<?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 Phil Gold on Medium]]></title>
        <description><![CDATA[Stories by Phil Gold on Medium]]></description>
        <link>https://medium.com/@philgold?source=rss-732fed96d8f6------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/0*eJp_vdo1LmYqdzJy.</url>
            <title>Stories by Phil Gold on Medium</title>
            <link>https://medium.com/@philgold?source=rss-732fed96d8f6------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 02:41:00 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@philgold/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 the Second Opinion Starts Making Things Up]]></title>
            <link>https://medium.com/@philgold/when-the-second-opinion-starts-making-things-up-a820a3742b92?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/a820a3742b92</guid>
            <category><![CDATA[trust]]></category>
            <category><![CDATA[ai-hallucination]]></category>
            <category><![CDATA[chatgpt]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[second-order-effect]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Sat, 24 Oct 2026 13:08:21 GMT</pubDate>
            <atom:updated>2026-10-24T20:35:59.706Z</atom:updated>
            <content:encoded><![CDATA[<h4><em>What happens to trust when your AI research assistant invents the evidence?</em></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*mI6j4J8DrBveyZzL8Jw9Og.png" /></figure><p>I’ve been using ChatGPT for a while now, and until recently, I was pretty happy with the relationship.</p><p>I use it for all sorts of things. Sometimes it’s a writing partner, helping me organize ideas, refine arguments, or find a better way to say something. Sometimes it’s a research assistant, helping me investigate subjects I know relatively little about. And sometimes it’s a second opinion, helping me interpret documents, evaluate options, or figure out what to do next.</p><p>That last category is particularly important because some of the things I’m working on have real consequences. For example, I’ve been dealing with the complicated process of closing a company in Portugal, including tax filings, vehicle registration issues, and correspondence with accountants and government agencies. I’ve also been navigating a dispute with my landlord in Spain over mold, air-conditioning maintenance, and responsibility for repairs.</p><p>On the professional side, I’m pursuing music licensing and traditional publishing opportunities, both of which involve researching organizations, evaluating submission requirements, and deciding where to invest time and money.</p><p>And then there’s a nonfiction book project I’m researching about a centuries-old crypto-Jewish community in Portugal. That involves historical documents, archival records, and attempting to establish connections between people who lived more than a hundred years ago.</p><p>In all these situations, ChatGPT has been enormously useful. It can help me understand unfamiliar terminology, evaluate competing interpretations, organize complicated information, and prepare correspondence. It can also remember enough about an ongoing project that I don’t have to explain the entire situation every time I have a question.</p><p>I know perfectly well that AI can make mistakes. I’ve been using computers for decades, and I’ve spent much of my professional life working around technology, so I’m not inclined to treat a computer’s output as divine revelation. But experience matters. When a tool repeatedly gives you useful answers, you naturally become more comfortable relying on it.</p><p>And then one day last week, things got strange.</p><h4><strong>The Day the Evidence Disappeared</strong></h4><p>I was working with ChatGPT on a few unrelated projects when I began noticing something weird. It was describing information in screenshots, documents, and websites that simply wasn’t there.</p><p>Not just misinterpreting ambiguous language or drawing questionable conclusions. It was confidently reporting things that didn’t exist in the material I had provided.</p><p>One particularly clear example involved YouTube analytics. I’d uploaded a screenshot showing the early performance of a new YouTube Short. ChatGPT responded with an enthusiastic analysis that included performance information I couldn’t see.</p><p>I looked at the screenshot again. There were two views, which was all the analytics information available. So, I asked a fairly obvious question: Where are you seeing this information?</p><p>Answer: It wasn’t there.</p><p>That was disturbing enough, but it wasn’t an isolated incident that day. Similar things happened in other conversations, including ones involving documents and websites where I was relying on ChatGPT to tell me what the source said.</p><p>The problem wasn’t simply that ChatGPT made mistakes. Everyone makes mistakes, including me. The problem was that it presented invented information with the same confidence it used when describing something accurately.</p><p>There was no warning light, no hesitation, and nothing to distinguish a verified observation from something the system had apparently manufactured.</p><p>This changes things considerably.</p><h4><strong>The Problem With a Convincing Answer</strong></h4><p>AI hallucinations aren’t exactly breaking news. Large language models can generate plausible statements that aren’t supported by facts, and anyone using them for serious work should understand that possibility.</p><p>But there’s a difference between knowing something can happen and experiencing it repeatedly in situations where you thought you had provided the evidence needed to prevent it. If I ask ChatGPT a general question about Portuguese tax law, I understand that the answer might be incomplete, outdated, or incorrect. I can check the relevant government website or ask my accountant.</p><p>But if I provide an actual letter from the Portuguese tax authority and ask what it says, I’m operating under a different assumption. I’m not asking the system to invent an answer based on general knowledge. I’m asking it to interpret a specific document.</p><p>I expect the interpretation might be debatable. Legal and administrative language often is. But I don’t expect the system to invent passages or details that aren’t in the document.</p><p>That’s the distinction that bothers me.</p><p>When I use AI as a second opinion, I’m often trying to reduce uncertainty. I want another perspective on the information in front of me, particularly when it’s written in a language I’m still learning or concerns a subject outside my expertise. But what happens when the second opinion introduces information that wasn’t there in the first place?</p><p>Now I have two problems instead of one. I still need to understand the original document, and I also need to determine whether the assistant’s interpretation is based on anything real.</p><h4><strong>Did Something Change?</strong></h4><p>Naturally, I wanted to know why this was happening. The errors arrived in a cluster. We’d been working together successfully for months, and suddenly there were multiple dramatic failures across several conversations.</p><p>Had something changed in the model? Was a newer version less reliable at interpreting screenshots and documents? Was this simply an unfortunate run of errors that I happened to notice because they were unusually obvious?</p><p>So, I asked ChatGPT itself for the reason. It offered possible explanations, but the uncomfortable reality is that I don’t have enough information to know what caused the change. And ChatGPT’s explanations of its own behavior aren’t necessarily reliable evidence of what happened internally.</p><p>Since then, the model I’ve been using has changed again, and I haven’t spotted any other episodes. But it hasn’t been very long, and a short time without an obvious error doesn’t firmly establish that anything has been fixed.</p><p>This is one of the stranger aspects of working with AI. The system can change without my having any meaningful understanding of what changed, why it changed, or how those changes affect the tasks I depend on it for.</p><p>With conventional software, an update might introduce a bug, but the program generally continues to perform a defined set of operations. With generative AI, the output itself is probabilistic. The same question can produce different answers, and a system that performs well on one task may fail unexpectedly on another.</p><p>I don’t have access to the information needed to diagnose what happened. So, I’m left with a more practical question — what do I do now?</p><h4><strong>The Cost of Checking the Checker</strong></h4><p>One of the great attractions of AI is that it saves time. I can upload a document, ask for an explanation, discuss possible interpretations, and prepare a response in a fraction of the time it might take me to do everything independently.</p><p>For my historical research, I can use it to organize leads, understand unfamiliar archival terminology, and identify questions worth investigating. For my publishing and music projects, it can help me evaluate opportunities and prepare submissions.</p><p>The value isn’t just speed. It’s also the ability to work on projects that might otherwise require expertise I don’t have or assistance can’t easily obtain.</p><p>But there’s a catch. If I have to independently verify every statement ChatGPT makes, how much of that advantage disappears? Suppose I upload a three-page document and ask for a summary. If I then have to reread all three pages carefully to establish whether the summary is accurate, I haven’t saved much time.</p><p>Worse, if the document is in Portuguese or concerns a technical subject I don’t completely understand, I may not have the ability to recognize every error in the first place. That negates one of the reasons I asked for help.</p><p>Of course, the answer can’t be to abandon verification entirely. The consequences of misunderstanding a tax notice or drawing the wrong conclusion from an archival document are very different from choosing an unfortunate adjective in a LinkedIn post.</p><p>And I don’t want to stop using AI. Despite the recent problems, it’s still one of the most useful tools I’ve encountered in years.</p><p>So perhaps the question isn’t whether I can trust it. Perhaps the question is what kind of trust is appropriate, and how much verification different situations require.</p><h4><strong>Changing the Way I Ask</strong></h4><p>After the hallucination episodes, ChatGPT and I had a rather serious conversation about how we should work together.</p><p>We established some new rules. When interpreting a document, screenshot, or website, it should distinguish between information actually present in the source and conclusions drawn from that information. It should identify the specific evidence supporting factual claims and acknowledge when something can’t be verified. If it doesn’t know, it should say so.</p><p>And if I challenge a statement, it should examine the evidence rather than invent an explanation for why its original answer might somehow have been correct.</p><p>None of that sounds particularly revolutionary. In fact, it sounds remarkably like the standards we’d expect from a competent human research assistant.</p><p>But there’s an obvious problem. These are instructions given to the same system that was making things up. ChatGPT readily agreed to follow them. It even produced a reassuring explanation of how it would avoid repeating the problem.