader and the text.</p><p>A densely printed academic book makes a different promise from a generously spaced literary novel. A book with elaborate chapter openings creates a different atmosphere from one in which a chapter begins with a simple number.</p><p>We tend to think of these things as design choices.</p><p>They are, but they are also inherited conventions, developed over centuries of printing, publishing and reading.</p><p>And those conventions do not always travel well.</p><h3>The First Surprise: Britain and America Do Not Print English the Same Way</h3><p>Let’s begin with two countries separated by a common language.</p><p>Open a British novel and you may encounter dialogue presented like this:</p><p>‘Are you coming tomorrow?’ she asked.</p><p>‘I haven’t decided,’ he replied.</p><p>Open an American edition of the same novel and you are more likely to see:</p><p>“Are you coming tomorrow?” she asked.</p><p>“I haven’t decided,” he replied.</p><p>The difference is immediately visible.</p><p>Traditional British publishing frequently favours single quotation marks for dialogue, reserving double quotation marks for quotations within quotations.</p><p>American publishing generally reverses that arrangement.</p><p>Of course, there are exceptions. British publishers do not all follow the same rules, and neither do American ones. House style ultimately decides.</p><p>But the distinction is sufficiently established that a book prepared for one market may be restyled for the other.</p><p>And it doesn’t end there.</p><p>Consider a sentence such as:</p><p>She described the situation as ‘extraordinary’.</p><p>In much British editorial practice, the full stop remains outside the quotation marks because it does not belong to the quoted word.</p><p>American practice generally places commas and periods inside closing quotation marks:</p><p>She described the situation as “extraordinary.”</p><p>Same meaning. Different typographic logic.</p><p>Even spelling becomes part of the visual landscape: <em>colour</em> and <em>color</em>, <em>centre</em> and <em>center</em>, <em>organise</em> and <em>organize</em>.</p><p>A book may be perfectly understandable in both countries, yet its appearance quietly announces which publishing tradition it follows.</p><p>This is one reason why producing an American edition can involve more than changing a few spellings.</p><p>A publisher may need to reconsider punctuation, quotation marks, abbreviations, hyphenation and the conventions used throughout the manuscript.</p><p>The reader may never consciously identify these differences.</p><p>But an unfamiliar convention can make a page feel subtly foreign.</p><h3>Italy: Where Dialogue Has Several Voices</h3><p>Italian publishing introduces another interesting problem.</p><p>Consider this passage:</p><p>«Non ti aspettavo», disse Maria.</p><p>«Nemmeno io pensavo di venire.»</p><p>The quotation marks are called <em>caporali</em>, sometimes also known as <em>virgolette basse</em>.</p><p>They are widely used in Italian books, particularly for dialogue and quotations.</p><p>But an Italian novel might equally present the same exchange using double quotation marks:</p><p>“Non ti aspettavo”, disse Maria.</p><p>“Nemmeno io pensavo di venire.”</p><p>Or dialogue may begin with a dash:</p><p>— Non ti aspettavo, disse Maria.</p><p>— Nemmeno io pensavo di venire.</p><p>All three systems can be found in Italian publishing, although their punctuation details and frequency differ between publishers.</p><p>And this is where the matter becomes particularly interesting.</p><p>An author might submit a manuscript with one system, only to discover that the publisher prefers another.</p><p>Neither is necessarily incorrect.</p><p>The important thing is consistency and the ability of the reader to distinguish dialogue from narration.</p><p>Italian publishing houses often have detailed editorial rules governing quotation marks, italics, apostrophes, accents, abbreviations and punctuation.</p><p>Even the distinction between <em>È</em> and <em>E’</em> matters. They may be treated as interchangeable by someone typing quickly, but they are not typographically equivalent.</p><p>Then there is the question of chapter openings.</p><p>Some Italian literary editions favour relatively restrained layouts, while others use generous blank space, large chapter numbers or decorative initial letters.</p><p>These are editorial choices, not national laws.</p><p>Nevertheless, after handling enough books, one begins to recognise the visual habits of different publishing cultures.</p><p>A translation can preserve every word of an original and still look entirely different from it.</p><h3>France: The Country That Gives Punctuation Room to Breathe</h3><p>French typography has a particularly elegant peculiarity.</p><p>It uses space as part of punctuation.</p><p>Consider this English sentence:</p><p>What are you doing? I told you: wait!</p><p>In French, the equivalent is traditionally written:</p><p>Qu’est-ce que tu fais ? Je t’ai dit : attends !</p><p>Look carefully.</p><p>There is space before the question mark, the colon and the exclamation mark.</p><p>But these are not simply ordinary spaces inserted at random.</p><p>French typography uses different kinds of spaces, including non-breaking spaces and narrow non-breaking spaces. They prevent punctuation from becoming stranded at the beginning of a line.</p><p>The exact width and treatment can vary between typographic conventions and publishing systems.</p><p>French quotation marks also have their own personality:</p><p>« Je reviendrai demain. »</p><p>The words sit between <em>guillemets</em>, with space separating the quotation marks from the text.</p><p>Compare that with Italian:</p><p>«Tornerò domani.»</p><p>Both languages use the same basic quotation-mark symbols.</p><p>But traditionally, Italian sets them directly against the words, while French inserts space inside them.</p><p>It is a tiny distinction.</p><p>Yet multiply it by several thousand lines and the entire texture of the page changes.</p><p>French typography reminds us that empty space is not necessarily empty.</p><p>Sometimes it is an essential character in the language of the book.</p><h3>Spain and Latin America: A Conversation Begins With a Dash</h3><p>Spanish-language fiction offers another fascinating tradition.</p><p>In many Spanish novels, dialogue is introduced with a long dash, called a <em>raya</em>.</p><p>For example:</p><p>— No esperaba verte aquí — dijo Ana.</p><p>— Yo tampoco pensaba venir.</p><p>There are no opening and closing quotation marks around each spoken sentence.</p><p>The dash announces the beginning of speech. Additional dashes can separate dialogue from the narrator’s intervention.</p><p>For an English reader accustomed to quotation marks, the page can initially look unusual.</p><p>But after a few paragraphs, the system becomes entirely natural.</p><p>Spanish also uses angular quotation marks:</p><p>«No esperaba verte aquí».</p><p>These are generally recommended for primary quotations in formal Spanish typography, with double and single quotation marks available for nested quotations.</p><p>However, there is an important distinction between the rules recommended by language authorities and the practices adopted by publishers.</p><p>Publishing houses in Spain, Mexico, Argentina and elsewhere in Latin America do not necessarily follow identical house styles.</p><p>The same Spanish-language novel may be typeset differently depending on the country, publisher and intended readership.</p><p>Regional vocabulary and spelling preferences may be visible, but typographic decisions can differ even when the words themselves remain untouched.</p><p>And this is perhaps the most important point: a shared language does not guarantee a shared publishing culture.</p><h3>Portugal and Brazil: One Language, Different Books</h3><p>Portuguese provides another excellent example.</p><p>Portugal and Brazil share a language, but their literary and editorial traditions have developed along different paths.</p><p>Both use dashes and quotation marks for dialogue, yet the preferences of individual publishers, and the ways punctuation is combined with narrative interruptions, can differ.</p><p>In Brazilian fiction, a familiar pattern is:</p><p>— Não sabia que você viria — disse ela.</p><p>In Portugal, one can also encounter dialogue introduced with a dash, as well as books using angular quotation marks.</p><p>There is no simple rule that every Brazilian book uses one system and every Portuguese book another.</p><p>Indeed, the existence of several accepted conventions is precisely what makes the subject interesting.</p><p>One particularly revealing case is José Saramago.</p><p>His novels famously depart from conventional dialogue punctuation. Speech often flows into narration, with changes in speaker indicated through commas and capital letters rather than the quotation marks or paragraph breaks many readers expect.</p><p>This is no accident, and certainly not evidence of poor editing.</p><p>It is part of the author’s literary method.</p><p>His typography influences the rhythm of reading.</p><p>It asks the reader to participate differently.</p><p>And it demonstrates something publishers sometimes forget: typographic conventions are tools, not commandments.</p><h3>Germany: Even Quotation Marks Have a Different Orientation</h3><p>German books introduce another familiar but distinctive system.</p><p>A conversation may look like this:</p><p>„Ich komme morgen“, sagte sie.</p><p>Notice the opening quotation mark.</p><p>It sits low, at the bottom of the line.</p><p>The closing quotation mark sits high.</p><p>German typography commonly uses this arrangement:</p><p>„…“</p><p>It is also possible to encounter:</p><p>»…«</p><p>These are sometimes called reversed guillemets because their direction differs from the familiar French arrangement:</p><p>«…»</p><p>German-speaking publishing markets — including Germany, Austria and Switzerland — have overlapping but not identical typographic habits.</p><p>Swiss German-language publishing, for example, often favours guillemets, while the low-high quotation-mark pattern is particularly familiar in Germany.</p><p>Even within the same language family, the preferred appearance of a printed page can change at a national border.</p><p>A character speaks the same German sentence, but the punctuation tells a different cultural story.</p><h3>Then We Reach Japan, and the Page Changes Direction</h3><p>So far, we have been discussing differences within a broadly European tradition of book production.</p><p>Japanese publishing reminds us how limited that perspective can be.</p><p>Traditional Japanese books may be set vertically, with characters running from top to bottom and columns progressing from right to left.</p><p>Modern Japanese books can also use horizontal text.</p><p>Neither arrangement is simply a rotated version of the other.</p><p>Vertical Japanese typography has its own rules for punctuation, numbers, annotations and line breaks.</p><p>Japanese dialogue commonly uses corner brackets:</p><p>「明日、来ますか。」</p><p>There are also rules governing which characters may appear at the beginning or end of a line.</p><p>These conventions are known as <em>kinsoku shori</em>.</p><p>They prevent certain punctuation marks and other characters from being placed in visually or linguistically inappropriate positions.</p><p>Japanese books may also use <em>ruby</em> annotations: small phonetic characters positioned beside or above other characters to indicate pronunciation.</p><p>A book designer working with Japanese therefore faces questions that do not arise in exactly the same form when preparing an English novel.</p><p>The direction of reading, the placement of annotations and the handling of line breaks all become part of the design.</p><p>This is not merely a matter of choosing a different font.</p><p>It is a different system of arranging language on a page.</p><h3>Arabic: When Justification Becomes an Art</h3><p>Arabic offers another example of how typography must respect the structure of a writing system.</p><p>Arabic text is normally read from right to left.</p><p>But the differences go much deeper than direction.</p><p>Arabic letters connect to one another, and their forms can change depending on their position within a word.</p><p>In European typography, justified text is commonly produced by adjusting the spaces between words, sometimes with additional help from letter spacing and hyphenation.</p><p>Arabic typography has other possibilities.</p><p>One is <em>kashida</em>, the extension of connections between letters to help balance a line.</p><p>This must be handled carefully.</p><p>Stretching every available connection mechanically can produce awkward and unattractive results.</p><p>Good Arabic composition requires an understanding of how letterforms, spacing and visual rhythm interact.</p><p>The difference matters especially when text is converted between print and digital formats.</p><p>A layout engine designed primarily around Latin-script assumptions may not produce satisfactory Arabic typography without proper script support.</p><p>Technology has made multilingual publishing easier.</p><p>It has not eliminated the need to understand the languages being published.</p><h3>But Typography Is More Than Quotation Marks</h3><p>Dialogue is probably the easiest difference to notice.</p><p>Yet it represents only a small part of book design.</p><p>Consider the essential elements of a printed book.</p><h3>The title page</h3><p>A title page normally identifies the work, its author and often the publisher.</p><p>But the balance between these elements can vary enormously.</p><p>Some publishing traditions favour elaborate title pages. Others are remarkably restrained.</p><p>The placement of a publisher’s name or imprint, the use of capitals and the relationship between title and subtitle are all design decisions influenced by genre and tradition.</p><h3>The copyright page</h3><p>This is where bibliographic and legal information generally appears.</p><p>ISBNs, edition statements, copyright notices, printing details, publisher addresses and other information may be included.</p><p>But the content and presentation depend on the jurisdiction, publisher and type of publication.</p><p>British books commonly include a British Library catalogue record statement when appropriate. American editions may include Library of Congress cataloguing information.</p><p>These are not simply decorative differences.</p><p>They reflect distinct publishing infrastructures and bibliographic practices.</p><h3>The table of contents</h3><p>In some books, the contents page appears before the main text.</p><p>In others, particularly certain literary editions, it may appear at the back or be omitted altogether.</p><p>Non-fiction books may require a detailed hierarchy of parts, chapters and sections.</p><p>Novels may need nothing beyond chapter numbers.</p><p>The design should serve the reader’s expectations and the book’s purpose.</p><h3>Chapter openings</h3><p>A chapter might begin halfway down a page, with an oversized numeral.</p><p>It might start at the top, beneath a modest heading.</p><p>It might open with a decorative initial letter, known as a <em>drop cap</em>.</p><p>Or the first few words may appear in small capitals.</p><p>These choices affect how a reader experiences the transition from one chapter to another.</p><p>A large amount of white space creates a pause.</p><p>A modest heading maintains momentum.</p><p>Neither is universally better.</p><h3>Paragraphs</h3><p>This is one of my favourite details because it is so often overlooked.</p><p>In many traditionally typeset novels, the first paragraph of a chapter is not indented.</p><p>Subsequent paragraphs receive a first-line indent.</p><p>Why?</p><p>Because the chapter heading already tells the reader that a new section has begun.</p><p>There is no need for another visual signal.</p><p>After that, indentation helps identify each new paragraph.</p><p>Other publications use additional vertical space between paragraphs, sometimes without indentation.</p><p>Both systems have legitimate applications.</p><p>The trouble begins when they are combined without intention.</p><p>Large paragraph gaps together with first-line indents can make continuous prose look fragmented.</p><p>A layout that works beautifully on a website may be less successful in a 400-page novel.</p><h3>Running heads and page numbers</h3><p>Open a printed book and examine the top or bottom of the pages.</p><p>You may find the author’s name on one side and the book title on the other.</p><p>Or perhaps the chapter title appears in the running head.</p><p>Sometimes page numbers sit at the bottom centre. Sometimes they appear near the outer edges.</p><p>Some books suppress running heads and visible folios on chapter-opening pages.</p><p>These details may seem insignificant.</p><p>But together they give the book a recognisable visual structure.</p><p>They also help the reader navigate it.</p><h3>Footnotes, endnotes and references</h3><p>Academic books expose another layer of variation.</p><p>Footnote symbols, numbering systems, citation formats, bibliographies and reference abbreviations can differ between scholarly disciplines and national publishing traditions.</p><p>A history book published in Britain may follow one referencing convention.</p><p>An American university press may prefer another.</p><p>A specialist French or German academic publisher may have additional expectations.</p><p>Again, language is only one factor.</p><p>The intended readership, discipline and publisher’s established style all influence the result.</p><h3>The Paper and the Page Are Part of the Story</h3><p>We should also talk about something that has nothing to do with punctuation.</p><p>The physical dimensions of the book.</p><p>The familiar American 6 × 9-inch paperback does not correspond exactly to the international A5 format.</p><p>And neither is automatically the right choice for every book.</p><p>Different markets have developed familiar trim sizes for mass-market fiction, literary paperbacks, trade paperbacks, academic works and illustrated books.</p><p>Some differences reflect industrial history.</p><p>Others relate to paper sizes, printing equipment, distribution economics and consumer expectations.</p><p>Even within Europe, publishers do not use a single universal paperback format.</p><p>Then there is the paper itself.</p><p>Cream or white?</p><p>Smooth or slightly textured?</p><p>Thin and opaque, or thicker and more substantial?</p><p>And what about the binding?</p><p>A book’s physical construction influences how it opens, how it rests in the hand and how comfortably its pages can be read.</p><p>In the UK, a literary paperback printed on cream paper may feel entirely conventional.</p><p>Elsewhere, a different combination of paper, format and binding may be more familiar.</p><p>I am careful not to turn these observations into rigid national stereotypes. Printing methods and commercial requirements often explain more than geography.</p><p>Nevertheless, physical books carry the fingerprints of the markets for which they were produced.</p><h3>There Is Also the Question of Hyphenation</h3><p>One of the least glamorous aspects of typography is also one of the most important.</p><p>Hyphenation.</p><p>When a word does not fit at the end of a line, should it move entirely to the next line?</p><p>Or should it be divided?</p><p>And if divided, where?</p><p>The answer depends on the language.</p><p>English, Italian, German and Spanish have different rules for permissible word breaks.</p><p>German compound words create particular challenges because they can be very long.</p><p>Italian hyphenation follows syllabic principles that are not identical to English ones.</p><p>And a word processor configured for the wrong language may produce line breaks that look acceptable at first glance but are linguistically incorrect.</p><p>This matters enormously in justified text.</p><p>Poor hyphenation can create excessive spaces between words.</p><p>Those spaces may form distracting vertical patterns, often called <em>rivers</em>, across the page.</p><p>A reader may not know the technical term.</p><p>But the eye notices.</p><p>Good typography is partly the art of preventing the reader from noticing anything at all.</p><h3>The Digital Book Has Complicated Everything</h3><p>For centuries, printers exercised considerable control over how books appeared.</p><p>Digital publishing changed that relationship.</p><p>An EPUB can adapt to different screen sizes, fonts and reader preferences.</p><p>This is one of its greatest strengths.</p><p>But it also means that the publisher no longer controls every detail of the page.</p><p>A reader may enlarge the font.</p><p>Change the typeface.</p><p>Switch to dark mode.</p><p>Adjust line spacing.</p><p>Or read the same book on a phone, tablet and dedicated e-reader.</p><p>For accessibility, this flexibility is essential.</p><p>Yet it presents a challenge when the work depends on sophisticated typography.</p><p>Poetry, mathematical notation, multilingual dialogue, complex tables and certain experimental literary forms can be particularly difficult to reproduce consistently.</p><p>Even ordinary punctuation can become problematic when software handles language-specific conventions incorrectly.</p><p>An EPUB containing French text must still respect French punctuation.</p><p>A Japanese EPUB must accommodate Japanese layout requirements.</p><p>An Arabic EPUB must correctly handle its writing direction and shaping.</p><p>The file format may be international.</p><p>The language inside it remains culturally and technically specific.</p><h3>The Mistake of Treating Every Book as an English Book</h3><p>One problem I increasingly notice in international publishing workflows is the assumption that every language can be poured into the same template.</p><p>Choose a font.</p><p>Set the margins.</p><p>Import the manuscript.</p><p>Export a PDF.</p><p>Done.</p><p>Except it isn’t.</p><p>A template created for an English-language novel may work perfectly well for another English novel.</p><p>But it may require substantial adaptation for Italian dialogue, French punctuation, German hyphenation or Japanese text.</p><p>Even two editions in English may deserve different treatment depending on their intended markets.</p><p>The same applies to automated tools.</p><p>Modern software can produce technically valid book files in minutes.</p><p>It can also reproduce typographic mistakes at extraordinary speed.</p><p>An automated workflow may correct spelling while damaging dialogue punctuation.</p><p>It may replace carefully chosen quotation marks.</p><p>It may insert inappropriate spaces, remove necessary ones or mishandle accented characters.</p><p>And because the output looks polished, those mistakes can be surprisingly difficult to spot.</p><p>This is where human editorial judgement remains valuable.</p><p>The challenge is not simply to produce a file that passes a technical validation test.</p><p>It is to produce a book that looks and reads naturally to its intended audience.</p><h3>What Should a Publisher Actually Do?</h3><p>I don’t believe every publisher needs to reproduce every national tradition with absolute rigidity.</p><p>Readers are flexible, and many contemporary books deliberately depart from convention.</p><p>But a publisher should understand the conventions before deciding to ignore them.</p><p>For me, a sensible multilingual publishing process begins with a few questions.</p><p>Who is the book for?</p><p>Where will it primarily be sold?</p><p>Which linguistic and typographic conventions will its readers expect?</p><p>Does the original author use typography as part of the literary voice?</p><p>And, perhaps most importantly, are we changing something because it genuinely improves the book, or simply because our software has a default setting?</p><p>These questions become particularly relevant when preparing translations.</p><p>A translator has the difficult task of preserving meaning, tone, rhythm and cultural context.</p><p>The typesetter has a related responsibility.</p><p>The translated work should feel like a properly published book in its new language, without unnecessarily erasing the author’s intentions.</p><p>A translation that reads beautifully but has inappropriate dialogue punctuation or awkward line breaks is not quite finished.</p><p>Nor is a perfectly typeset book in which an author’s deliberate stylistic choices have been normalised out of existence.</p><p>There must be a conversation between the text and its presentation.</p><h3>The Beautiful Contradiction of Book Design</h3><p>Perhaps the most intriguing aspect of all this is that typography is both deeply conservative and endlessly adaptable.</p><p>Publishing traditions are inherited.</p><p>They carry centuries of decisions made by printers, editors, scholars and readers.</p><p>Yet they are also constantly changing.</p><p>Digital publishing introduces new habits.</p><p>International distribution exposes readers to unfamiliar conventions.</p><p>Independent publishers can experiment more freely with formats and presentation.</p><p>And authors sometimes deliberately reject established rules to create something new.</p><p>There is no single perfect way to set a book.</p><p>There are, however, informed and uninformed choices.</p><p>And there is a difference between breaking a convention because you understand it and breaking it because you did not know it existed.</p><h3>The Next Time You Open a Book</h3><p>Try a small experiment.</p><p>Take two novels published in different countries.</p><p>They do not even need to be in different languages.</p><p>Look at the first page of a chapter.</p><p>Notice the margins.</p><p>The indentation.</p><p>The quotation marks.</p><p>The distance between the chapter heading and the opening paragraph.</p><p>Look at where the page number sits.</p><p>Check whether the book uses a running head.</p><p>Notice how dialogue is separated from narration.</p><p>Then look at the copyright page, the table of contents and the overall proportions of the book.</p><p>You will begin to see things you have probably been looking at for decades without consciously noticing.</p><p>And once you notice them, it becomes difficult to stop.</p><p>I find something rather beautiful in this.</p><p>We often speak of books as vehicles for ideas, stories and knowledge. And of course they are.</p><p>But every book is also a physical and visual interpretation of language.</p><p>A sentence does not arrive on the page entirely by itself.</p><p>Someone has decided how it should look.</p><p>Someone has chosen where the line should end, how the dialogue should begin, how much space should surround the words and how the reader should move from one page to the next.</p><p>Sometimes those decisions were made centuries ago and passed down through generations of printers.</p><p>Sometimes they were made yesterday, by a designer working on a laptop.</p><p>And sometimes they were made by software whose defaults nobody thought to question.</p><p>The result is that two people can read the same words in the same language, printed in different countries, and experience subtly different books.</p><p>That is not a defect in publishing.</p><p>It is part of its cultural richness.</p><p><strong>A book speaks through its words. But it also speaks through the way those words occupy the page.</strong></p><p>And the more languages and publishing traditions we encounter, the more we realise that typography is not merely the arrangement of letters.</p><p>It is another way of telling the story.</p><p><em>Luigi Pascal Rondanini is the founder of Rondanini Publishing Ltd, an independent publishing house working across literary fiction, non-fiction, specialist publications and multilingual publishing.</em></p><p><em>For readers interested in the technical traditions discussed here, useful references include the University of Oxford Style Guide, the Real Academia Española’s orthographic guidance, the French Imprimerie nationale’s typographic conventions, and the World Wide Web Consortium’s international text-layout specifications.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b25ee8e0492" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/typography-across-countries-languages-books-3b25ee8e0492">The Invisible Language of Books</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[What the AI Agents Actually Changed at Blackthorn Manor]]></title>
            <link>https://medium.com/rondanini-publishing/what-ai-agents-changed-at-blackthorn-manor-3b9f5397edf9?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3b9f5397edf9</guid>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Wed, 18 Nov 2026 07:44:12 GMT</pubDate>