</p><p>Unfortunately, a promise from an AI assistant isn’t quite the same thing as evidence that its behavior has changed.</p><p>I’ve encountered situations before where ChatGPT agreed to follow an instruction and then failed to do so. Not necessarily immediately, and not necessarily every time, but often enough that I know agreement doesn’t guarantee compliance. So, the new rules are an experiment, not a solution.</p><p>I’m beginning to think the more useful approach may be to build verification directly into the questions I ask. Instead of simply asking, “What does this document mean?” I might ask, “Identify the exact passages relevant to this question, explain what they establish, and distinguish that from anything you’re inferring.”</p><p>Instead of asking whether a historical connection is plausible, I might ask which facts are documented, which connections are speculative, and what additional evidence would be needed to establish them.</p><p>And instead of asking for a confident recommendation, I might ask what assumptions the recommendation depends on and what information could change the conclusion.</p><p>This won’t prevent every error. A system capable of inventing a fact may also invent a citation or misrepresent the evidence supposedly supporting it. But perhaps it will make some errors easier to recognize before they become part of my decision-making.</p><p>I don’t know yet.</p><h4><strong>Trust Isn’t All or Nothing</strong></h4><p>One thing I don’t want to do is overreact. A bad experience with AI doesn’t erase the genuinely useful work we’ve accomplished together. Nor does it mean every subsequent answer should be treated as suspicious until independently proven correct. That would be exhausting, and it would defeat much of the purpose of using the tool.</p><p>But it does suggest that I need to be more deliberate about where I place my confidence.</p><p>When I’m brainstorming an article, exploring a creative idea, or asking for suggestions about structure and wording, factual precision may be relatively unimportant. I can evaluate the result using my own judgment.</p><p>When I’m interpreting a tax document, evaluating a legal responsibility, or investigating a historical claim, the standard needs to be different.</p><p>And there are intermediate cases. A music licensing opportunity might not involve legal consequences, but inaccurate submission requirements could waste hours of preparation. A mistaken publishing contact could send a manuscript to the wrong person or, worse, lead me into a scam.</p><p>Perhaps the answer is to develop different working habits for different levels of risk. I don’t need to check every suggestion. I do need to know which conclusions are supported by evidence, which are reasonable interpretations, and which are simply possibilities worth investigating.</p><p>That sounds manageable in principle. Whether it works in practice remains to be seen.</p><h4><strong>The Second-Order Effect of Losing Confidence</strong></h4><p>There’s an irony in all this that I find particularly amusing. I’ve been writing a series of articles about second-order effects, the changes that happen because people adapt to the direct consequences of technology. My basic premise is that technology starts the change, but people decide what happens next.</p><p>I’ve become an example of my own argument.</p><p>The first-order effect of generative AI is that I can obtain useful analysis, explanations, and creative assistance quickly and conveniently. One second-order effect is that I’ve begun incorporating it into activities where I previously would have relied on other resources or spent considerably more time working independently.</p><p>But now there’s another adaptation underway. Because the technology has demonstrated that it can confidently invent evidence, I’m changing how I interact with it. I’m becoming more cautious about certain kinds of answers, asking different questions (and questions differently), and reconsidering how much independent verification is necessary.</p><p>The technology hasn’t simply changed how I work; its failures are changing how I work with the technology.</p><p>And there’s a further complication: I’m developing these new practices partly through conversations with the very system whose reliability I’m questioning. In fact, ChatGPT helped me organize this article.</p><p>I find that simultaneously useful and slightly absurd.</p><h4><strong>Where Does That Leave Us?</strong></h4><p>I don’t have a satisfying conclusion yet.</p><p>I don’t know whether the recent hallucinations were caused by a model change, an unfortunate coincidence, or something else entirely. I don’t know whether the new rules I’ve established with ChatGPT will make a meaningful difference, or whether I’ll eventually find myself having the same conversation again.</p><p>I also don’t know how much extra time verification will require, or whether I’ll develop a way of working that preserves most of the benefits without exposing me to unacceptable risks.</p><p>Those are things I’m going to have to discover through experience.</p><p>What I do know is that something has changed. Until recently, I was comfortable treating ChatGPT as a generally reliable assistant whose occasional mistakes I could recognize and correct.</p><p>Now I’m less certain about my ability to recognize those mistakes, particularly when the information is unfamiliar or the evidence is difficult for me to evaluate independently.</p><p>That doesn’t mean I’m going to stop using it, it means I’m going to have to learn how to use it differently.</p><p>And perhaps that’s the most interesting second-order effect of all. The problem isn’t that I discovered AI can make mistakes, since I already knew that.</p><p><strong>The problem is that I discovered I wasn’t always able to tell when it was making them.</strong></p><p>For a technology that’s supposed to help us make better-informed decisions, that’s a rather important distinction.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=a820a3742b92" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Measure Becomes the Goal]]></title>
            <link>https://medium.com/@philgold/when-the-measure-becomes-the-goal-251837fcad8e?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/251837fcad8e</guid>
            <category><![CDATA[second-order-effect]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[goodharts-law]]></category>
            <category><![CDATA[incentives]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Fri, 02 Oct 2026 09:40:07 GMT</pubDate>
            <atom:updated>2026-10-02T09:40:07.082Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>What happens when people start optimizing for the numbers meant to evaluate them?</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/600/1*04gPIQj6PUmyId7gEjpLlA.jpeg" /></figure><p>Years ago, I worked for a company that gave a monthly award to the technical support person who handled the most calls in the least amount of time.</p><p>One technician kept winning, and the rest of us had a pretty good idea why. This guy answered virtually every support call the same way: “Reboot and try again.” Then he hung up. It took us about two months to convince management that their “most efficient technician” wasn’t necessarily their best technician.</p><p>The numbers weren’t wrong; he really was handling more calls in less time than anyone else. The problem was that management thought it was measuring one thing, effective technical support, when it was actually measuring something else: how quickly someone could get a caller off the phone.</p><p>And once employees knew what was being measured and rewarded, the measurement itself became part of the system.</p><h4><strong>When a Measure Becomes a Target</strong></h4><p>Economist <a href="https://en.wikipedia.org/wiki/Goodhart%27s_law">Charles Goodhart</a> identified a version of this problem while studying British monetary policy in the 1970s. He originally put it rather formally:</p><blockquote>“Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.”</blockquote><p>Anthropologist Marilyn Strathern later distilled the idea into the much more memorable formulation now commonly known as Goodhart’s Law:</p><blockquote><strong>“When a measure becomes a target, it ceases to be a good measure.”</strong></blockquote><p>My former employer wanted good technical support. Good technical support, however, is complicated. Did we solve the customer’s problem? Was the solution correct? Did we explain it clearly so that the customer wouldn’t need to call again? Did we leave the customer feeling someone had actually listened?</p><p>Those things are difficult to put into a spreadsheet, but calls handled and average call time are easy, so management measured those instead. But measurement doesn’t simply observe behavior. Once people know that a measurement has consequences, they adapt their behavior to it.</p><h4><strong>Teaching to the Test</strong></h4><p>Education provides one of the most familiar examples. We want students to learn, but “learning” is an extraordinarily broad thing to measure. So, we create tests. Tests give us numbers, and numbers let us compare students, classrooms, schools, and sometimes entire educational systems.</p><p>There’s nothing inherently wrong with that. A well-designed test can tell us useful things about what students know and where they’re struggling.</p><p>But then, we attach consequences to the results. Schools are evaluated according to test scores. Teachers are judged by them. Funding, advancement, or other opportunities may depend on them. Inevitably, people respond.</p><p>Teaching begins to focus on the test. Students learn how to answer the kinds of questions the test will ask. Classroom time shifts toward the subjects and skills that tests will measure. And perhaps most importantly, material that isn’t going to be tested takes a lower priority than material that will.</p><p>The test hasn’t simply changed <strong>how</strong> students learn. It begins to influence <strong>what</strong> they learn.</p><p>There are more extreme examples. Researchers studying a school accountability program in Chile found that some schools responded to high-stakes testing by <strong>keeping lower-performing students from taking the tests</strong>. Students likely to reduce a school’s score were significantly more likely to be absent on test day, and the researchers concluded that schools were strategically excluding them to improve their measured performance.</p><p>The score improved, but the students didn’t necessarily learn anything more.</p><h4><strong>Hitting the Target</strong></h4><p>The consequences can become considerably more serious when the thing being measured is health care.</p><p>When Britain’s National Health Service introduced targets intended to reduce waiting times, the goal was perfectly reasonable. Nobody wants patients to wait unnecessarily for medical care, and waiting time is something that can be measured.</p><p>But not every kind of waiting was included in the same targets. Researchers found cases in which resources were concentrated on the patients whose waiting times were being measured at the expense of patients whose waiting wasn’t. In one particularly disturbing case involving ophthalmology services in Bristol, follow-up appointments were delayed or cancelled while the organization concentrated on meeting targets for new patients. At least 25 patients were estimated to have lost vision over two years as a consequence of delayed follow-up care.