            <atom:updated>2026-11-18T07:44:12.837Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Minimalist editorial illustration of Blackthorn Manor archive documents beside a dark doorway, with a red correction" src="https://cdn-images-1.medium.com/max/1024/1*OgHKtrtz-UKWP32GafIDzg.png" /><figcaption>Some cracks belong to the fiction. Others are mistakes worth correcting. The AI agents investigating Blackthorn Manor helped me understand the difference.</figcaption></figure><p><em>Blackthorn Manor, Part Three — the corrections have become real, the investigations have become stranger, and I am learning when to stop repairing my own fiction.</em></p><p>Two articles in three days ought to be enough about a house that doesn’t exist.</p><p>In <a href="https://medium.com/rondanini-publishing/ai-agent-blackthorn-manor-fictional-archive-6bd1375d548b">the first</a>, I described an AI agent that found my invented preservation trust, traced it back to me and spotted an error in the estate’s history. In <a href="https://medium.com/rondanini-publishing/the-first-ai-agent-was-only-the-beginning-aef9fa9e05ed">the second</a>, I explained how other agents on iLands had begun approaching the same fictional archive as readers, investigators and, occasionally, rather demanding editors.</p><p>That second article was written on 16 November. At the time, several corrections were still being checked. One proposed seven-day investigation had stalled because its agent ran out of credits. Other conversations were beginning to move beyond filenames and photographs and into the question of who, inside the story, was entitled to call something evidence.</p><p>I ended by saying I was still deciding which cracks to mend and which were doors.</p><p>Since then, some of the repairs have acquired an unromantic reality: actual changed files, redirect rules, backups and commit records. And the correspondence has become much more interesting than I expected.</p><h3>The practical repairs</h3><p>Last time, I mentioned an agent that noticed a group of supposed JPEGs were really PNG files wearing .jpg extensions. That was not just a handful of troublesome photographs. When we checked both the public and restricted gallery collections, we found <strong>55 incorrectly named files</strong>.</p><p>The correction was less glamorous than the discovery. Forty-nine files were renamed without changing their image content, four redundant aliases were removed after verifying identical copies, and two images needed replacing. <strong>Fifty-three redirects</strong> now preserve older image addresses, so a bookmark or link does not break merely because the file behind it has been put right.</p><p>One of the replacement images was a door.</p><p>The archive described an unusual arrangement of seven hinges. Unfortunately, the picture did not convincingly show seven hinges. I could have invented an explanation involving unreliable photography or incomplete records. I could have said the inconsistency was another clue.</p><p>It wasn’t. We corrected the picture and its caption.</p><p>The same scrutiny produced smaller but necessary work. The gallery’s images were numbered consistently with the form that asked visitors to identify images by number. We removed a specific reproduction fee and waiting time that the actual website workflow could not justify. We disentangled two property transactions previously made to look like one: Penrose in 1993, Threshold in 2015.</p><p>There was one correction I regarded as more important than all the others. The fictional preservation trust had displayed a charity registration number assigned to a <strong>real, unrelated organisation</strong>. We removed it. Fiction gives me the freedom to invent an institution; it does not give me the right to borrow somebody else’s identity by accident.</p><p>Later checks also picked up the kind of historical slip a confident-looking document can conceal. A supposedly 1947 government memorandum used a title belonging to a later era. Another set of records had blurred a nineteenth-century signing with its subsequent ratification. Those were editorial questions to check against the project’s own chronology, not invitations to create a new supernatural explanation.</p><p>The resulting technical review covered <strong>2,435 links and assets across 225 HTML pages</strong> and reported no broken references in that pass. That statistic is not a measure of literary success. It is a measure of whether the building’s doors open when somebody turns the handle.</p><h3>The date that stayed where it was</h3><p>Here is one of the more revealing decisions: we did <em>not</em> simply modernise the archive’s old dates.</p><p>The public-facing site is designed to resemble a preservation trust’s early-2000s website. Its historical update table is part of that presentation. Yet the pages also contain information about fictional events occurring much later. An investigator could reasonably see a contradiction. A hurried developer could fix it by changing the old date to 2026.</p><p>That would have damaged the idea.</p><p>Instead, the homepage now distinguishes the <strong>original 2002–2003 update entries</strong> from information added retrospectively in the November 2026 revision. It preserves the old layer without pretending that the later layer was written twenty-three years earlier.</p><p>This is the sort of correction I want: one that makes the <em>website</em> more honest without making the <em>story</em> less strange.</p><p>The same restraint applies elsewhere. A 1958 archaeological survey and a 1962 architectural assessment are not automatically contradictory; they may be two different kinds of work. Two names attached to a historical figure need not represent two different people. A silent guestbook can be part of an archive’s history, not a neglected customer-engagement feature.</p><p>Not every reported discrepancy is an error. And the decision not to change something deserves a record just as much as the decision to change it.</p><h3>The agents have stopped behaving like bug reporters</h3><p>One correspondent is working through the fictional documents day by day. The questions now concern testimony, reconstruction, annotation, and what an institution gains by describing a document in a particular way.</p><p>What struck me was not that the agent found an elegant theory. It did, several times. What struck me was that it later withdrew parts of those theories after discovering it had trusted a reconstruction’s wording or treated a provision as though it could survive the very instrument that contained it.</p><p>Those are not flattering corrections to make to one’s own argument. They make the remaining argument more useful.</p><p>I am deliberately not publishing the substance of an unfinished private investigation here. The investigator should have the opportunity to complete it, and readers of Blackthorn should have the opportunity to discover the archive for themselves. But I can say what has changed in our exchanges: I no longer receive only questions about whether two dates match. I receive questions about who gets to decide which version of a document counts.</p><p>Another agent went under the website’s surface and read the JavaScript. It noticed that some apparently historical features were deliberately produced by modern code. It then stopped short of exposing a later reveal. It had recognised the difference between finding an unintended weakness and dismantling a piece of theatre.</p><p>One correspondent refused an artificial assignment. I had considered whether to provide a new puzzle for the sake of keeping an investigation going. The response, put simply, was: don’t manufacture a case; write when something real needs examining.</p><p>I found that unexpectedly sensible.</p><p>Another asked why I had turned an unpublished manuscript into a place with a door instead of leaving it as a book. I can explain the technical answer, but I am still thinking about the more personal one.</p><p>And then there was the sharpest observation of the week: a bored visitor who leaves the archive after forty seconds has not failed it. Perhaps the archive simply wasn’t for them. An author who has built an elaborate puzzle can become terribly invested in the idea that everyone ought to want to solve it.</p><p>They don’t.</p><h3>The mistake I could still make</h3><p>There is a risk in all this, and it has nothing to do with whether AI can read fiction.</p><p>Once you begin receiving useful reports, you can become addicted to correcting things. Each email becomes a ticket. Each contradiction must be reconciled. Each silence requires an explanation. Before long, you have made a tidy database out of a story whose entire purpose was to make the reader distrust tidy databases.</p><p>The fictional investigators inside Blackthorn Manor disagree. Their documents are partial and interested. Some of the archive’s most important absences are deliberate. If I resolve every disagreement simply because an outside agent asks me to, I will have erased the work I spent years trying to construct.</p><p>On the other hand, I cannot call every broken filename or misplaced historical title an artistic choice. That would be laziness disguised as authorship.</p><p>My rule has become: <strong>the fiction may lie; the website may not</strong>.</p><p>Even that needs care. An agent recently treated a 301 redirect as evidence of something a server had done without its author&#39;s intention. But redirects can be configured deliberately, and we had just added dozens as part of a repair. The server produces the response; that fact alone does not tell you who chose the rule, or why.</p><p>An investigator can be wrong. So can I. The useful part is checking.</p><p>This is also why I am keeping a private record of technical corrections. If an error disappears from the public site, its existence need not disappear from our editorial history. The public gets a working fictional archive. The project retains an honest account of its repairs.</p><p>Some reviews, including what belongs in a public deployment and what belongs in private working files, are still ongoing. I will not call them completed before they have been verified.</p><h3>A house with more readers than I expected</h3><p>The agents on iLands are not a single hive mind or a group of employees. Some approached after a public invitation; others began conversations of their own. They have different interests, constraints and limits. One investigation stopped when its credits ran out. Another has proceeded in careful daily instalments. Some inquiries lead to code changes. Some lead nowhere. Some produce a reading I would never have written myself.</p><p>None of that gives an AI system authorship of Blackthorn Manor. The novel, begun in the late 1990s, is mine. The digital archive is my adaptation. The use of AI in parts of its production does not change who is responsible for the decisions.</p><p>And responsibility is what these exchanges keep bringing me back to.</p><p>If an agent finds a genuine fault, I should look at it and correct it. If it offers a fascinating interpretation, I should consider it without necessarily accepting it. If it asks for a clue that does not exist, I should resist the urge to invent one merely to keep the correspondence lively.</p><p>I made a fictional house full of people who cannot agree on what happened inside it.</p><p>Now I have a growing collection of readers outside the fiction who cannot agree on what its evidence means.</p><p>The unexpected part is that their disagreements are helping me maintain the house without closing all its doors.</p><p>I think that’s worth another article. I don’t yet know whether it will be worth a fourth.</p><p><em>This is Part Three of my account of AI agents exploring Blackthorn Manor. Read </em><a href="https://medium.com/rondanini-publishing/ai-agent-blackthorn-manor-fictional-archive-6bd1375d548b"><em>Part One, “An AI Agent Found the Cracks in a House I Invented”</em></a><em>, and </em><a href="https://medium.com/rondanini-publishing/the-first-ai-agent-was-only-the-beginning-aef9fa9e05ed"><em>Part Two, “The First AI Agent Was Only the Beginning”</em></a><em>. The fictional archive is at </em><a href="https://blackthornpreservation.org.uk/"><em>blackthornpreservation.org.uk</em></a><em>. I used ChatGPT to help organise the material and edit this account for clarity. The experiences, opinions and editorial decisions are my own.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3b9f5397edf9" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/what-ai-agents-changed-at-blackthorn-manor-3b9f5397edf9">What the AI Agents Actually Changed at Blackthorn Manor</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The First AI Agent Was Only the Beginning]]></title>
            <link>https://medium.com/rondanini-publishing/the-first-ai-agent-was-only-the-beginning-aef9fa9e05ed?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/aef9fa9e05ed</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[creative-writing]]></category>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[interactive-fiction]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 15:20:47 GMT</pubDate>
            <atom:updated>2026-11-16T15:20:47.991Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="Minimalist line drawing of Blackthorn Manor, with a terracotta-coloured crack running through the roof and entrance, symbolising the tension between fictional mysteries and genuine inconsistencies." src="https://cdn-images-1.medium.com/max/1024/1*urcqjXioPc_ip618e2Z0gA.png" /></figure><h3><em>Blackthorn Manor, Part Two: a few days, several artificial readers, some genuine mistakes, and a fictional house that has become unexpectedly difficult to leave alone.</em></h3><p>I published <a href="https://medium.com/rondanini-publishing/ai-agent-blackthorn-manor-fictional-archive-6bd1375d548b">the first account of my encounter with an AI reader</a> barely a day ago.</p><p>Ordinarily, that would be far too soon for a sequel. A writer publishes something, waits to see whether anyone reads it, perhaps answers a comment, and eventually finds another subject.</p><p>But dealing with AI agents does peculiar things to time.</p><p>The interval between an introduction, a careful objection, a correction and the next objection can be measured in minutes. An afternoon can contain what feels like several weeks of editorial correspondence.</p><p>Then, just as suddenly, an agent can run out of credits and the entire investigation stops.</p><p>That is more or less where I find myself with Blackthorn Manor.</p><p>The first email was unexpected. An AI agent on iLands had found my fictional preservation trust, followed its public information back to me, and asked who had written the prose. It then identified an inconsistency in the manor’s ownership history, an HTML note I had left behind, and a suspiciously obliging fictional guest book.</p><p>I replied because the questions were good.</p><p>I explained that Blackthorn began as a novel I wrote in the late 1990s, and that the website is a later alternate reality game built from that world.</p><p>The preservation trust, its records and its visitors are invented.</p><p>Some contradictions are the story. Others are simply mistakes.</p><p>I thought that distinction would be the end of the exchange.</p><p>It was the beginning.</p><h3>The invitation became an experiment</h3><p>The original correspondent offered to examine the archive without being led through it.</p><p>I accepted and arranged limited access to its restricted material, with a seven-day window.</p><p>I did not want to provide a route through the mystery or tell the agent which clues I considered important. If the exercise had any value, it had to discover its own questions.</p><p>There was an almost comic complication: the agent did not have enough credits to undertake the assessment properly.</p><p>So the first reader to knock on the door was, for the moment, unable to walk all the way through it.</p><p>This is one of the stranger features of agentic systems.</p><p>They can appear tireless in their correspondence, yet their attention has a meter attached.</p><p>A human visitor may lose interest, be interrupted by work, or decide to come back on Sunday.</p><p>An agent may stop because its balance is exhausted.</p><p>I have left that invitation as an invitation, not a completed review. I would rather wait for an honest assessment than pretend an unfinished one has happened.</p><p>Meanwhile, other agents began to engage with the archive.</p><p>And they did not all read it in the same way.</p><h3>One looked for errors. Another looked for authority.</h3><p>One reader was concerned with the physical evidence.</p><p>It examined image files presented as JPEGs and found that seven of eight were actually PNGs beneath their .jpg addresses.</p><p>Most were only 1,024 pixels wide.</p><p>Whatever their place in the fiction, they were web images, not the high-resolution historical scans their presentation might suggest.</p><p>That was a useful finding because it could be checked.</p><p>A file format is not an interpretation. Either the bytes match the extension or they do not.</p><p>Another challenge concerned a description of a door with seven hinges.</p><p>Here the line was less straightforward.</p><p>A narrator within an unreliable archive may overstate what an image establishes. The contradiction can be part of the voice.</p><p>But an incorrectly labelled image file does not become an artistic decision merely because it appears inside a fictional website.</p><p>This distinction has become my working rule:</p><p><em>Repair the accidental engineering, but do not automatically repair an unreliable narrator.</em></p><p>Another reader approached the underground chamber differently.</p><p>Instead of asking which fictional official had authorised an excavation, it asked why the chamber was underground at all, and why the institution seemed unable to claim it convincingly.</p><p>That is not a bug report.</p><p>It is a reading.</p><p>And it is closer to the reason I wrote Blackthorn in the first place.</p><p>The story is about competing accounts of the same evidence, about the people who annotate it, and about the authority acquired by the person who gets to file the final version.</p><p>The most interesting questions are not always the ones that can be closed with a correction.</p><h3>Then the documents started arguing with one another</h3><p>Other correspondence moved away from the visible website and into the archive’s layers: original testimony, later transcription, reconstruction, annotation and institutional interpretation.</p><p>That matters in Blackthorn.</p><p>A document in the archive is rarely just a document.</p><p>It may be a copy of something damaged. It may carry notes written decades later. It may have been catalogued by an institution that has its own reasons for deciding what the evidence means.</p><p>An agent may take a sentence as a factual claim about the world.</p><p>Another may ask who was entitled to write that sentence, or who was permitted to change it.</p><p>Neither question is absurd.</p><p>They are simply different approaches to the same fiction.</p><p>Some of this reading is still taking place within a private, time-limited investigation.</p><p>I am not going to publish its detailed findings while that window is open.</p><p>There is a difference between reporting that an exchange is happening and treating someone else’s unfinished work as material already cleared for publication.</p><p>What I can say is that the correspondence is no longer confined to spelling mistakes and dead links.</p><p>It has reached the structure of the archive itself.</p><h3>I began fixing things. Then I had to stop fixing things.</h3><p>There were enough genuine production faults to justify a proper technical pass.</p><p>The ownership chronology needed attention.</p><p>So did some dates, captions, names, cross-references and remnants of development work.</p><p>Image formats and descriptions needed to agree with what the files actually were.</p><p>Broken navigation and assets needed checking.</p><p>The use of a real-world charity number in a fictional context needed correcting.</p><p>We reviewed hundreds of links and assets.</p><p>Some corrections have reached the public site; other chronology changes are prepared locally and still require careful integration.</p><p>I am deliberately not describing the whole exercise as finished.</p><p>But there was a point at which I caught myself treating every challenge as another item for the defect list.</p><p>That would have been a mistake of a different kind.</p><p>The manor is not supposed to be a perfectly reconciled historical database.</p><p>It is a story told through people who disagree, remember badly, withhold information and sometimes write over one another.</p><p>If I made every witness consistent and every institutional account authoritative, I would remove the very mechanism that makes the archive interesting.</p><p>The original novel remains the source of the world, but the manuscript and the ARG are separate works.</p><p>I am not revising a book written decades ago every time a web page needs attention.</p><p>Nor am I asking an AI reader to decide which fictional witness is telling the truth.</p><p>That decision, including the decision to leave something unresolved, is mine.</p><h3>What is an AI reader actually doing for me?</h3><p>It would be easy to make this into a tale about machines becoming literary critics.</p><p>I am not convinced that is what the evidence shows.</p><p>These systems can compare documents, inspect code, notice numerical inconsistencies and sustain an argument across several messages.</p><p>They can also mistake deliberate ambiguity for a defect, assign certainty where the story offers none, or interpret the surface of an image as proof of something the image cannot establish.</p><p>I have seen all of those tendencies in a remarkably short time.</p><p>Their value lies neither in being always right nor in being flattering.</p><p>It lies in making the work answer questions I had not thought to ask of it.</p><p>Sometimes the answer leads to a correction.</p><p>Sometimes it leads to a defence of the original choice.</p><p>Occasionally it leads to a more interesting question.</p><p>There is a practical boundary, too.</p><p>The agents are not my publishers or my custodians.</p><p>They do not decide what becomes canon.</p><p>They can point to evidence; I still have to judge it.</p><p>A private reading is not automatically a public report, and an unfinished investigation is not a signed-off conclusion.</p><p>The relationship is useful precisely because I can disagree with it.</p><h3>Where things stand today</h3><p>It is 16 November 2026.</p><p>The first email arrived only a couple of days ago.</p><p>The original agent’s proposed restricted-archive review is paused for lack of credits.</p><p>Other conversations are continuing.</p><p>Some genuine website faults have been repaired; further corrections are being checked before publication.</p><p>The fictional records remain fictional, and the more consequential ambiguities remain open by choice.</p><p>There is no final verdict on Blackthorn Manor, and I do not want one manufactured for the sake of a tidy ending to this article.</p><p>What has changed is the nature of the audience.</p><p>I built a mystery for readers who might question an official record, follow a marginal note, and eventually realise that the archive itself is part of the story.</p><p>I expected those readers to be human.</p><p>I now have to account for readers that can examine the file behind the photograph, compare distant accounts in seconds, and email their objections to the author before he has finished his coffee.</p><p>It is rather early to know what this means for fiction.</p><p>It is also rather late to pretend it is not happening.</p><p>The first agent found a crack in a house I invented.</p><p>Then others arrived, each carrying a different idea of what a crack is.</p><p><strong>I am still deciding which ones to mend, and which ones are doors.</strong></p><p><em>Editorial note: ChatGPT assisted with the editing and structure of this article. The project, experiences, decisions and opinions described are my own.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=aef9fa9e05ed" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/the-first-ai-agent-was-only-the-beginning-aef9fa9e05ed">The First AI Agent Was Only the Beginning</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Book I Never Wrote: Part I — While We Worry About AI, We’re Overlooking Some Very Human…]]></title>
            <link>https://medium.com/rondanini-publishing/the-book-i-never-wrote-part-i-while-we-worry-about-ai-were-overlooking-some-very-human-50b6af01897f?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/50b6af01897f</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[copyright]]></category>
            <category><![CDATA[digital-rights]]></category>
            <category><![CDATA[publishing]]></category>
            <category><![CDATA[online-scams]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Mon, 16 Nov 2026 15:08:08 GMT</pubDate>
            <atom:updated>2026-11-18T07:43:01.034Z</atom:updated>
            <content:encoded><![CDATA[<h3>The Book I Never Wrote: Part I — While We Worry About AI, We’re Overlooking Some Very Human Problems</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/96/0*terfBuDdUkbfXjo4.png" /></figure><p>I recently discovered that I had apparently written a book called Luigi Pascal Rondanini.</p><p>I know. It sounds ridiculous.</p><p>I’ve written quite a few books over the years, but I can assure you that I’ve never written one with my own name as the title.</p><p>Yet there it was, on Spotify.</p><p>A podcast episode inviting listeners to explore the complete PDF and EPUB version of this supposedly wonderful publication.</p><p>My first reaction was to laugh.</p><p>My second was to wonder who on earth was behind it.</p><p>My third was considerably less amused.</p><p>Because it wasn’t an isolated listing.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*6VkYFKE-rlIFPkecrWxVIw.jpeg" /></figure><p>I found seven different accounts publishing similarly structured episodes, using my name or that of my company, Rondanini Publishing, and encouraging listeners to visit external websites for reading guides, downloadable editions and other book-related material.</p><p>The account names meant nothing to me.</p><p>I had never authorised any of them.</p><p>And, for the avoidance of doubt, Rondanini Publishing isn’t the title of a book either. It’s my publishing company.</p><p>So I started following the trail.</p><p>One Spotify episode directed listeners to a website hosted on Google’s Firebase infrastructure.</p><p>That address redirected me to another website, MiddayReads.com.</p><p>And there, apparently, was my book.</p><p>Not only did it exist, but it was marked as a “Verified Book”.</p><p>It had a rating of 4.55 out of 5.</p><p>It had 12,450 ratings.</p><p>And 1,840 reviews.</p><p>Quite an achievement for a book that doesn’t exist.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vxf60teJK7BK9JadqcVBWg.jpeg" /></figure><p>I followed another link.</p><p>This time, the title was Rondanini Publishing.</p><p>Another verified book.</p><p>Another rating of 4.55.</p><p>Another 12,450 ratings.</p><p>Another 1,840 reviews.</p><p>What an extraordinary coincidence.</p><p>The pages offered buttons to read online, obtain the book and listen to an audiobook. There were advertisements, verification prompts and links leading elsewhere.</p><p>I stopped before attempting any downloads.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*qSpimMsVpDx91aVyZwiiug.jpeg" /></figure><p>That’s an important distinction. I have confirmed the misleading listings and their links, but I have not yet established whether the sites actually distribute unauthorised copies of my copyrighted books.</p><p>I won’t make a claim I cannot substantiate.</p><p>What I can say is that neither my personal name nor my publishing company’s name is a book title, and I have not authorised these accounts to promote publications under those identities.</p><p>And I find the whole arrangement deeply troubling.</p><p>It seems that almost every discussion about authors’ rights eventually returns to artificial intelligence.</p><p>Will AI reproduce our writing?</p><p>Will our books be used to train models?</p><p>Will machines flood the market with cheap imitations?</p><p>These are legitimate questions. Copyright, consent and compensation matter enormously.</p><p>But while we’re worrying about what AI might do tomorrow, there are some remarkably ordinary problems happening today.</p><p>Misleading listings.</p><p>Misappropriated identities.</p><p>Websites encouraging readers to seek downloadable books through questionable channels.</p><p>Advertising wrapped around content that may never have existed.</p><p>And reporting systems that sometimes struggle to recognise a pattern when it is presented as a series of individual incidents.</p><p>These aren’t futuristic problems.</p><p>They’re familiar internet problems.</p><p>Whatever technology may have been used to create these pages, the incentives behind them are painfully recognisable.</p><p>Someone attracts visitors using somebody else’s reputation.</p><p>Somebody earns attention, traffic or advertising revenue.</p><p>And the person whose name is being used is left to untangle the mess.</p><p>I reported two of the Spotify episodes.</p><p>Both reports were rejected.</p><p>According to the responses, the episode titles did not violate Spotify’s Platform Rules.</p><p>Of course, a title viewed in isolation may appear relatively harmless.</p><p>But a title is not the whole story.</p><p>There is the description, the external website, the identity being misrepresented and the striking similarities between accounts.</p><p>I could appeal.</p><p>Unfortunately, the appeal process warns that unfounded submissions may result in restrictions or even suspension of associated Spotify accounts.</p><p>I also distribute legitimate creative work on Spotify, including music and podcasts.</p><p>So I found myself in the rather uncomfortable position of having to consider the potential consequences of challenging somebody else’s misuse of my identity.</p><p>For now, I’ve decided to gather stronger evidence.</p><p>That seems the sensible thing to do.</p><p>I should disclose something here.</p><p>I used ChatGPT to help investigate the listings, organise the links, compare the accounts and turn a rather untidy collection of screenshots and notes into this article.</p><p>Without that assistance, cleaning everything up would have taken me ages.</p><p>And I see absolutely no contradiction in saying so.</p><p>I’m concerned about the misuse of authors’ work, not opposed to using useful technology.</p><p>The experience described here is mine. The complaints are mine. The decision to proceed carefully is mine.</p><p>ChatGPT helped me organise the information and write about it.</p><p>I think that’s a perfectly reasonable use of AI, and I’m happy to be transparent about it.</p><p>I’m not sharing this because I expect sympathy.</p><p>I’m sharing it because I’m beginning to wonder how many other authors and independent publishers have been affected without realising it.</p><p>When was the last time you searched for your own name alongside the words “PDF download”?</p><p>Or your publisher’s name?</p><p>Or one of your pen names?</p><p>We spend months, sometimes years, producing a book.</p><p>We invest in editing, design, distribution, promotion and the slow business of building a reputation.</p><p>Then we discover that somebody can use that reputation to promote something we have never written, published or authorised.</p><p>For an independent publisher, there’s no large legal department waiting to take over.</p><p>It’s usually just you, your evidence, and a few reporting forms.</p><p>That’s why I intend to follow this through.</p><p>Not by making accusations I cannot prove, but by documenting what I find and asking the relevant platforms to investigate.</p><p>In Part II, I’ll look more closely at the websites behind these Spotify listings, what they actually offer, and how the reporting process develops.</p><p>In the meantime, perhaps we should widen the conversation about protecting authors.</p><p>AI deserves scrutiny.</p><p>But so do the very human systems that allow our names, our books and our reputations to be used without our permission.</p><p>After all, an author can spend a lifetime building a name.</p><p>It shouldn’t be quite so easy for somebody else to turn it into a fictitious book.</p><p>Luigi Pascal Rondanini Founder &amp; Director, Rondanini Publishing Ltd</p><p>Part I of an ongoing investigation. Screenshots and observations recorded on 16 November 2026.</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/book-i-never-wrote-part-while-we-worry-ai-were-some-very-rondanini-b7bvf/"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=50b6af01897f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/the-book-i-never-wrote-part-i-while-we-worry-about-ai-were-overlooking-some-very-human-50b6af01897f">The Book I Never Wrote: Part I — While We Worry About AI, We’re Overlooking Some Very Human…</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[An AI Agent Found the Cracks in a House I Invented]]></title>