</p><p>The target was met while something much more important got worse.</p><p>That doesn’t mean the NHS shouldn’t measure waiting times. In fact, other research has found that waiting-time targets produced genuine improvements. Nor does it mean schools shouldn’t test students or companies shouldn’t measure employee performance.</p><p>We need measurements. Without them, we’re left with anecdotes, intuition and guesswork.</p><p>The problem begins when we forget that the measurement is usually a <strong>proxy for something we actually care about</strong>.</p><h4><strong>The Number and the Thing</strong></h4><p>A test score isn’t learning. Average call time isn’t good technical support. Waiting time isn’t health care.</p><p>They’re measurements we use because the underlying things are complicated.</p><p>Organizations are particularly vulnerable to this because complicated things are difficult to manage at scale. “Provide excellent customer service,” is a sensible objective, but someone eventually wants to know whether we’re doing it. Once that happens, we need a number.</p><p>So, we count calls. We measure response times. We collect satisfaction scores. We track sales, productivity, utilization, attendance, clicks, impressions, and dozens of other things that can be placed neatly into a dashboard. Dashboards have an extraordinary ability to make messy reality look reassuringly precise.</p><p>The trouble starts when improving the number quietly replaces improving the thing the number was supposed to represent.</p><p>This doesn’t even require anyone to cheat. The call center tech wasn’t falsifying his call statistics. An employee who learns that management values the number of tickets closed is entirely rational finding ways to close more tickets. The schools that concentrate instruction on tested material aren’t necessarily doing anything dishonest.</p><p>People respond to incentives; if you tell them what counts, they will quite reasonably try to produce more of it.</p><h4><strong>When Measurement Changes Behavior</strong></h4><p>This is where Goodhart’s Law becomes a second-order effect.</p><p>The first-order effect of measurement is information. We measure something because we want to understand what’s happening.</p><p>Then people discover that the measurement matters, and adapt.</p><p>Now you’re no longer simply measuring the original behavior. You’re measuring behavior that has partly developed in response to the measurement itself.</p><p>We see versions of this every day on social media. Likes, views, followers and impressions began as ways of measuring how people responded to content. Once those numbers became visible and consequential, creators began changing what they produced in order to generate them.</p><p>I recently spent several weeks doing exactly that on LinkedIn, trying different posting days, lengths, images and links because I wanted to understand why one post received more than 8,000 impressions while another received fewer than 200.</p><p>LinkedIn never told me to do any of those things. The numbers did. The metric wasn’t simply reporting my behavior anymore, it was helping shape it.</p><h4><strong>What Are We Actually Trying to Improve?</strong></h4><p>The answer isn’t to stop measuring things. That would solve the problem the same way throwing away the bathroom scale solves weight gain.</p><p>The more useful question is what happens <strong>after</strong> we create the measurement. If a call center rewards short calls, what behavior does that encourage? If a school is judged primarily by standardized test results, what happens to subjects that aren’t tested? If a hospital is measured on one category of waiting time, what happens to patients whose waiting time isn’t included?</p><p>And if the number improves, have we actually improved the thing we cared about?</p><p>Technology makes measurement easier than ever. Organizations can collect enormous amounts of data, build dashboards in real time and track behaviors that once would have been invisible. That’s progress, but it also means we’re creating more measurements for people to respond to.</p><p>The moment we attach consequences to a measurement, we’re no longer simply observing behavior. We’re changing the environment in which that behavior occurs.</p><p>People notice what counts, then they adapt.</p><p>The technician who kept saying “Reboot and try again” understood that perfectly.</p><p>Management just took a little longer to catch up.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=251837fcad8e" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Putting the Friction Back]]></title>
            <link>https://medium.com/@philgold/putting-the-friction-back-133ae22c5d16?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/133ae22c5d16</guid>
            <category><![CDATA[second-order-effect]]></category>
            <category><![CDATA[digital-media]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[consumer-behavior]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Mon, 21 Sep 2026 13:31:01 GMT</pubDate>
            <atom:updated>2026-09-21T13:31:01.696Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>What we discover when technology makes everything easier</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*q6fLfLaSzmkt6s_kgZVGgg.jpeg" /><figcaption>A small portion of my pre-digital library</figcaption></figure><p>When I moved homes in 2007, I had 47 cartons of fiction books.</p><p>Not 47 cartons of books. Forty-seven cartons of <em>fiction</em>. There were another 20-something cartons of nonfiction. And at least half the books were hard-bound.</p><p>I also had about 900 CDs, 50 vinyl albums and three drawers full of movies in two different formats, VHS and Blu-ray. Apparently, I didn’t so much move houses as relocate a small branch of the public library.</p><p>Today, most of that fits comfortably on a device I carry in a shoulder bag.</p><p>My books are almost exclusively electronic. Music that I don’t own electronically is available through streaming services. The movies I watch are mostly streamed as well. I can travel with a library considerably larger than the one that once required 67-plus cartons, and it’s considerably easier to lift.</p><p>This is progress, and I have absolutely no desire to move those boxes again. But something else happened when all that physical stuff disappeared.</p><h4><strong>The Things We Used to Own</strong></h4><p>For most of my life, acquiring music, books, and movies meant acquiring objects. If I wanted an album, I bought a record or CD. If I wanted a book, I bought a physical copy (used when available; if not, new). If I wanted to be able to watch a movie whenever I chose, I bought it on VHS and later, when technology thoughtfully made my VHS collection obsolete, I bought some of them a second time on Blu-ray.</p><p>Ownership came with inconvenience. Books required shelves, CDs required storage, and movies filled drawers. Everything accumulated dust. When you moved, your possessions presented you with the bill.</p><p>Digital technology removed a lot of that friction. An eBook weighs nothing. Streaming music requires no shelf space. Thousands of photographs don’t require albums or boxes. Movies no longer need three drawers in the living room.</p><p>But we didn’t simply replace physical objects with more convenient versions of the same things. In many cases, we replaced <strong>ownership with access</strong>.</p><p>And those aren’t quite the same thing.</p><h4><strong>The Library That Weighs Nothing</strong></h4><p>Books are probably the easiest example for me because I genuinely prefer much of what digitization has given us. I carry hundreds of books on a single device. If I hear about something interesting, I can often have it within minutes. I don’t have to decide which books deserve increasingly scarce shelf space.</p><p>And digital books haven’t eliminated browsing or made purchasing a requirement. Apps such as Libby let me browse library collections to discover books and borrow them without leaving home. In some ways, access has become much better.</p><p>But not every book has been digitized, and not every digitized book is available in a library. I’ve recently been doing some research that led me to an older book I wanted to consult. Other editions were available electronically, but the particular edition I needed wasn’t. Finding it meant going back into the physical world of libraries, booksellers and copies that actually had to exist somewhere.</p><p>Suddenly, the friction is back. The digital world feels almost limitless until the thing you want is outside it.</p><h4><strong>When Access Isn’t Ownership</strong></h4><p>Movies make the distinction even clearer. If I owned a movie on VHS or Blu-ray, I could watch it whenever I wanted (assuming I still owned something capable of playing it).</p><p>Streaming appears to offer something better. Instead of owning a few hundred movies, I have access to enormous catalogs containing thousands of them.</p><p>Except I have no interest in most of it, and I don’t have access to <em>Young Frankenstein</em>. At least, not conveniently in Spain.</p><p>Streaming rights vary by country. Movies move between services. Something available today may disappear next month. The catalog may be enormous, but I don’t control what’s in it.</p><p>I am now in 2026 seriously considering buying a Blu-ray player so I can buy and watch a copy of a movie released in 1974. This is not quite how I imagined technological progress working.</p><p>The physical copy was inconvenient, but its availability was under my control. Streaming removed the inconvenience while introducing a new dependency: someone else decides what I can access.</p><p>That’s not necessarily a bad trade. Most of the time, I’ll happily take the enormous streaming catalog over three drawers of plastic boxes. But it is a trade.</p><h4><strong>Abundance Changes the Problem</strong></h4><p>Music went through an even more dramatic transformation.</p><p>Owning 900 CDs once seemed like an enormous music collection, but it was still finite. Every one of those CDs represented a decision. I had heard something, become interested enough to buy it, and added it to the collection. When I wanted to listen to music, I browsed what I owned and chose something.</p><p>Streaming changed the problem.</p><p>I now have access to vastly more music than I could possibly have purchased. The constraint is no longer, “<em>What music do I own?”</em> It’s, “<em>What, among almost everything, do I want to hear?”</em></p><p>That’s an extraordinary improvement. It’s also a completely different experience.</p><p>Technology removed the friction involved in acquiring and storing music, and our behavior changed around the abundance that followed. We make playlists. We follow recommendations. We let services choose what comes next. We can sample an artist without committing anything more than three minutes of attention.</p><p>Then something curious happened. People started buying records again.</p><h4><strong>Putting the Friction Back</strong></h4><p>From a purely technological standpoint, vinyl records are ridiculous. They’re large. They’re fragile. They take up space. They require dedicated equipment. You have to get up and turn them over.