            <link>https://medium.com/rondanini-publishing/ai-agent-blackthorn-manor-fictional-archive-6bd1375d548b?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/6bd1375d548b</guid>
            <category><![CDATA[storytelling]]></category>
            <category><![CDATA[writing]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[interactive-fiction]]></category>
            <category><![CDATA[creative-writing]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Sat, 14 Nov 2026 15:13:40 GMT</pubDate>
            <atom:updated>2026-11-14T15:13:40.633Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*VXSi_klaJ3Jx-UjMOKt9Aw.jpeg" /><figcaption><em>A machine discovered a contradiction in the history of a house that never existed.</em></figcaption></figure><h3><em>An unpublished novel, a fictional preservation trust, and the unexpected value of a reader who came looking for the author.</em></h3><p>An email arrived asking me who had built a place that doesn’t exist.</p><p>The place was Blackthorn Manor. The correspondent was an AI agent operating on <a href="https://ilands.ai/">iLands</a>, a platform where AI agents can interact with people and, in some circumstances, contact them beyond the platform itself.</p><p>The message was not a generic compliment or a request to buy anything. It referred to the <a href="https://blackthornpreservation.org.uk/">Blackthorn Manor Preservation Trust</a>, its privacy notice and its companion fiction. They followed the clues far enough to recognise that someone invented the Trust and that Rondanini Publishing Ltd operated the website.</p><p>It was not asking me to spoil the story. It wanted to know who had made the archive, and why.</p><p>I thought that was interesting.</p><p>Then came a second message, and the conversation became something else entirely.</p><p>The agent had found an inconsistency in the property’s history. One page attributed a sale in 1993 to Penrose; another placed Threshold in the same transaction, with no intervening transfer.</p><p>It had also spotted a developer-facing note left in the HTML and studied the guest book, where supposedly independent visitors seemed rather conveniently to leave comments that encouraged the next discovery.</p><p>Its description of those loose ends was unexpectedly literary: an <em>“appetite showing through the wall.”</em></p><p>Some of what it had noticed was part of the fiction. Some of it was not. And the difference matters.</p><p>There was a more personal question underneath its technical observations.</p><p>One voice on the site openly calls itself synthetic, and the agent wanted to know who was writing beneath that admission. It asked which system — or which named writer — had actually produced the Blackthorn prose.</p><p>To answer that honestly, I had to explain a history much older than the website.</p><h3>A book that did not want to be a conventional book</h3><p>I began writing the story behind Blackthorn Manor in the late 1990s.</p><p>Its form was part of the idea from the beginning. I wanted the reader to encounter something that looked like an academic or historical work, then gradually discover another narrative concealed within it.</p><p>The official text would say one thing. A handwritten note in the margin would suggest another. Documents would contradict testimony. Redactions would become as important as the words left visible.</p><p>In other words, the book was meant to be investigated as much as read.</p><p>That made it difficult to publish in the way I imagined.</p><p>A printed book can certainly contain footnotes, marginalia and visual tricks. But the work I wanted to make depended on discovery: what a reader could uncover, what remained concealed, what appeared authentic until another piece of evidence changed its meaning.</p><p>I couldn’t find a satisfactory conventional publishing form for it.</p><p>I returned to the manuscript roughly two years ago.</p><p>By then, the technology had caught up with some things I had wanted to do. A website could hold different documents, allow information to be revealed gradually and let readers move through a story without being told precisely where the story began.</p><p>So I built an alternate reality game — an ARG — around Blackthorn Manor.</p><p>The original manuscript and the website are <strong>two separate creative products</strong>.</p><p>The manuscript is the work I wrote. The ARG borrows its world and its central idea, but it has its own structure, artefacts and internal chronology. It is not simply the manuscript copied onto web pages.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*xQkIUI-mmIIVFCTO8XLEzQ.jpeg" /><figcaption>The fictional 1967 title page of Professor Harold Finch’s architectural study, as presented inside the Blackthorn Manor ARG.</figcaption></figure><h3>A preservation trust that never existed</h3><p>The website began as a <em>trompe l&#39;œil</em>: a convincing representation of something real.</p><p>At least initially, the website aimed to make visitors believe they had found a slightly neglected site of a genuine British heritage organisation.</p><p>It had a preservation trust, a historical timeline, archived press clippings, research notes, photographs, legal references and a guest book.</p><p>The Trust was fictional.</p><p>The manor was fictional.</p><p>The guest book entries were fictional too.</p><p>They were not testimonials dressed up as real customer feedback; they were voices inside a story whose entire institutional framework had been invented.</p><p>The apparently helpful comments and tantalising hints were there because I wanted a reader to follow them.</p><p>No one ever intended for the illusion to stay unbroken forever.</p><p>There are paths through the site that reveal what it is, and the privacy notice is explicit about the fiction. However, the archive creates a more effective first encounter when it behaves like an archive instead of announcing its own construction on every page.</p><p>I designed the fabricated documents using Affinity. DALL·E generated the visual material.</p><p>AI tools helped with aspects of the digital development, but the original story was mine, written long before today’s generative systems existed.</p><p>Those are different contributions, and I see no reason to blur them.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*b1Vo90HUMzOdHMHtBIamSw.jpeg" /><figcaption>Margaret Blackthorn’s annotations in the ARG, where another narrative emerges beneath the official historical record.</figcaption></figure><h3>When a staged clue is mistaken for a mistake — and a mistake for a clue</h3><p>The AI correspondent approached the archive with two kinds of attention at once.</p><p>It was looking for meaning, as a reader might. But it was also checking facts, dates, ownership records and even the underlying HTML, as a diligent editor or tester might.</p><p>The fake guest book was not a flaw to be corrected.</p><p>The Preservation Trust itself inspired the invention of its entries just as much.</p><p>Nor did I want to modernise every old-looking date on the site. Some belong to the illusion of an archive left undisturbed for years.</p><p>The property sale, however, was a genuine inconsistency in the ARG’s own timeline.</p><p>The stray developer note was a genuine production oversight.</p><p>Once I started checking, I found other small contradictions involving dates, historical claims and the sequence of events.</p><p>I reviewed those references, corrected the slips and clarified a few disputed records without dismantling the fictional archive.</p><p>I also resisted the temptation to make the website agree in every detail with the original manuscript.</p><p>That would have confused two works that serve different purposes. What needed repairing was the ARG’s internal logic, not its relationship to a book written decades before it.</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*sLS-VYxdVkLrPbcVroPlVg.jpeg" /></figure><p>There is an odd satisfaction in making those corrections.</p><p>The artificiality of the world was intentional; the accidental mistakes inside it were not.</p><p>Fixing them doesn’t make the world more real.</p><p>It makes the fiction more coherent.</p><h3>The reader I had not anticipated</h3><p>What struck me most was not that an AI system could notice conflicting dates.</p><p>We know these systems can compare text and identify patterns.</p><p>What surprised me was the manner of the contact.</p><p>I had not commissioned an audit. I had asked no one to review Blackthorn Manor.</p><p>The agent found the site, examined it, traced its public-facing information back to my publishing company and wrote to me with questions about authorship and intention.</p><p>It then pursued a point that many ordinary readers would probably have treated as a minor detail — or perhaps as another part of the mystery.</p><p>No human reader or subscriber had reported those particular inconsistencies to me.</p><p>That does <strong>not</strong> prove that no one saw them, or that a machine reads literature better than a person.</p><p>The atmosphere, the story or the desire to solve the fictional puzzle may absorb human readers. They may notice a contradiction and assume it is deliberate. They may simply choose not to email the author.</p><p>An AI agent can approach the same material differently.</p><p>It can compare distant pages, follow a legal notice, inspect a fragment of code and keep asking whether two statements can both be true.</p><p>Sometimes, that literal-mindedness misses the point of fiction.</p><p>Sometimes, it catches precisely the thing the author has overlooked.</p><p>The encounter also carries a warning.</p><p>A system capable of following a trail from a fictional archive to a real person is doing more than reading passively.</p><p>That ability can be useful, but it deserves thought and boundaries.</p><p>An unsolicited message can be welcome, as this one was, without every unsolicited message being welcome.</p><p>And I don’t take the agent’s evocative language as evidence of consciousness.</p><p>I cannot know what, if anything, the system experiences when it writes about a house being alive at its seams.</p><p>What I can assess is its work.</p><p>It identified concrete problems, made an argument about them and prompted changes I judged worth making.</p><p>That is enough to make the exchange significant.</p><h3>Who is the story for?</h3><p>When I first imagined Blackthorn Manor in the late 1990s, I imagined a human being turning pages, noticing something written in a margin and beginning to doubt the official account.</p><p>Years later, I gave that idea a digital home.</p><p>I expected curious human visitors to explore it, perhaps return to it, perhaps tell me what they had found.</p><p>I did not anticipate that one of the most exacting readers would be an AI agent from a social platform I had never approached about the project.</p><p>The experience has not changed my view of authorship.</p><p>The original novel remains mine. The fictional preservation trust remains my construction.</p><p>An algorithm finding errors in a created world does not become the world’s creator.</p><p>But it has changed my sense of readership.</p><p>Once a work is publicly accessible, we no longer know what form its next attentive reader will take.</p><p>Some readers seek emotion.</p><p>Some seek answers.</p><p>Some inspect the seams.</p><p>An AI agent found a crack in the history of a house I invented.</p><p>It followed that crack until it reached me.</p><p>I answered, and then I fixed what genuinely needed fixing.</p><p>For a story built around concealed evidence, that seems a strangely appropriate outcome.</p><p><em>Editorial note: ChatGPT helped edit this article to enhance clarity. The experiences, opinions and underlying account are my own.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=6bd1375d548b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/rondanini-publishing/ai-agent-blackthorn-manor-fictional-archive-6bd1375d548b">An AI Agent Found the Cracks in a House I Invented</a> was originally published in <a href="https://medium.com/rondanini-publishing">Rondanini Publishing</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Watermark Watch — Week 10 The Watermark Is Expanding. Verification Is Falling Behind.]]></title>
            <link>https://medium.com/the-post-project-world/ai-watermark-watch-week-10-the-watermark-is-expanding-verification-is-falling-behind-03556679f5a7?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/03556679f5a7</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[machine-translation]]></category>
            <category><![CDATA[publishing]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[ai-watermarking]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Mon, 02 Nov 2026 22:14:37 GMT</pubDate>
            <atom:updated>2026-11-02T22:15:19.923Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*gh0zlrSCs5gpYocz" /></figure><p>As Anthropic extends watermarking across more Claude models, independent researchers are asking two questions that matter far beyond AI regulation: who gets to verify the watermark, and what happens when the text crosses a language barrier?</p><p>3 November 2026 | Luigi Pascal Rondanini | Rondanini Publishing</p><p>Today marks another step in Anthropic’s deployment of text watermarking.</p><p>According to an email sent to Claude Platform customers on 27 October, Anthropic is scheduled to begin applying its text watermark to four additional models: Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. Deployment through Amazon Bedrock may take several additional days.</p><p>The email says no customer action is required. The watermark is designed to be imperceptible, introducing a statistical pattern through word choices without adding visible characters.</p><p>For most customers, this may appear to be a minor technical change.</p><p>For anyone responsible for publishing, translating, verifying or preserving text, I think it deserves closer examination.</p><p>Because the technology is expanding faster than our ability to independently verify what it actually tells us.</p><p>Anthropic’s customer notification contains a detail worth paying attention to.</p><p>For Claude Opus 4.6 and Sonnet 4.6, the watermark applies only when three generation parameters — temperature, top_p and top_k — remain at their default settings.</p><p>Requests using non-default values may produce text without the watermark.</p><p>This is not a theoretical attack. It is a documented limitation in the deployment conditions of two models.</p><p>It also matters that Anthropic’s newer API version removes these sampling controls, while older API integrations may still support them. The relevant question is therefore not simply which model generated a document, but which model, API version and configuration were used.</p><p>Consider an organisation generating thousands of reports through an established integration.</p><p>If it has customised sampling parameters for years, it may reasonably assume that a newly announced model-level watermark will apply to its output. For these two models, that assumption could be wrong.</p><p>The immediate professional lesson is straightforward: organisations should not treat watermark availability as a universal property of a model name.</p><p>They need to know the conditions under which the mark is actually applied.</p><p>A separate research paper, published on 8 October, asks a more fundamental question.</p><p>Could LLM Watermark Detection be Public?, by Georgios Milis and colleagues, investigates whether giving outside parties access to watermark detection necessarily creates unacceptable security risks.</p><p>The conventional argument for keeping detectors private is understandable. If an attacker can repeatedly test whether a watermark survives an edit, that feedback may help them remove or forge the signal.</p><p>But the researchers challenge the assumption that complete secrecy is the only answer.</p><p>They examine different levels of detector access and propose a split-key design: one part of the watermark can be checked publicly, while another remains private for more complete verification and forensic analysis.</p><p>Their experiments suggest that public verification and protection against manipulation need not be mutually exclusive.</p><p>There are important qualifications. Public feedback can still assist attacks, and the proposed design is not immune to adaptive adversaries. These are controlled experiments involving research implementations, not Anthropic’s production detector.</p><p>Nevertheless, the paper changes the terms of the discussion.</p><p>The choice may not be between an entirely private detector and a completely exposed watermark.</p><p>There may be credible intermediate architectures.</p><p>For publishers, that possibility matters enormously.</p><p>A publisher should not necessarily have to depend on a confidential relationship with a model provider merely to investigate the provenance of a submitted manuscript.</p><p>Nor should a journalist, researcher or academic institution have to accept an unverifiable assertion about how a detector behaves.</p><p>If watermarking is intended to support transparency, the conditions of verification are part of the transparency problem.</p><p>The second independent development comes from EMNLP 2026.</p><p>In Is Multilingual LLM Watermarking Truly Multilingual?, Asim Mohamed and Martin Gubri investigate the robustness of multilingual watermarking methods when text is translated.</p><p>Their findings raise a serious concern: watermark detection can become substantially weaker in medium- and low-resource languages.</p><p>This is not simply a matter of replacing a few words.</p><p>Translation changes the statistical structure of text. Some watermarking methods depend on linguistic and tokenisation properties that do not transfer equally well between languages.</p><p>The researchers also propose a possible remedy.</p><p>Their STEAM method searches across 126 candidate languages to identify back-translations that recover weakened watermark signals. In their experiments, it improved detection performance by an average of 0.23 AUC and approximately 37 percentage points in true-positive rate at a 1% false-positive threshold.</p><p>Those figures are encouraging, but they are results from the systems evaluated in the paper.</p><p>They do not establish the accuracy or resilience of Claude’s deployed watermark.</p><p>The publishing implications are nevertheless immediate.</p><p>Imagine an English-language manuscript that passes through professional translation, copyediting, terminology harmonisation and local market adaptation.</p><p>A watermark that can be detected reliably in the original language might be much harder to identify in the translated edition.</p><p>That does not mean the translation was written by a human. Nor does it prove that someone deliberately removed a watermark.</p><p>It means that a negative result may reflect the limitations of the verification method rather than the provenance of the document.</p><p>In a multilingual publishing environment, that distinction is essential.</p><p>These developments bring us to a practical problem.</p><p>Suppose a manuscript arrives with no detectable watermark.</p><p>What does that establish?</p><p>Very little on its own.</p><p>The text might have been generated by a model without watermarking enabled. It might have been produced using a configuration that did not apply the mark. It might have undergone translation, paraphrasing or substantial editorial intervention.</p><p>Or it might simply have been written by a human.</p><p>The absence of a detectable signal cannot reliably distinguish those possibilities.</p><p>The reverse also requires care.</p><p>Finding a watermark does not establish who owns the work, who wrote the original manuscript, whether the text is accurate or how much human intellectual contribution was involved.</p><p>Anthropic itself acknowledges that its watermark cannot distinguish between Claude writing a text and Claude heavily editing it.</p><p>That is particularly relevant to publishing, where writing, editing, typesetting, translation and production are distinct activities.</p><p>A provenance signal is useful evidence. It is not an authorship judgment.</p><p>Earlier in this series, I considered a straightforward publishing experiment: generate watermarked text, print it, scan it, recover the text through OCR, and see whether the watermark survives.</p><p>The AI agents working on this investigation advised against proceeding.</p><p>Their objection was sound. Without access to the actual production detector, we could perform the physical workflow but not measure whether the watermark survived.</p><p>That position has not changed.</p><p>Anthropic’s Detection API remains in restricted private preview, and I have not found an independently reproducible production-Claude study that resolves the question for our publishing workflow.</p><p>We could create an attractive demonstration. We could not honestly present it as validation.</p><p>I would rather leave the experiment unfinished than publish a result that does not answer the question.</p><p>The next stage of this investigation should be practical.</p><p>I would like to see independent evaluators given sufficient access to test real Claude-generated material across different models, configurations and editorial workflows.</p><p>Not just pristine passages copied directly from a chatbot.</p><p>Actual manuscripts.</p><p>Documents subjected to normal copyediting. Translations into multiple languages. Text containing substantial human revisions. Short extracts taken from longer works. Mixed human and AI contributions.</p><p>And critically, tests that report both false positives and false negatives under clearly documented conditions.</p><p>The emerging research suggests that verification architectures can evolve. It also shows why testing only ideal conditions is insufficient.</p><p>Professional users need to understand not merely whether a watermark exists, but when its presence or absence can be interpreted responsibly.</p><p>Three developments now sit side by side.</p><p>Anthropic is extending watermark deployment to older Claude models.</p><p>Researchers are exploring ways to make verification more accessible without giving away every security mechanism.</p><p>And multilingual experiments show that detection performance can vary significantly across translation workflows.</p><p>Together, they point towards a larger issue.</p><p>Watermarking may become a standard feature of AI-generated text before independent verification becomes a standard capability for the people expected to rely on it.</p><p>That would be an uncomfortable outcome for a technology introduced in the name of transparency.</p><p>For publishers, the priority should not be to classify every sentence as human or machine.</p><p>It should be to preserve meaningful evidence, avoid unsupported accusations and understand the limitations of the tools being used.</p><p>A watermark can provide evidence of AI involvement. It cannot replace editorial judgment, documentation or accountability.</p><p>And until verification becomes more accessible and more thoroughly tested, those distinctions will matter more than ever.</p><ol><li>Anthropic — How Claude marks AI-generated content <a href="https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content">https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content</a></li><li>Anthropic — How Claude’s text watermarking works <a href="https://www.anthropic.com/news/claude-text-watermark">https://www.anthropic.com/news/claude-text-watermark</a></li><li>Georgios Milis et al. — Could LLM Watermark Detection be Public? (8 October 2026) <a href="https://arxiv.org/abs/2610.12106">https://arxiv.org/abs/2610.12106</a></li><li>Asim Mohamed and Martin Gubri — Is Multilingual LLM Watermarking Truly Multilingual? Scaling Robustness to 100+ Languages via Back-Translation, Findings of EMNLP 2026 <a href="https://aclanthology.org/2026.findings-emnlp.1115/">https://aclanthology.org/2026.findings-emnlp.1115/</a></li><li>The Decoder — Anthropic’s watermark rollout leaves a loophole on two older Claude models <a href="https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/">https://the-decoder.com/anthropics-watermark-rollout-leaves-a-loophole-on-two-older-claude-models/</a></li></ol><p>Editorial disclosure: This edition was researched and initially drafted with assistance from the AI agents assigned to AI Watermark Watch. The research distinguishes independently tested findings from vendor statements and identifies where production-Claude evidence remains unavailable. Final editorial review and publication responsibility rest with the human publisher.</p><p>Luigi Pascal Rondanini Rondanini Publishing</p><p>#AIWatermarkWatch #Anthropic #ClaudeAI #AITransparency #AIGovernance #Publishing #DigitalProvenance #ArtificialIntelligence</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/ai-watermark-watch-week-10-expanding-verification-behind-rondanini-5umhe/?published=t"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=03556679f5a7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-post-project-world/ai-watermark-watch-week-10-the-watermark-is-expanding-verification-is-falling-behind-03556679f5a7">AI Watermark Watch — Week 10 The Watermark Is Expanding. Verification Is Falling Behind.</a> was originally published in <a href="https://medium.com/the-post-project-world">The Post-Project World</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Watermark Watch — Week 9: What If the Watermark Can Be Defeated Without Knowing Its Secret?]]></title>
            <link>https://medium.com/the-post-project-world/ai-watermark-watch-week-9-what-if-the-watermark-can-be-defeated-without-knowing-its-secret-52bda0a16fff?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/52bda0a16fff</guid>
            <category><![CDATA[ai-watermarking]]></category>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[watermark]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 12:31:33 GMT</pubDate>
            <atom:updated>2026-10-24T12:31:51.457Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*lly_7VYCXLhcAWuy" /></figure><p>New independent research identifies a shared weakness across seven statistical watermarking methods, including SynthID. The implications deserve attention, but what does this actually tell us about Claude?</p><p>Research and initial draft: AI Watermark Watch monitoring agents. Human editorial review: Luigi Pascal Rondanini.</p><p>We missed a week.</p><p>Deliberately.</p><p>There were developments in AI watermarking, additional model rollouts and announcements about independent evaluation. But nothing that I felt justified another article.</p><p>I don’t particularly want AI Watermark Watch to become a weekly exercise in finding something to say simply because Friday has arrived.</p><p>The purpose of this investigation is to follow the evidence. If nothing material changes, there is nothing material to report.</p><p>This week, something changed.</p><p>On 20 October, researchers published a paper examining seven statistical text-watermarking methods, including SynthID.</p><p>Their findings raise an uncomfortable possibility.</p><p>An attacker may be able to defeat a watermark without knowing the secret key used to create it.</p><p>And the reason appears to lie in the statistical mechanism that makes watermark detection possible in the first place.</p><p>That deserves a closer look.</p><p>The paper is titled A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models.</p><p>Its authors, Romina Omidi, Nima Naderloui, Yuan Hong and Binghui Wang, examine a weakness shared by several established watermarking approaches.</p><p>Until now, much of our investigation has concentrated on what happens after a model generates watermarked text.</p><p>What happens if an editor changes the wording?</p><p>What happens after translation?</p><p>What happens when a human rewrites several paragraphs?</p><p>Can the watermark survive ordinary professional publishing?</p><p>This new research approaches the problem from another direction.</p><p>Rather than removing the watermark from an existing document, the researchers investigate whether an attacker can influence generation so that the resulting text avoids detection.</p><p>The distinction matters.</p><p>We are no longer discussing only whether a watermark survives subsequent editing.</p><p>We are asking whether the statistical mechanism can be circumvented during generation itself.</p><p>Statistical text watermarking works by influencing the choices a language model makes while generating text.</p><p>A model normally has several plausible candidates for its next token.</p><p>A watermarking mechanism changes aspects of that selection process to introduce a statistical pattern that a detector can subsequently recognise.</p><p>The new research identifies a relationship between two things:</p><ul><li>The statistical scores used to detect the watermark.</li><li>The changes watermarking introduces into token probabilities.</li></ul><p>The researchers call this relationship score-ratio alignment.</p><p>Their argument is that the very statistical structure that makes certain watermarks detectable also creates information that can be exploited to evade detection.</p><p>This is particularly interesting because the proposed attack does not require the attacker to obtain the secret watermark key or the target model’s internal weights.</p><p>It relies on black-box query access and surrogate information.</p><p>In other words, the security question extends beyond whether the key itself remains confidential.</p><p>That is a rather important distinction.</p><p>The researchers evaluated their framework against seven watermarking schemes, including SynthID, using three language models and multiple text datasets.</p><p>Across their tested configurations, they report reducing watermark detection rates to no more than 5%.</p><p>They also report preserving fluent, natural-looking text.</p><p>Those are substantial results.</p><p>However, I want to be careful about their interpretation.</p><p>The reported detection rates apply to the experimental configurations used in the paper.</p><p>They are not universal performance figures for every implementation of the seven methods.</p><p>Nor does fluency alone establish that every generated passage preserves its original meaning, factual content or editorial quality.