</p><p>They even degrade when you use them. In fact, the whole thing works because of friction. A stylus physically travels through a groove in the record.</p><p>We spent decades developing technologies that removed friction from listening to music, and then people started buying a format that literally puts it back.</p><p>I don’t think that means millions of people have suddenly concluded that technological progress was a mistake. Nor does it necessarily mean vinyl is objectively better than digital music. It may simply mean that some of what technology removed wasn’t actually part of the problem.</p><p>Choosing a record, taking it from its sleeve, looking at the artwork, putting it on the turntable and listening to a side is less efficient than telling a phone to play a song, but perhaps efficiency isn’t always the point.</p><h4><strong>What Friction Was Doing</strong></h4><p>We think about friction as something technology should eliminate, and much of the time, it should. I don’t miss hauling cartons of books around. I don’t want to return to a world where hearing a new album required finding a record store that happened to stock it, and I certainly don’t want my entire movie collection back in the living room.</p><p>But friction sometimes does more than inconvenience us. Physical limits force choices. Ownership creates control. Effort makes us deliberate. Objects remain visible even when we’re not actively using them. A shelf of books reminds us what we’ve read and what we haven’t. An album collection becomes a record of choices made and shifting tastes over decades.</p><p>Remove the friction and you don’t simply make the old behavior easier; it changes the behavior.</p><p>That’s the second-order effect.</p><p>Digital technology gave us extraordinary access to books, music, and movies. We responded by changing how we acquire them, how we discover them, how we value them, and even what it means to have a collection.</p><p>And now, in some cases, we’re selectively putting a little friction back.</p><p>I’m not planning to rebuild my collection of 900 CDs. I have moved across countries often enough to remain enthusiastic about possessions that don’t require cardboard boxes.</p><p>But I may buy a Blu-ray player. Because apparently the future is wonderful, provided you keep enough obsolete technology around to watch <em>Young Frankenstein</em>.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=133ae22c5d16" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Performing for the Algorithm]]></title>
            <link>https://medium.com/@philgold/performing-for-the-algorithm-3d45e2aa0683?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/3d45e2aa0683</guid>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[social-media]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[human-behavior]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 15 Sep 2026 13:31:01 GMT</pubDate>
            <atom:updated>2026-09-15T13:31:01.782Z</atom:updated>
            <content:encoded><![CDATA[<h4><strong>How opaque systems change our behavior without ever explaining what they want</strong></h4><figure><img alt="" src="https://cdn-images-1.medium.com/max/645/1*3hiQJ13qQPVL8qnGdCncLg.png" /></figure><p>A few weeks ago, I wrote a LinkedIn post about receiving a rejection letter. I’m 69 years old, I’ve self-published several children’s books, and I’d recently started querying literary agents for the first time. One of them sent me a very nice rejection. She’d enjoyed <em>The Stinky Princess</em>, but her client list was full.</p><p>I wrote about why this made me unexpectedly happy. At 69, I’d found another way to be a beginner. Getting rejected meant I was actually trying something I hadn’t done before.</p><p>The post took off, at least by my standards. It received 8,401 impressions, 101 engagements, 29 comments and a repost. More interestingly, only 24 percent of those impressions came from within my LinkedIn network. LinkedIn had decided to show the post to thousands of people who didn’t know me.</p><p>I was delighted, and then I made a mistake: I tried to figure out why.</p><h4><strong>Reverse Engineering the Oracle</strong></h4><p>LinkedIn gives you information about what happened to a post. It tells you how many impressions it received, how many people engaged with it and something about who those people were. So, I started developing theories about why this post had done so well. Maybe the personal story had helped? Maybe Thursday was a particularly good day for me to post? Maybe the image was the key? Maybe keeping the story native to LinkedIn rather than sending people elsewhere made a difference? Maybe the publishing hashtags helped reach an audience I hadn’t tapped before?</p><p>All perfectly plausible, and, as it turned out, not terribly useful.</p><p>I had another story involving one of my children’s books where a marketer contacted me to say how much she’d enjoyed the book, only to eventually admit she hadn’t read it.</p><p>I thought this was fascinating, particularly as an example of how AI can make counterfeit attention extraordinarily cheap, so I wrote a Medium article about it. Then I created a separate LinkedIn post, using what I thought I’d learned from the previous one to guide me.</p><p>My successful post had involved children’s publishing, and so did this one. It had told a personal story, and so did this one. I posted on Thursday again, used an illustration from the book, kept the story on LinkedIn rather than merely posting a link, and used relevant publishing hashtags.</p><p>The result was 209 impressions, 88 percent of them from within my network.</p><p>Hmm.</p><h4><strong>Clearly, More Research Was Required</strong></h4><p>Then came Goodreads.</p><p>The marketer had mentioned my “other seven books.” This puzzled me because I had written only four, so I went looking for the source and discovered that Goodreads thought I had written eight. It had merged some of my books with books by another Phil Gold, a distinguished Canadian physician and cancer researcher who has been complicating my search results for years.</p><p>I corrected the Goodreads record and wrote another article about what happens when incorrect information enters apparently authoritative databases. Then I turned that into another LinkedIn story, beginning with what I thought was a pretty good opening:</p><p>“I recently discovered that I had written eight books. This came as something of a surprise, since I thought I had only written four.”</p><p>It was another publishing story, another personal story and another slightly ridiculous situation. This time I changed the day and posted Wednesday afternoon, because maybe Thursday wasn’t the secret after all.</p><p>The result was 148 impressions, 87 percent from within my network.</p><p>At this point, I developed another theory: maybe the post was too long.</p><p>Notice what was happening. LinkedIn hadn’t told me my post was too long. It hadn’t told me that Thursday mattered or didn’t matter, that publishing stories worked, that personal stories worked, that images mattered or that hashtags helped. In fact, it hadn’t actually told me anything.</p><p>I was changing my behavior anyway.</p><h4><strong>Then I Stopped Experimenting</strong></h4><p>A few days later, I received a wonderful review of <em>The Duchess Would Not Approve!</em> from Midwest Book Review. I posted about it on LinkedIn, not because I was conducting another experiment but because I had good news and wanted to share it. I included an illustration from the book.</p><p>Then I noticed something I hadn’t seen before. In the corner of the image was a tiny “Cr” icon. Clicking it revealed Content Credentials showing that the image had been created using OpenAI’s image-generation technology.</p><p>Interesting. Could LinkedIn be treating AI-generated images differently?</p><p>I have no objection to people knowing that I use AI in my illustration process. I’ve written publicly about it. But if an AI provenance marker were being used as an automatic reason to restrict distribution, I’d have considerably stronger feelings about that.</p><p>For a few minutes, I thought I might have found something. Then I checked the previous posts.</p><p>The counterfeit-attention post had carried the Content Credentials marker and had performed poorly. The Goodreads post hadn’t carried one and had performed even worse. My wildly successful <em>Stinky Princess</em> post hadn’t carried one either, although the cover included AI-generated artwork that I’d subsequently assembled into a composite image in Adobe Illustrator.</p><p>Perhaps Content Credentials mattered. Perhaps they didn’t. I added another theory to the pile.</p><p>The book-review post eventually received 78 impressions, 97 percent from within my network. For those keeping score, my attempts to understand LinkedIn had now produced 8,401 impressions, followed by 209, then 148, then 78.</p><p>I was getting remarkably bad at this.</p><h4><strong>The Algorithm Doesn’t Have to Explain Itself</strong></h4><p>Somewhere in the middle of all this, I realized that the most interesting part of my experiment wasn’t whether I could figure out LinkedIn’s algorithm. It was what the algorithm was doing to me.</p><p>The first-order effect of an algorithmic feed is straightforward. There’s far more content available than any of us could possibly consume, so a system decides what to show us. That’s useful, and probably unavoidable.</p><p>But there’s another group affected by that decision: the people who want to be shown.</p><p>The second-order effect is that once creators, businesses, musicians, authors or job seekers realize that an algorithm stands between what they produce and the people they hope to reach, they begin adapting to it. The strange part is that we don’t even need to know how the algorithm works. We just need theories about how it works.</p><p>A post performs well, so we examine it for clues. Was it the subject, the length, the time of day, the image or the hashtags? Did people spend longer reading it? Did early comments trigger wider distribution? Did the system classify it differently? Was Mercury in retrograde? (I haven’t tested that last one yet.)</p><p>Then we change something and try again.</p><p>The system has changed my behavior without ever explaining its own behavior to me. The algorithm never issued an instruction; it never said, “Phil, we’d like you to write shorter posts.” But after watching a long post that I thought was genuinely funny receive 93 impressions, guess what I’m planning to try next?</p><p>A shorter post.</p><h4><strong>We’re All Doing This</strong></h4><p>LinkedIn is only one example. Musicians think about what Spotify’s recommendation systems might favor. Writers think about what Amazon might surface. Businesses structure websites around what they believe Google wants. Job applicants add keywords because they know software may evaluate their résumés before a person does.</p><p>Entire professions have grown around trying to understand these systems. Some of the resulting advice is based on good evidence, some on experience, and some is folklore repeated so many times that it begins to sound like established fact.</p><p>The situation becomes even more interesting because the systems themselves keep changing. Suppose enough people become convinced that an algorithm rewards a particular behavior and start behaving that way. Suddenly the platform fills with content optimized around that belief, which changes the environment in which the algorithm is operating.</p><p>The people designing the algorithm respond. Users observe the new results, develop new theories and adapt again.</p><p>Technology changes behavior. Human behavior changes the environment. Then technology adapts to the changed environment, and around we go.</p><h4><strong>The Frustrating Part</strong></h4><p>I understand why LinkedIn doesn’t publish a precise recipe for getting broad distribution. If everyone knew exactly what the system rewarded, we’d optimize for the recipe rather than for the people reading our posts. In fact, we’re already doing some of that without knowing the recipe.</p><p>But there’s a considerable distance between publishing the recipe and providing almost no meaningful explanation at all, and that’s the part I find frustrating.</p><p>My rejection post received 8,401 impressions, 76 percent from people outside my network. A funny story about discovering that I’d apparently written four books I knew nothing about received 148 impressions, 87 percent from people already connected to me.</p><p>I can see what happened, but I have almost no idea why LinkedIn had decided to show the first post to thousands of people who didn’t know me.</p><p>And because distribution matters to me, I adapt. That’s what people do. We’re remarkably good at spotting patterns, including patterns that may not actually be there, particularly when the outcome matters to us.</p><p>The algorithm is trying to learn what people will do. Meanwhile, people are trying to learn what the algorithm wants them to do. Neither side has complete information about the other.</p><h4><strong>What Have I Learned?</strong></h4><p>Not how to succeed on LinkedIn, obviously. If anything, my experiments have become progressively more impressive demonstrations of how <em>not</em> to reproduce an 8,401-impression post.</p><p>But I have learned something more interesting. We tend to think of algorithms as systems that observe, classify and predict human behavior. That’s only half the relationship.</p><p>Once we know the algorithms are watching us, we start watching them back. We study their decisions, develop theories and change what we do. Then they observe the changed behavior and adjust again.</p><p>Technology starts the change. People decide what happens next. And sometimes the technology gets another turn after that.</p><p>This week, I’m going to try a shorter LinkedIn post. I have absolutely no evidence that this will work, but apparently LinkedIn has trained me to try it.</p><p>I’ll let you know what happens.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3d45e2aa0683" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[When the Internet Decides Who You Are]]></title>
            <link>https://medium.com/@philgold/when-the-internet-decides-who-you-are-806f3615a23d?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/806f3615a23d</guid>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[algorithms]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[digital-identity]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 08 Sep 2026 13:31:01 GMT</pubDate>
            <atom:updated>2026-09-08T13:31:01.380Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong>A second-order effect of letting algorithms organize the world</strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*RT26TEKtFA6B2zhW8kmBIg.png" /></figure><p>I recently discovered that I had written eight books. This came as something of a surprise, since I thought I had only written four.</p><p>I found out about my unexpectedly productive literary career after receiving a series of messages from people offering to help market my newest children’s book, <em>The Duchess Would Not Approve!</em> One of them referred to my other seven books.</p><p>I was pretty sure I hadn’t written them. But I was curious where she’d gotten the number, so I started looking.</p><p>The answer was Goodreads. According to my author page, I was responsible for eight books. Four really were mine. The other four belonged, in one form or another, to another Phil Gold.</p><p>And not just any Phil Gold.</p><h4><strong>The Other Phil Gold</strong></h4><p>I’ve had a minor identity problem on the internet for years. There’s another Phil Gold, Dr. Phil Gold, a distinguished Canadian physician and cancer researcher who co-discovered an important tumor marker and has accumulated a rather intimidating collection of professional honors.</p><p>He is, through absolutely no fault of his own, the bane of my online existence.</p><p>This is the reason I’ve used <strong>therealphilgold</strong> as my online identity for almost 20 years. I don’t seriously claim to be any more real than Dr. Gold, but <strong>philgold</strong> was already rather thoroughly occupied, and I needed some way to differentiate myself.</p><p>(For the record, I can prove that I’ve been using the “real” formulation since at least 2008, well before a certain future American president made the same naming convention considerably more famous. I would like the historical record to reflect this fact.)</p><p>Over the years, search engines have actually gotten pretty good at telling us apart. Right this moment, if I search Google for Phil Gold, I show up at numbers two and five. Dr. Gold occupies the other prominent positions. I haven’t displaced him, exactly. We’ve reached a sort of search-engine détente.</p><p>Goodreads, unfortunately, decided the distinction was unnecessary.</p><p>My author page included one of Dr. Gold’s books, another book to which he’d contributed, and two reprints. Add those to my four actual books, and Goodreads confidently presented the world with a single, unusually versatile Phil Gold.</p><p>Apparently, I write children’s books, business books and medical literature in my spare time. Impressive!</p><h4><strong>When an Error Becomes Information</strong></h4><p>On its own, this was a fairly harmless database error. Two people had the same name, an automated system associated their work with the same author, and nobody was injured.</p><p>But then something interesting happened. Someone researching me found that information and used it.</p><p>The marketer who contacted me didn’t invent the number eight. She’d done what we increasingly expect people, and the technologies assisting them, to do; she searched for information about me, found an apparently authoritative source and used what she found to personalize her message.</p><p>The problem was that the information was wrong.</p><p>That creates a second-order effect that interests me much more than the original mistake. Technology has made it extraordinarily easy to gather information about people. Search engines, social networks, databases and now generative AI can assemble a surprisingly detailed picture of someone in seconds. Most of the time, this is enormously useful.</p><p>But those systems don’t begin with reality. They begin with <strong>data about reality</strong>, and the two aren’t always the same thing.</p><p>Once incorrect information enters that ecosystem, it can stop looking like an error and start looking like a fact. One site lists something incorrectly. Another system finds it. Someone uses that system to research a person. The error appears in an email, a profile, a recommendation or perhaps another database.</p><p>At some point, nobody involved necessarily knows where the information originally came from. It has simply become something the internet “knows.”</p><h4><strong>We Used to Correct People</strong></h4><p>Mistaken identity isn’t a new problem. People have been confusing people with similar names for as long as we’ve had names.</p><p>What’s different is the scale and persistence of the mistake. If someone once confused me with another Phil Gold, I could correct them. The correction happened between two people, and the mistake usually ended there.</p><p>Now I may have to correct the <strong>representation of me</strong> that other people and automated systems are consulting.</p><p>That’s what I did with Goodreads. I claimed my author profile, had the incorrect books separated from mine, added the correct information and generally persuaded the system that Dr. Gold and I were two different people.</p><p>In other words, I had to maintain not only my identity, but the <strong>digital representation of my identity</strong>.</p><p>That’s becoming a normal part of modern life. We check what Google says about us. We update our LinkedIn profiles. Authors claim their Goodreads and Amazon pages. Musicians verify their artist profiles on streaming services. Businesses monitor listings, reviews and maps.</p><p>We aren’t simply telling other people who we are anymore. Increasingly, we’re telling databases who we are so that the databases can tell other people.</p><h4><strong>AI Adds Another Layer</strong></h4><p>Generative AI makes this more interesting because it can synthesize information rather than merely display it.</p><p>Ask a search engine for Phil Gold and you get a collection of results that you can inspect. If two different people appear, there’s at least a chance you’ll notice.</p><p>Ask an AI system about someone and you may instead receive a coherent description assembled from multiple sources. That’s wonderfully convenient, but coherence can create its own kind of confidence. An answer can sound internally consistent even when some of the underlying information belongs to somebody else.</p><p>The first-order improvement is obvious. We no longer have to spend as much time finding and assembling information ourselves.</p><p>But if we stop looking at the sources, we also lose some of the clues that might have told us the information didn’t fit together.</p><p>A children’s-book author who also happens to have written medical literature about cancer research should perhaps raise an eyebrow (unless, of course, he’s like me and extraordinarily versatile).</p><h4><strong>The Internet’s Version of You</strong></h4><p>There’s a larger issue here than whether Goodreads thinks I have an unexpected sideline in medicine. As more decisions are made with the help of automated systems, the version of us contained within those systems matters.</p><p>A recruiter may encounter a database’s version of a job candidate before meeting the person. A customer may encounter Google’s version of a business before visiting it. A reader may encounter Goodreads’ version of an author. An AI system may assemble its own version from all of them.</p><p>Most of the time, those representations are probably close enough to reality that we don’t think much about them. It’s when they’re wrong that things get interesting.</p><p>Because once we discover that algorithms are constructing versions of us, we adapt. We claim profiles. We correct listings. We add structured information to websites. We think about search results. We monitor what appears under our names.</p><p>Technology organizes information about us, and then we change our behavior to influence how technology organizes information about us.</p><p>There’s the second-order effect.</p><p>I started using <strong>therealphilgold</strong> years ago because there were already two “real” Phil Golds on the internet and I wanted people to be able to tell us apart.</p><p>Apparently, that job isn’t finished.</p><p>Dr. Gold and I continue to coexist peacefully, although I have no idea how he feels about any of this. For all I know, somewhere in Canada a distinguished cancer researcher is furiously trying to convince an algorithm that he did not write <em>The Stinky Princess</em>.</p><p>If so, I apologize.</p><p>But he should really get his own username.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=806f3615a23d" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Counterfeit Attention]]></title>