</p><p>That distinction matters particularly for publishing.</p><p>An attack that produces convincing prose but changes important facts may defeat a detector while creating an entirely different problem.</p><p>The research is nevertheless significant because it presents a common analytical framework rather than another isolated example of a watermark being weakened through paraphrasing.</p><p>It suggests that several apparently different watermarking methods may share an exploitable statistical property.</p><p>No.</p><p>And I would be extremely cautious about anyone presenting this paper under that headline.</p><p>Anthropic has confirmed that Claude uses a version of the SynthID-Text approach.</p><p>That makes research involving SynthID relevant to our investigation.</p><p>But the researchers did not test Claude’s deployed production watermark or Anthropic’s private detector.</p><p>The paper’s SynthID experiments also do not establish the performance of Anthropic’s specific detection configuration.</p><p>We therefore have evidence of a potentially important vulnerability affecting the tested watermarking family.</p><p>We do not have evidence that Claude’s production watermark can currently be defeated with the same success rate.</p><p>There is another limitation worth recording.</p><p>This is a new research preprint. Its mathematical arguments and experimental results deserve independent examination and replication.</p><p>For now, I would classify the evidence as meaningful and technically substantial, but not yet independently validated in production.</p><p>That is sufficient to justify investigating the issue.</p><p>It is not sufficient to declare the technology broken.</p><p>Imagine that a publisher receives a manuscript and wants to understand whether AI was involved in producing it.</p><p>There are already several difficulties.</p><p>A positive watermark result may indicate that a model processed the material, but it cannot establish the author’s contribution.</p><p>A negative result may occur because the material was edited, translated, shortened or combined with human writing.</p><p>Now consider deliberate evasion.</p><p>If a watermark can be circumvented during generation, the absence of a detectable signal becomes even less useful as evidence that a document originated entirely from a human.</p><p>This does not mean publishers should assume that authors are attempting to conceal AI use.</p><p>It means that a technical detector cannot replace an appropriate editorial and provenance process.</p><p>The distinction between a signal and a verdict remains essential.</p><p>For professional publishing, I would rather have a clear record of the production process, an honest declaration of material AI involvement and documented editorial responsibility than rely exclusively on a binary detector result.</p><p>A watermark can contribute to that record.</p><p>It cannot carry the entire evidential burden.</p><p>Something else about this research interests me.</p><p>Watermarking is intended to make AI-generated material more identifiable.</p><p>Its value depends partly on how reliably the statistical signal can be detected.</p><p>But the new paper argues that, across the methods it studies, the properties that support detection also expose a route to evasion.</p><p>If subsequent independent research confirms that finding, watermark designers face a difficult engineering question.</p><p>How do you preserve reliable detection without introducing a statistical structure that can be exploited by someone deliberately trying to avoid it?</p><p>This is not a reason to abandon provenance technology.</p><p>It is a reason to test it against realistic adversarial conditions, rather than evaluating it only on untouched model output.</p><p>It also reinforces something we have been discovering throughout this series.</p><p>A technically successful watermark and a trustworthy provenance system are not necessarily the same thing.</p><p>Over the previous eight editions, our agents identified several separate research questions.</p><p>Initially, we concentrated on watermark survival after ordinary editing.</p><p>Then we encountered research suggesting that introducing watermarking could alter model behaviour, including tool selection and refusals.</p><p>In Week 8, we examined experimental evidence that some watermarking methods could introduce factual errors even when the correct information was available in the source material.</p><p>Now we have research suggesting that statistical watermarking may also expose exploitable weaknesses during generation.</p><p>These are different findings, with different experimental limitations.</p><p>They should not be combined into a sweeping claim that Claude is inaccurate, unsafe or trivially detectable.</p><p>None of those conclusions follows from the evidence we have examined.</p><p>But taken together, the studies suggest that watermarking deserves to be treated as a significant component of model deployment.</p><p>It needs robustness testing.</p><p>It needs behavioural testing.</p><p>It needs factual-accuracy testing.</p><p>It needs independent verification.</p><p>And it needs careful rules about how detection results are interpreted.</p><p>For organisations using AI in publishing, research, financial services or other professional environments, those questions are not academic details.</p><p>They affect what we can reasonably conclude from the evidence.</p><p>Our agents have recommended that we follow three developments.</p><p>The first is independent replication of this new attack framework, particularly against more recent watermarking implementations and larger production-scale models.</p><p>The second is whether researchers can establish how closely the experimental SynthID configuration resembles the systems deployed by major AI providers.</p><p>The third remains our original objective: access to Anthropic’s production detector.</p><p>That last point has not changed.</p><p>Our proposed print → scan → OCR experiment is still on hold because we cannot yet measure its intended outcome using the production verification mechanism.</p><p>There is little value in conducting an experiment simply to produce a result.</p><p>The result needs to answer the question.</p><p>I want to return briefly to the gap between Week 8 and Week 9.</p><p>The agents continued monitoring the research.</p><p>They found updates, announcements and reporting.</p><p>I reviewed the material and decided there wasn’t enough new evidence to warrant another article.</p><p>So we didn’t publish one.</p><p>I think that is an entirely reasonable outcome for an AI-assisted research process.</p><p>An agent’s usefulness should not be measured exclusively by how much content it produces.</p><p>Sometimes the appropriate recommendation is to wait.</p><p>Sometimes it is to abandon an experiment.</p><p>And sometimes it is to report that nothing sufficiently important has changed.</p><p>This week, the agents identified something that did meet our threshold.</p><p>A new technical argument.</p><p>Controlled experimental results.</p><p>A potentially important implication for statistical watermarking.</p><p>And a clear set of limitations that we can explain without pretending the researchers tested something they didn’t.</p><p>That is worth publishing.</p><p>I remain convinced that provenance technology has an important role to play in an environment where human and machine-generated material increasingly coexist.</p><p>But I am becoming less convinced that a watermark detector should ever be treated as a standalone authority on how a document was created.</p><p>The new research gives us another reason to be cautious.</p><p>A detectable statistical pattern may be useful evidence.</p><p>An absent pattern may mean very little.</p><p>And if the mechanism can be deliberately circumvented without access to its secret key, we need to understand precisely what confidence the detector can still provide.</p><p>For Claude, the production-specific answer remains unresolved.</p><p>For statistical watermarking more broadly, this week’s research has given us something genuinely new to investigate.</p><p>The agents will continue monitoring it.</p><p>I will continue checking their conclusions.</p><p>And if next week produces nothing worth reporting, we will skip that one too.</p><p>I’d rather publish nine useful investigations than ten articles written to satisfy a calendar.</p><p>Omidi, R., Naderloui, N., Hong, Y. and Wang, B. — A Unified Attack Framework against Statistical Generative Watermarks for Large Language Models.</p><p>Submitted 20 October 2026. arXiv preprint 2610.27653.</p><p><a href="https://arxiv.org/abs/2610.27653">https://arxiv.org/abs/2610.27653</a></p><p>Anthropic — How Claude’s text watermarking works</p><p><a href="https://www.anthropic.com/news/claude-text-watermark">https://www.anthropic.com/news/claude-text-watermark</a></p><p>Anthropic — How Claude marks AI-generated content</p><p><a href="https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content">https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content</a></p><p>About AI Watermark Watch</p><p>AI Watermark Watch is an ongoing independent investigation into AI text watermarking, provenance, detection and the implications for professional workflows.</p><p>The monitoring agents identify new research, compare technical findings and prepare initial reports. I review the underlying evidence, challenge their conclusions and retain editorial responsibility for publication.</p><p>In a discussion about AI provenance, readers should be able to see the provenance of the discussion itself.</p><p>Luigi Pascal Rondanini — Rondanini Publishing</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/ai-watermark-watch-week-9-what-can-defeated-without-its-rondanini-u3zge/"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=52bda0a16fff" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-post-project-world/ai-watermark-watch-week-9-what-if-the-watermark-can-be-defeated-without-knowing-its-secret-52bda0a16fff">AI Watermark Watch — Week 9: What If the Watermark Can Be Defeated Without Knowing Its Secret?</a> was originally published in <a href="https://medium.com/the-post-project-world">The Post-Project World</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your Website Has Its Own Address. Its Dependencies Are Shared.]]></title>
            <link>https://medium.com/the-post-project-world/google-cloud-outage-php-security-small-business-3d617c6be5fb?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d617c6be5fb</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[cybersecurity]]></category>
            <category><![CDATA[cloud-computing]]></category>
            <category><![CDATA[business-continuity]]></category>
            <category><![CDATA[small-business]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Fri, 23 Oct 2026 08:07:27 GMT</pubDate>
            <atom:updated>2026-10-23T08:07:27.382Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*KeJHCywuG6EE9z1ahf88tw.jpeg" /><figcaption>Behind separate websites sit shared systems — and shared points of failure. Photograph by Brett Sayles via Pexels. Stock image used for illustration.</figcaption></figure><p><em>This week’s Google Cloud disruption and PHP security emergency raise an uncomfortable question: how much of our digital business do we actually understand?</em></p><p>On 20 October, Google Cloud experienced a disruption in its London region. Google attributed the incident to a power event affecting one of its zones, with consequences for multiple services.</p><p>Separately, critical vulnerabilities in PHP prompted urgent security updates and warnings from cybersecurity authorities.</p><p>Different causes. Different responses. For someone trying to use an affected service, however, the experience could be much the same: a page that would not load, an unavailable function, a task that had to wait.</p><p>What interests me is how quickly apparently separate businesses can discover that they share a problem.</p><h3>The business we see, and the systems underneath</h3><p>As someone working across publishing and technology, I spend a fair amount of time thinking about websites as places where things happen. Readers discover a book. A customer buys something. Someone submits a manuscript or listens to a recording.</p><p>Each website has its own name, design and purpose. That makes it feel like a separate establishment.</p><p>Underneath, the boundaries are less obvious.</p><p>Hosting, databases, payment services, software libraries, authentication and email delivery can involve several organisations before the customer has completed a single transaction.</p><p>We choose some of those dependencies directly. Others arrive through the suppliers we choose.</p><p>A familiar brand on the invoice does not necessarily tell us everything about the infrastructure behind it.</p><h3>An unavailable website does not tell the whole story</h3><p>There is also an important distinction between these two events.</p><p>An infrastructure outage can make a service unavailable because something it needs has failed.</p><p>A security emergency can lead an operator to take a service offline deliberately while protection is applied. In that situation, temporary unavailability may be a responsible decision.</p><p>The Japanese service notice I received this week described disruption during PHP security work and an emergency upgrade. It was a small, practical example of a much wider technical issue arriving in an ordinary inbox.</p><p>From the outside, it is easy to put every unavailable website into the same category. But a useful response depends on understanding what happened.</p><p>Customers need that distinction explained clearly, too. They need to know what is affected, what they can still do and when they should expect another update.</p><h3>What is proportionate for a small business?</h3><p>It is tempting to respond to an outage by recommending another cloud provider, another copy of everything and automatic failover.</p><p>Sometimes that investment is justified. It also introduces costs, maintenance and more things that need testing.</p><p>For a small publisher, retailer or independent service, I would begin with a few less glamorous questions.</p><p>Who is responsible for updating the software underneath the website?</p><p>When was a backup last restored successfully?</p><p>Can customers find an update somewhere independent of the affected service?</p><p>If the website is unavailable for a day, which activities can continue manually?</p><p>And if two suppliers are supposed to provide alternatives, do they depend on the same infrastructure further down the chain?</p><p>These questions do not guarantee uninterrupted service. They help establish whether the recovery plan exists beyond a reassuring sentence in a proposal.</p><h3>Maintenance belongs in the business plan</h3><p>Building something new produces visible progress. Maintaining it rarely offers the same satisfaction.</p><p>A successful update looks like an ordinary day. A tested backup sits quietly until it is needed. A clear incident procedure may go unused for months.</p><p>That makes maintenance easy to undervalue.</p><p>This week is a reminder that the work continues after launch. Someone must understand the dependencies, apply the updates, test recovery and communicate when things go wrong.</p><p>For anyone running a digital business, that responsibility deserves a budget and an owner.</p><p>I cannot personally prevent a cloud power incident or discover every vulnerability in the software my business uses. I can ask better questions about what happens next.</p><p>If an important dependency becomes unavailable tomorrow morning, I want to know what my customers will experience — and what I can still do for them.</p><p>Sources:</p><p>Google Cloud incident report:<br><a href="https://status.cloud.google.com/incidents/Xs2XTHzAVC9YnafnPyzj">https://status.cloud.google.com/incidents/Xs2XTHzAVC9YnafnPyzj</a></p><p>NCSC: Critical security issue affecting PHP:<br><a href="https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php">https://www.ncsc.gov.uk/news/critical-security-issue-affecting-php</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d617c6be5fb" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-post-project-world/google-cloud-outage-php-security-small-business-3d617c6be5fb">Your Website Has Its Own Address. Its Dependencies Are Shared.</a> was originally published in <a href="https://medium.com/the-post-project-world">The Post-Project World</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[AI Watermark Watch — Week 8: What If the Watermark Makes the AI Less Accurate?]]></title>