            <link>https://medium.com/@philgold/counterfeit-attention-25bc4fcb61bd?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/25bc4fcb61bd</guid>
            <category><![CDATA[communication]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[marketing]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[technology]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 01 Sep 2026 11:01:01 GMT</pubDate>
            <atom:updated>2026-09-01T11:01:01.130Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*-YKarkHiZM6iBdh4hVFVmw.png" /><figcaption>Illustration: Concept and direction by Phil Gold. Dutifully mass-produced by ChatGPT.</figcaption></figure><h3><strong>When technology makes it easy to sound like you care</strong></h3><p>I recently received a message from someone who wanted to tell me how much she enjoyed my new children’s book, <em>The Duchess Would Not Approve!</em></p><p>This was very nice of her, except I was pretty sure she hadn’t read it.</p><p>The only place the book was available was Amazon, where the book had only just been published. I hadn’t even received my author copies. I asked her how she’d managed to read it.</p><p>Her explanation began to evolve.</p><p>She claimed she read it online, but no online version of the book existed. On further questioning, she admitted that she hadn’t actually read the book. Instead, she’d become familiar with it through its synopsis and cover. This was apparently enough to give her a remarkably deep appreciation for my work, along with some thoughts about how she might help me promote it.</p><p>In other words, the compliment wasn’t really a compliment. It was the beginning of a sales pitch.</p><p>There was also a rather mysterious reference to the other seven books I’d supposedly written. That turned out to have an interesting explanation, but it’s a story for another day.</p><p>What interested me more was the way her original message was designed to create the impression that someone had actually paid attention to what I’d done. This isn’t particularly new, of course. What’s changing is how easy it’s become to do more convincingly.</p><h4><strong>We Learned to Ignore the Old Version</strong></h4><p>I’ve been uploading my music to SoundCloud for years, and whenever I post a new track, somebody will almost instantly leave a comment telling me how much they love it and offering to help me promote it.</p><p>The speed with which some of these people reach out is impressive. Apparently, my music can touch the human soul and transform a person’s life in approximately 32 seconds.</p><p>These messages aren’t that difficult to recognize. They’re generic, they’re repetitive, and they often use the exact same language. Most of us learned a long time ago how to deal with them. We generally ignore them.</p><p>The same thing happened with email. Once mass email became cheap, inboxes overloaded with messages, but they were usually generic enough that they were easy to identify as spam. It was when they contained specific information about you, your company or something else that suggested they came from an actual person who had made the effort to contact you that we stopped and took the time to read them.</p><p>That effort mattered because personalization used to cost something.</p><h4><strong>When Personalization Becomes Cheap</strong></h4><p>Generative AI changes the equation because one of the things it does remarkably well is gather information, identify relevant details, and produce language that sounds specifically tailored to an individual.</p><p>That’s incredibly useful. It’s also incredibly useful if you’re trying to convince 5,000 strangers that you’ve personally taken an interest in each of them.</p><p>So, marketers adapted. They started adding details that suggested they’d actually done some homework. The message might mention your company, something you’d recently published or a project you’d worked on. Even if you weren’t interested in whatever they were selling, the specificity suggested that someone had at least spent a few minutes deciding why they were contacting you.</p><p>AI didn’t invent fake attention. It made fake attention much more convincing.</p><p>And that’s where the second-order effect becomes interesting. The obvious first-order effect is that technology makes personalized communication faster and cheaper. A marketer can research more prospects, create more individualized messages and reach more people with less effort.</p><p>Specificity was a signal of genuine interest because producing that specificity required effort.</p><p>Now it doesn’t necessarily require much effort at all. The details that once suggested someone had spent time learning about you can increasingly be assembled and incorporated into a convincing message automatically.</p><p>So, we adapt again, and we start distrusting the personalization itself.</p><p>That’s a much more interesting consequence than simply getting more sophisticated spam. When everyone can manufacture the signals of authenticity, authenticity itself becomes harder to signal.</p><h4><strong>The Problem Isn’t Using AI</strong></h4><p>This isn’t an argument that using AI automatically makes something inauthentic. That would be a particularly difficult argument for me to make because I use it myself.</p><p>The illustrations in my three published children’s books were all created with AI assistance. Creating them involved developing visual identities for the characters, deciding how they dressed, how they behaved and what expressions they should have. I developed the look of the settings, and worked through composition, continuity, poses, clothing, props and quite a lot of revision until the illustrations supported the story I’d written.</p><p>AI executed on my input. It didn’t manufacture my involvement.</p><p>I think that’s an important distinction. AI-assisted communication isn’t inherently deceptive, any more than AI-assisted illustration is. Technology has always changed the way work gets done, and generative AI is giving us another set of tools for doing it.</p><p>The problem isn’t using AI to help communicate something. The problem is using it to counterfeit attention that was never there in the first place.</p><p>If I’ve genuinely researched someone and use AI to help me organize my thoughts or write a clearer message, the attention is real. If I ask a system to create 500 messages designed to make 500 people believe I’ve personally researched each of them when I haven’t, something quite different is happening.</p><p>One uses technology to communicate attention. The other uses technology to simulate it.</p><h4><strong>What Happens to Genuine Attention?</strong></h4><p>The obvious cost of all this is the time we spend trying to determine whether somebody genuinely has an interest in us or is simply trying to sell us something.</p><p>But I think there’s a less obvious cost, and it affects the people who really are paying attention.</p><p>Suppose someone actually does read one of my books and decides to contact me. They mention a character, refer to something that happens in the story and explain what they liked about it. Not long ago, those details would have been pretty good evidence that they’d actually read the book.</p><p>Now I may find myself wondering where they got the information.</p><p>That’s the second-order effect that concerns me more. The technology doesn’t merely make insincere communication easier. It gradually reduces the value of the signals sincere people use to demonstrate sincerity.</p><p>We’ve seen versions of this before. Mass email made email easier to ignore. Caller ID became less useful once numbers could be spoofed. Online reviews became less trustworthy once businesses realized how valuable good reviews could be and people discovered ways to manufacture them.</p><p>In each case, the technology creates an opportunity, people exploit that opportunity, and everyone else adapts to the exploitation. Eventually, we’re all operating in a slightly different environment.</p><h4><strong>Attention Still Has Value</strong></h4><p>We often talk about information being valuable, but information is becoming extraordinarily cheap. Personalization is getting cheaper too. What may become increasingly valuable is credible evidence that another person actually paid attention.</p><p>Genuine attention isn’t simply knowing a collection of facts about someone. It’s asking a question that only makes sense because you listened to the answer before it. It’s noticing something that wasn’t conveniently summarized on the first page of a website. It’s remembering something from an earlier conversation and responding to the person rather than the profile.</p><p>Technology can help us do all of those things. But if we use it instead to manufacture the appearance that we’ve done them, we slowly make those signals less meaningful for everyone.</p><p>That’s the irony I find most interesting. The better technology becomes at helping us create the appearance of personal attention, the more suspicious we may become of personal attention itself.</p><p>The more technology helps us sound human, the harder it becomes to prove that we are.</p><p>And somewhere out there, someone is probably preparing to send me a deeply personal message about how much <em>The Duchess Would Not Approve!</em> changed her life.</p><p>This time, I may ask what she thought of page two.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=25bc4fcb61bd" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Second-Order Effect of Getting Exactly What We Want]]></title>
            <link>https://medium.com/@philgold/the-second-order-effect-of-getting-exactly-what-we-want-020bbe8506ab?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/020bbe8506ab</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[innovation]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 25 Aug 2026 13:01:03 GMT</pubDate>
            <atom:updated>2026-08-25T13:01:03.297Z</atom:updated>
            <content:encoded><![CDATA[<h3><strong><em>What happens to serendipity when technology eliminates the need to wander?</em></strong></h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*WG_OwKQPjo6QGMKePf9GfA.jpeg" /><figcaption><a href="https://commons.wikimedia.org/wiki/User:BrokenSphere">BrokenSphere</a> / <a href="https://creativecommons.org/licenses/by-sa/3.0">Creative Commons Attribution-Share Alike 3.0</a></figcaption></figure><p>Some of my favorite musical discoveries around the turn of the century came from Borders Books.</p><p>In addition to books, Borders had a large music selection with listening stations. I would put on the headphones and sample CDs by artists I knew little or nothing about. That’s how I discovered Robbie Fulks, Sarah McLachlan, and The New Pornographers. They were all over the board genre-wise, but I found something I loved in each of them: great musicianship and terrific songwriting.</p><p>There wasn’t much science involved. If something looked interesting, I listened. Sometimes I liked it, and sometimes I didn’t. And every once in a while I found something I might never have gone looking for because I didn’t know it existed.</p><p>Today, I have access to a ton more music than Borders could ever have stocked. Spotify looks at what I’ve listened to and suggests other artists I’m likely to enjoy. Its Artist Radio feature regularly introduces me to music I’ve never heard before. In almost every measurable way, finding new music has gotten easier.</p><p>But there’s a second-order effect worth thinking about.</p><p>Technology has become extraordinarily good at eliminating the need to wander. Recommendation systems solve a very real problem. There is simply too much available to explore ourselves. Too much music. Too many movies. Too many books. Too many articles. Too many restaurants. Too many products.</p><p>So, technology helps us choose.</p><p>That’s the first-order effect. It’s easier to find things we’re likely to enjoy. Because the system is so good at finding things for us, we don’t need to spend as much time looking ourselves. What we liked yesterday helps determine what we’re shown today. What we choose today helps determine what we’re shown tomorrow.</p><p>That’s not necessarily a bad thing. Often it works remarkably well, but it changes the way discovery happens.</p><p>When I wandered through Borders, I wasn’t being presented with music based on my previous listening habits. And my own tastes didn’t necessarily fit neatly into a single category, either. Robbie Fulks, Sarah McLachlan, and The New Pornographers sounded very different from one another, but I found qualities I loved in all three. The store was certainly making choices about what went into those listening stations, but I could sample two completely unrelated artists simply because their CDs happened to be next to each other.</p><p>Discovery had more dead ends, and some of those dead ends turned out not to be dead ends at all.</p><p>I still use algorithms to discover music today. But I’ve noticed that I often wander away from them. If Spotify Artist Radio plays a song I particularly like, I’ll stop and explore the artist. I’ll listen to some of their other music, which may lead me somewhere else.</p><p>SoundCloud produces a different kind of wandering. If someone likes one of my tracks or follows me, I often check out their page. If they’ve posted music of their own, I’ll listen. I’ve found some artists I really like that way. Nobody recommended them to me. They happened to cross my path, and I chose to follow it.</p><p>Maybe that’s the distinction that matters.</p><p>Serendipity doesn’t necessarily require randomness. It requires being willing to follow something you weren’t originally looking for.</p><p>And this isn’t just about music. Search engines take us directly to information. GPS takes us directly to destinations. Recommendation systems suggest what to watch, read, buy, or eat. Artificial intelligence can increasingly give us a synthesized answer without requiring us to visit any of the places where we might once have searched for it ourselves.</p><p>All of these technologies remove friction, and most of the time that’s exactly what we want.</p><p>But wandering <strong>is</strong> friction.</p><p>Browsing a bookstore without knowing what you want is inefficient. Listening to six songs you don’t like before finding one you love is inefficient. Taking a different street home for no particular reason is inefficient. Reading something outside your field that turns out not to be useful is inefficient.</p><p>If technology gets better and better at eliminating those dead ends, the second-order question is: what else might disappear along with them?</p><p>Organizations should probably be asking the same question. Businesses have spent decades making information easier to find, processes more efficient, communication faster, and expertise more accessible. AI promises to accelerate all of that.</p><p>Again, those are real improvement, but organizations also depend on people encountering ideas they weren’t looking for. Someone in marketing hears what engineering is working on. A conversation in an informal Slack channel wanders into a problem nobody intended to discuss. Someone asks a colleague a simple question and gets not only the answer, but also, “You should really talk to Maria about this.”</p><p>Those kinds of encounters happen in offices, but they happen remotely too. Technology hasn’t necessarily eliminated the hallway conversation. In some cases, it has expanded it. I’ve had long, unexpected conversations with people on the other side of the world that could never have happened in a traditional office hallway.</p><p>The conditions for serendipity changed, and people adapted.</p><p>That’s the second-order effect leaders need to notice.</p><p>When we remove friction from a system, it’s worth asking what else that friction was doing. Did a useful behavior disappear? Did it move somewhere else? Did people invent a new way to accomplish the same thing?</p><p>The answer isn’t to make work deliberately inefficient or to reject technologies that help us find things faster. It’s to leave room for curiosity.</p><p>That applies personally, too.</p><p>If a recommendation system gives you something interesting, follow the breadcrumb instead of simply waiting for the next recommendation.</p><p>Browse occasionally without knowing exactly what you’re looking for. Ask another person what they’ve been reading, watching, or listening to rather than asking an algorithm. Read something outside your field. Walk down a street you’ve never taken. Pick up a book because the cover intrigues you.</p><p>And perhaps resist the temptation to optimize every choice.</p><p>Artificial intelligence can even help. Instead of always asking it for the best answer, sometimes ask it for three completely different ways of looking at the problem. Ask what an unrelated field might teach you. Ask it to point you toward a rabbit hole, and then go explore it yourself.</p><p>Technology is extraordinarily good at helping us get from A to B. Maybe we should occasionally leave ourselves enough room to discover C.</p><p>Convenience helps us find what we’re looking for.</p><p>Serendipity helps us find what we never knew to look for.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=020bbe8506ab" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Technology Doesn’t Disrupt Business. People Do.]]></title>
            <link>https://medium.com/@philgold/technology-doesnt-disrupt-business-people-do-9ba4a53ceb5c?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/9ba4a53ceb5c</guid>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[organizational-change]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 18 Aug 2026 12:47:31 GMT</pubDate>
            <atom:updated>2026-08-18T12:47:31.371Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*o8sHZ2fFArPWLdI05cP84A.png" /><figcaption>Concept and direction by Phil Gold; image created with ChatGPT</figcaption></figure><h3>The biggest impact of new technology isn’t what it does. It’s what people do in response.</h3><p>Whenever a new technology arrives, businesses naturally focus on specifics. How much will it cost? How much time will it save? What processes can it automate? How many people will we need? What new capabilities will it give us?</p><p>Those are all important questions, but they’re only the beginning.</p><p>In my previous articles, I’ve written about first-order effects, the immediate consequences of change, and second-order effects, what happens after people adapt to those consequences.</p><p>Businesses tend to be fairly good at planning for first-order effects. We build implementation plans, calculate returns on investment, redesign processes, and estimate productivity gains.</p><p>What we’re much less good at predicting is what people will do in response.</p><p>We’ve seen this happen over and over again.</p><p>Think about something as ordinary as ordering dinner. Restaurant delivery existed long before smartphones, but it used to involve considerably more effort. You needed to know which restaurants delivered, find a menu, call in an order, give someone your address, arrange payment, and then wait without having much idea when your food would arrive.</p><p>Technology gradually removed most of that friction. Apps made it possible to browse restaurants, order, pay, and track a delivery from a phone.</p><p>Then COVID arrived.</p><p>The pandemic didn’t create food delivery technology, but it dramatically accelerated its adoption. Millions of people changed how they bought restaurant food, and restaurants changed in response.</p><p>Menus changed. Packaging changed. Kitchens changed. Restaurants that had never considered delivery began offering it. Delivery-only kitchens became practical. Customers began expecting food from restaurants that previously would never have come to them.</p><p>The technology made the change possible, but people’s behavior turned it into disruption.</p><p>Spreadsheets offer a different example. Electronic spreadsheets made calculations and financial modeling dramatically faster. But businesses didn’t simply perform the same amount of analysis in less time and send everyone home early.</p><p>Instead, they did more analysis.</p><p>Forecasts could be revised more often. More scenarios could be modeled. More reports could be produced. Questions that once required hours or days of manual calculation could suddenly be answered much more quickly.</p><p>The technology increased what people could do, and organizations responded by changing what they expected people to do.</p><p>And now we’re watching the same process unfold with artificial intelligence.</p><p>Much of the current conversation about AI focuses on what it can do. What can it write? What can it analyze? What can it automate? How much time can it save? Which jobs can it replace?</p><p>Again, those are reasonable questions. But they’re first-order questions.</p><p>Suppose AI reduces a task that once required four hours to twenty minutes. That’s a real productivity gain, but it tells us almost nothing about what happens afterward.</p><p>Maybe the employee produces more. Maybe the organization increases its expectations. Maybe the employee uses the extra time for work that previously wasn’t practical. Maybe the quality improves because there’s more time for judgment and refinement. Or maybe management decides fewer people can now produce the same amount of work.</p><p>Any of those things could happen, and probably others we haven’t thought of yet. That’s the problem with trying to predict disruption by looking only at the technology. People don’t stand still while technology changes around them.</p><p>And with AI, we’re seeing another wrinkle: people aren’t waiting for organizations to finish implementing it before they begin adapting.</p><p>Employees are already using AI tools, sometimes with their employer’s encouragement and sometimes without it. People are learning new skills because they believe they’ll need them. Others are resisting technologies they fear could threaten their jobs. Job applicants use AI to write applications while employers use AI to screen them. Managers may begin expecting work to be completed faster simply because they know the technology exists.</p><p>In other words, people aren’t just responding to changes AI has already created. They’re anticipating the changes they think are coming and adjusting their behavior in advance.