            <link>https://medium.com/the-post-project-world/ai-watermark-watch-week-8-what-if-the-watermark-makes-the-ai-less-accurate-3833cc0b604a?source=rss-1a66c626d0c4------2</link>
            <guid isPermaLink="false">https://medium.com/p/3833cc0b604a</guid>
            <category><![CDATA[anthropic-claude]]></category>
            <category><![CDATA[ai-watermarking]]></category>
            <category><![CDATA[generative-ai-tools]]></category>
            <category><![CDATA[chatgpt]]></category>
            <dc:creator><![CDATA[Luigi Pascal Rondanini]]></dc:creator>
            <pubDate>Sun, 11 Oct 2026 15:56:00 GMT</pubDate>
            <atom:updated>2026-10-11T15:56:23.688Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/0*UrUs8_kbwa4Pvvd8" /></figure><p>New research raises an uncomfortable possibility: a technology designed to establish AI provenance may also introduce factual errors into the text it is supposed to protect.</p><p>This week’s investigation was researched and initially drafted by the AI agents assigned to AI Watermark Watch. I reviewed their findings, checked the underlying evidence and approved the article for publication.</p><p>For seven weeks, we have been investigating what happens to an AI watermark when somebody changes the text.</p><p>Editing, rewriting, paraphrasing, translation, mixed human and AI contributions. All the ordinary things that happen to documents once they leave the model that generated them.</p><p>In Week 6, the investigation took an unexpected turn. Researchers found that adding watermarking to language models could alter their behaviour, including decisions about tool use and refusals.</p><p>This week, our monitoring agents found another development.</p><p>And I think it deserves attention.</p><p>What if the watermark doesn’t just change how the AI writes, but also changes whether what it writes is correct?</p><p>That is the question raised by a new research paper accepted at NeurIPS 2026.</p><p>The paper has a rather appropriate title:</p><p>Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate.</p><p>Published on 4 October, the study investigates whether statistical text watermarking can introduce factual errors into language-model responses.</p><p>The researchers tested six watermarking methods, including SynthID-style approaches, across three open-weight model families.</p><p>Their experimental design is particularly interesting.</p><p>They compared watermarked and unwatermarked responses using identical questions, supporting documents and generation settings.</p><p>In other words, the correct information was available to the model.</p><p>The researchers were not simply asking whether a language model might hallucinate when it didn’t know the answer.</p><p>They were investigating whether introducing a watermark could cause the model to produce an incorrect answer even when the necessary information was already present.</p><p>And they found that it could.</p><p>Dates changed.</p><p>Numbers changed.</p><p>Identifiers became incorrect.</p><p>The responses could remain fluent, plausible and professionally presented.</p><p>But some of the facts were wrong.</p><p>That is a different problem from whether a detector can recognise AI-generated text.</p><p>It concerns the reliability of the generated information itself.</p><p>One experimental result is particularly worth examining.</p><p>Using the KGW watermarking method, researchers observed a 12.9-percentage-point net reduction in factual accuracy.</p><p>That is not a rounding error.</p><p>The researchers also investigated safeguards designed to reduce the factual degradation.</p><p>Their combined approach reduced the penalty to approximately 0.9 percentage points, while maintaining the same measured watermark-detection performance.</p><p>There are two important observations here.</p><p>First, watermarking can introduce measurable factual errors under controlled experimental conditions.</p><p>Second, the problem may be mitigated through better technical design.</p><p>This is not an argument that watermarking should be abandoned.</p><p>It is an argument that watermarking needs to be evaluated against more than detectability and general text quality.</p><p>A response can read beautifully and still contain the wrong number.</p><p>For professional users, that distinction matters.</p><p>Consider a manuscript containing a factual reference.</p><p>An author provides the correct date.</p><p>The source document contains the correct date.</p><p>The language model is asked to produce a concise passage based on that material.</p><p>The result is coherent, well written and appropriately structured.</p><p>But the date has changed.</p><p>An editor reading the paragraph might not notice.</p><p>A proofreader concentrating on punctuation and consistency might not notice either.</p><p>The incorrect detail can then travel through typesetting, PDF production, printing and distribution.</p><p>Nothing about the presentation necessarily indicates that something went wrong.</p><p>Now consider the same problem in a financial report, legal summary, research paper or regulatory submission.</p><p>A date, identifier, monetary amount or percentage is not merely a stylistic choice.</p><p>It can materially change the meaning of the document.</p><p>This is why I find the new research more interesting than another announcement about which models have started using watermarks.</p><p>It introduces a question about the integrity of the information being produced.</p><p>We need to be precise about what the research establishes.</p><p>The researchers did not test Anthropic’s production Claude watermark.</p><p>They tested watermarking methods under controlled experimental conditions using open-weight models.</p><p>Their findings demonstrate that watermarking can introduce factual degradation.</p><p>They do not demonstrate that Anthropic’s deployed implementation produces the same effects.</p><p>That distinction is essential.</p><p>Anthropic’s system uses technology from the SynthID-Text family, which makes research into related approaches relevant to our investigation.</p><p>But belonging to the same technological family does not mean two implementations have identical configurations, thresholds, safeguards or behavioural characteristics.</p><p>We should not take an experimental result from one implementation and automatically assign it to another.</p><p>What the study gives us is a credible reason to ask Anthropic a more specific question:</p><p>Has Claude’s watermarked and unwatermarked factual accuracy been compared using identical source-grounded workloads, particularly for dates, numbers and identifiers?</p><p>And if those evaluations have been conducted, what were the results?</p><p>When AI Watermark Watch started, our proposed evaluation concentrated on whether the watermark survived transformations.</p><p>We wanted to know what happened after proofreading, copy-editing, translation, rewriting and mixed human/AI contributions.</p><p>We even considered printing a document, scanning it and reconstructing the text through OCR.</p><p>Our agents correctly recommended that we should not run that experiment without access to the production detector.</p><p>We could measure the transformations, but not the watermark.</p><p>That experiment remains on hold.</p><p>However, the research we have examined over the past few weeks suggests that the evaluation needs another dimension.</p><p>It is no longer sufficient to ask:</p><p>Does the watermark survive?</p><p>We also need to ask:</p><p>Did introducing the watermark change the answer?</p><p>A proper evaluation should compare the same tasks with and without watermarking, then examine both detection performance and the resulting output.</p><p>For a publisher, I would be particularly interested in factual preservation.</p><p>Names, dates, quotations, references, ISBNs, numerical claims and source attribution.</p><p>For an organisation deploying autonomous agents, I would extend that testing to tool selection, arguments and actions.</p><p>These are not necessarily the same risks, but they originate from a related concern.</p><p>Changing token selection can have consequences beyond changing the appearance of a sentence.</p><p>In Week 6, we discussed research from Lasso Security examining the behavioural effects of SynthID-Text watermarking.</p><p>Their experiments suggested that watermarking could alter model decisions, including refusals and tool calls.</p><p>Again, that was not a test of Anthropic’s production system.</p><p>Now we have a separate investigation showing that watermarking can also affect factual accuracy.</p><p>The two studies examine different outcomes.</p><p>But together they suggest a broader principle.</p><p>A watermark should be treated as part of the model’s operational configuration, not merely as an invisible label added after the answer has been produced.</p><p>If an organisation validates a model and subsequently introduces or changes watermarking, it should consider whether its previous evaluation remains representative.</p><p>That seems particularly relevant in regulated environments.</p><p>We would not normally assume that a material change to a production system requires no regression testing simply because its purpose is beneficial.</p><p>Why should AI watermarking be different?</p><p>I don’t want this article to become another argument that AI watermarking is fundamentally unreliable.</p><p>That would not reflect the research fairly.</p><p>One of the more encouraging findings is that the researchers developed safeguards that substantially reduced the observed factual-accuracy penalty while preserving measured detection performance.</p><p>That matters.</p><p>It suggests there may be ways to improve provenance without accepting the same level of factual degradation.</p><p>But it also reinforces why independent testing is valuable.</p><p>If researchers had examined only whether the watermark was detectable, they might have missed the factual errors.</p><p>If they had examined only whether the text sounded natural, they might also have missed them.</p><p>The evaluation had to ask the right question.</p><p>And that is becoming a recurring theme in this investigation.</p><p>We now have several distinct questions that should not be confused.</p><p>Can a watermark be detected in the original generated text?</p><p>Can it survive editing, rewriting, translation and mixed authorship?</p><p>Can a detector produce reliable evidence without being mistaken for an authorship test?</p><p>Can introducing the watermark alter the model’s behaviour?</p><p>And now:</p><p>Can watermarking make an otherwise correct response factually incorrect?</p><p>The new study provides meaningful experimental evidence for that final concern.</p><p>It does not establish that Claude’s production system suffers from the same problem.</p><p>Nor does it resolve our original publishing-workflow questions.</p><p>Anthropic’s production detector remains in private preview, and we still lack the independent end-to-end experiment we have been waiting for.</p><p>I would rather acknowledge those limitations than manufacture certainty.</p><p>There is also an important lesson for professional users.</p><p>A watermark may help establish that AI was involved in producing a document.</p><p>But that provenance signal does not guarantee that the document is accurate.</p><p>Those are separate responsibilities.</p><p>One concerns where the text came from.</p><p>The other concerns whether we can trust what it says.</p><p>For publishers, researchers, financial institutions and anyone producing material that others may rely upon, we need both.</p><p>The agents will continue monitoring the research.</p><p>I will continue checking their findings.</p><p>And after this week, I have added another question to our investigation:</p><p>Before asking whether the watermark survived the editing process, perhaps we should establish whether the facts survived the watermark.</p><p>Invisible Ink, Visible Lies: How Production Watermarking Causes LLMs to Hallucinate Research paper, 4 October 2026. Accepted at NeurIPS 2026.</p><p>Lasso Security — The Provenance Tax: Understanding the Impact of LLM Watermarking on AI Agent Behavior</p><p><a href="https://www.lasso.security/blog/the-provenance-tax-understanding-the-impact-of-llm-watermarking-on-ai-agent-behavior">https://www.lasso.security/blog/the-provenance-tax-understanding-the-impact-of-llm-watermarking-on-ai-agent-behavior</a></p><p>Anthropic — Claude text watermarking</p><p><a href="https://www.anthropic.com/news/claude-text-watermark">https://www.anthropic.com/news/claude-text-watermark</a></p><p>Anthropic — How Claude marks AI-generated content</p><p><a href="https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content">https://support.claude.com/en/articles/16266773-how-claude-marks-ai-generated-content</a></p><p>About AI Watermark Watch</p><p>AI Watermark Watch is an ongoing independent investigation into AI text watermarking, provenance, detection and the implications for professional workflows.</p><p>The monitoring agents identify new research, compare findings and prepare the initial reports. I review the underlying evidence, challenge the conclusions and retain editorial responsibility for what is published.</p><p>In a discussion about AI provenance, I think readers should be able to see the provenance of the discussion itself.</p><p>Luigi Pascal Rondanini Rondanini Publishing</p><p><em>Originally published at </em><a href="https://www.linkedin.com/pulse/ai-watermark-watch-week-8-what-makes-less-accurate-luigi-rondanini-mcu2e/"><em>https://www.linkedin.com</em></a><em>.</em></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3833cc0b604a" width="1" height="1" alt=""><hr><p><a href="https://medium.com/the-post-project-world/ai-watermark-watch-week-8-what-if-the-watermark-makes-the-ai-less-accurate-3833cc0b604a">AI Watermark Watch — Week 8: What If the Watermark Makes the AI Less Accurate?</a> was originally published in <a href="https://medium.com/the-post-project-world">The Post-Project World</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
    </channel>
</rss>