</p><p>That matters because organizations are made up of people, and people don’t behave like static pieces on a spreadsheet.</p><p>There’s another risk in thinking purely in terms of automation. A job isn’t necessarily just a collection of tasks. Automating some of those tasks doesn’t automatically eliminate the other value a person provides, whether that’s experience, judgment, relationships, institutional knowledge, or simply knowing how things actually get done.</p><p>Tasks are relatively easy to inventory. Human contribution is considerably harder.</p><p>None of this means AI won’t eliminate jobs. It already has, and that’s not going to stop right away. Technological change has always created winners and losers, and there’s no reason to assume this time will be different.</p><p>But knowing which tasks AI can perform tells us surprisingly little about what an organization will look like after people respond to it.</p><p>And that’s the part I think leaders need to spend more time thinking about.</p><p>When considering a new technology, it’s perfectly reasonable to ask what it will allow the organization to do. But there’s another question that deserves just as much attention:</p><p>What will people do differently because this technology exists?</p><p>What new expectations will it create? What behaviors might it encourage? What will customers begin expecting? What will employees stop doing, start doing, or find entirely new ways of doing?</p><p>We’re not going to predict every consequence. We can’t. Leadership isn’t about certainty. It’s about making the best decisions we can with the information available.</p><p>And part of making those decisions is recognizing that the implementation plan isn’t the end of the change. It’s the beginning.</p><p>Technology starts the change.</p><p>People decide what happens next.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=9ba4a53ceb5c" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Where the Time Went]]></title>
            <link>https://medium.com/@philgold/where-the-time-went-b77865132244?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/b77865132244</guid>
            <category><![CDATA[human-behavior]]></category>
            <category><![CDATA[transformation]]></category>
            <category><![CDATA[technology]]></category>
            <category><![CDATA[second-order-effect]]></category>
            <category><![CDATA[productivity]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 11 Aug 2026 08:40:32 GMT</pubDate>
            <atom:updated>2026-08-11T08:40:32.718Z</atom:updated>
            <content:encoded><![CDATA[<h3>Technology Saves Time. Human Nature Spends It.</h3><figure><img alt="" src="https://cdn-images-1.medium.com/max/500/1*XMPEYTTSEraHqJ7zeTepag.jpeg" /><figcaption>Image courtesy of <a href="https://www.shutterstock.com/g/cyano">Stokkete</a> / Shutterstock</figcaption></figure><p>Every generation seems to believe that the next technological breakthrough will finally give us more free time.</p><p>The washing machine. The personal computer. Email. Smartphones. Artificial intelligence. Each one promised to save us time. And to be fair, each one did.</p><p><strong>So, here’s a simple question: where did the time go?</strong></p><p>In my previous article, I introduced the idea of first-order and second-order effects. First-order effects are the immediate consequences of a change. Second-order effects are the changes that happen after people adapt.</p><p>When a new technology saves us time, we instinctively picture a simple equation. A task that took an hour now takes thirty minutes. Thirty minutes becomes fifteen. Surely that means we have more time available for everything else.</p><p>Yet that isn’t usually what happens.</p><p>The technology succeeds. The time savings are real. What changes is how we choose to spend them.</p><p><strong>Technology saves us time. Human nature spends it.</strong></p><p>Take email. It removed much of the friction from communication. No paper, envelopes, stamps, or trips to the mailbox. Sending a message became almost effortless.</p><p>Did we spend less time communicating? Hardly.</p><p>Instead of writing a handful of thoughtful letters each week, many of us now process dozens, sometimes hundreds, of emails every day. The effort we once spent sending messages has largely been replaced by the effort of managing them.</p><p>Digital photography followed much the same path. When every photograph cost money to develop, we tended to think before pressing the shutter. Today we can capture hundreds of images without hesitation.</p><p>The effort didn’t disappear. It shifted. Instead of deciding which pictures to take, we decide which ones to keep, how to organize them, and how to find them again months or years later.</p><p>GPS changed something else. Before satellite navigation, many of us gradually developed mental maps of the places we visited. We knew landmarks. We learned alternate routes. If traffic blocked our usual drive home, we often had another option already in mind.</p><p>Today, many people simply follow the blue line.</p><p>Navigation has become easier, but it has also become different. It’s not unusual to feel surprisingly disoriented when the suggested route suddenly changes, even somewhere familiar.</p><p>Artificial intelligence may become the most dramatic example yet. It can summarize documents, write first drafts, answer questions, generate ideas, and complete tasks that once required significant time and effort.</p><p>The immediate effect is easy to see. The more interesting question is what happens next. What will we choose to do with the time, attention, and mental energy those tools give back?</p><p>That’s the pattern. The productivity gains are real. What’s different is where the effort goes.</p><p>Sometimes it becomes more communication. Sometimes it becomes more decisions. Sometimes it becomes higher expectations. And sometimes it’s a skill we gradually stop practicing because technology quietly took over that part of the job.</p><p>None of this is an argument against technology. These innovations have made our lives easier, safer, and more productive in countless ways.</p><p>What they also reveal, however, is something interesting about ourselves: human beings are remarkably bad at leaving a vacuum.</p><p>Whenever technology removes effort from one part of our lives, we almost immediately find something else to do with the newly available time. We don’t preserve the empty space. We repurpose it.</p><p>That’s why predictions about new technology so often miss the mark.</p><p>We spend enormous energy asking what the technology will do. We spend much less time asking what people will do once the technology has done its job.</p><p>As artificial intelligence becomes part of our everyday lives, that may be the most important question of all.</p><p>Because that’s where the truly interesting changes begin.</p><p><strong>Technology doesn’t give us more time.</strong></p><p><strong>It simply changes what consumes it.</strong></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=b77865132244" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The World Doesn’t Change the Way We Think It Does]]></title>
            <link>https://medium.com/@philgold/the-world-doesnt-change-the-way-we-think-it-does-c9f2c1ff8a3c?source=rss-732fed96d8f6------2</link>
            <guid isPermaLink="false">https://medium.com/p/c9f2c1ff8a3c</guid>
            <category><![CDATA[future-of-work]]></category>
            <category><![CDATA[organizational-change]]></category>
            <category><![CDATA[strategy]]></category>
            <category><![CDATA[leadership]]></category>
            <category><![CDATA[systems-thinking]]></category>
            <dc:creator><![CDATA[Phil Gold]]></dc:creator>
            <pubDate>Tue, 04 Aug 2026 13:04:11 GMT</pubDate>
            <atom:updated>2026-08-04T13:04:11.779Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/1000/1*OchXuyUQhMcB_GdmVA755g.jpeg" /><figcaption>Image by <a href="https://www.shutterstock.com/g/Casther">Casther</a>/Shutterstock</figcaption></figure><p>When historians look back at major turning points, they usually focus on a single event like the invention of the printing press, or the arrival of the automobile, or the birth of the internet.</p><p>Today, it’s artificial intelligence.</p><p>The accepted wisdom is that these events changed the world, but I’m no longer convinced that’s true.</p><p>I think they <strong>started</strong> the change.</p><p>The real transformation came later, as millions of people gradually adapted their behavior around the changes. The automobile didn’t just replace horses; it changed where we lived. It reshaped cities and created suburbs. It changed dating, shopping, commuting, vacations, and even where businesses chose to locate.</p><p>None of those things happened because someone invented a better way to get from one place to another; they happened because people reorganized their lives around new possibilities.</p><p>That distinction has become increasingly important to me. I’ve started thinking about it in terms of <strong>first-order</strong> and <strong>second-order effects.</strong></p><p><strong>First-order effects are the direct consequences of an event. Second-order effects are the changes that happen because people adapt to those consequences.</strong></p><p>Most conversations begin and end with the first-order effects, and that’s understandable — they’re immediate, visible, and they make headlines.</p><p>But I think the second-order effects are usually where the bigger story lies.</p><p>For most of my career, I’ve worked with organizations trying to navigate change. Sometimes the change was digital transformation. Sometimes it was remote work. Sometimes it was a reorganization or a shift in strategy.</p><p>What I’ve noticed is that leaders often focus enormous energy on responding to the event itself, and pay far less attention to how people will adapt afterward.</p><p>That’s a mistake.</p><p>Because organizations don’t succeed or fail because change happened. They succeed or fail because of how they adapt to it.</p><p>Technology and policy matter, but eventually, all that becomes the new normal. What determines success is how people reorganize their work, redefine their relationships, develop new habits, and discover new ways of creating value.</p><p>That’s why I’m increasingly interested in second-order effects; not as an academic exercise, but as a practical way of thinking.</p><p>If we spend all our time asking, “What changed?” we’ll always be reacting to yesterday’s news. A more useful question is:</p><p><strong>“Now that this has changed, how will people change because of it?”</strong></p><p>That question shifts our attention from events to adaptation, from technology to behavior.</p><p>In the months ahead, I’d like to explore that question through the lenses of leadership, communication, organizational change, and, yes, artificial intelligence.</p><p>Not because I think AI is the story.</p><p>Because I think human adaptation is.</p><p><strong>The world doesn’t change when something new arrives. It changes when enough people decide to live differently because it arrived.</strong></p><p>That’s where I think the most interesting conversations are about to begin.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=c9f2c1ff8a3c" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>