<?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 Milnesycw on Medium]]></title>
        <description><![CDATA[Stories by Milnesycw on Medium]]></description>
        <link>https://medium.com/@milnesycw?source=rss-16f90032fb28------2</link>
        <image>
            <url>https://cdn-images-1.medium.com/fit/c/150/150/1*Tu1hdeqPVB2uEdW-n0qqoQ@2x.jpeg</url>
            <title>Stories by Milnesycw on Medium</title>
            <link>https://medium.com/@milnesycw?source=rss-16f90032fb28------2</link>
        </image>
        <generator>Medium</generator>
        <lastBuildDate>Fri, 20 Nov 2026 23:10:52 GMT</lastBuildDate>
        <atom:link href="https://medium.com/@milnesycw/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[An Accidental West End Encounter with Miriam Margolyes]]></title>
            <link>https://medium.com/@milnesycw/an-accidental-west-end-encounter-with-miriam-margolyes-697246336078?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/697246336078</guid>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 19:16:42 GMT</pubDate>
            <atom:updated>2026-10-30T19:16:42.917Z</atom:updated>
            <content:encoded><![CDATA[<p>Sometimes the best part of a story is the detail you only understand years afterwards.</p><p>I went into London with my mum’s friend for an evening at the theatre. At the time, she lived near Welwyn Garden City. Hertfordshire has long been part of my life, with family roots in the area; my close friend Shirley now lives up the road in Stevenage.</p><p>Our plan was straightforward: head into town, see Romeo and Juliet at an Off-West End theatre, and enjoy a night out. I certainly wasn’t expecting the person in the next seat to become the part of the evening I would still be talking about years later.</p><p>When I settled into the stalls, I exchanged a casual “hi” with the woman sitting beside me. We got chatting about the play and its themes. It was an easy, friendly conversation between two people who had come to watch the same performance.</p><p>I had no idea she was Miriam Margolyes.</p><p>That is what I like most about the memory. At the time, there was no moment of recognition to change the conversation. I wasn’t trying to think of something clever to say to a famous actor. She was simply a lovely person to chat with before the performance began.</p><p>There was another connection waiting to be noticed. I had studied Romeo and Juliet for my English Literature Certificate of Achievement, and I knew Baz Luhrmann’s film adaptation, in which Miriam played the Nurse. Yet none of that clicked while we were sitting there discussing the play.</p><p>I knew the work. I just didn’t connect it with the person beside me.</p><p>After the curtain came down, my mum’s friend and I rounded off our night in central London with a stop at the Hard Rock Cafe. The theatre, the conversation and the meal became one of those happy memories of a good evening out.</p><p>The recognition came much later.</p><p>Over the years, I saw Miriam appearing in the media. With help from my mum’s late best friend, the pieces eventually came together: the woman I had chatted with in the stalls really was Miriam Margolyes.</p><p>By then, the conversation was already a memory. I couldn’t go back and have the surprised reaction I might have had at the time. All I could do was look back at an ordinary exchange and realise it had contained an extraordinary coincidence.</p><p>I still find it hard to believe that we happened to sit directly next to each other. But I’m glad the recognition came afterwards. It left that conversation exactly as I experienced it: friendly, unforced, and about the play we were both there to see.</p><p>I went to London to watch Romeo and Juliet. Years later, I discovered that I had spent part of the evening chatting with its Nurse.</p><p>It remains one of my favourite London stories.</p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=697246336078" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[Your AI conversations shouldn’t move without telling you]]></title>
            <link>https://medium.com/design-bootcamp/your-ai-conversations-shouldnt-move-without-telling-you-e2e3d5584b7b?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/e2e3d5584b7b</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[equality]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[ai-agent]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Fri, 30 Oct 2026 12:57:32 GMT</pubDate>
            <atom:updated>2026-11-06T07:47:35.154Z</atom:updated>
            <content:encoded><![CDATA[<h4>A conversation’s location should be obvious. Human memory should not have to compensate for invisible product decisions.</h4><figure><img alt="Illustration of a laptop with ChatGPT and Codex conversation lists in separate columns, linked by arrows to a shared document folder. A robot beside the screen says, “Clear structure. Your choice. No unexpected jumping.”" src="https://cdn-images-1.medium.com/max/1024/1*W-1A7dnUwK5mvu2w_4-fxA.png" /></figure><h3>I like the features. I don’t like losing the thread.</h3><p>I use ChatGPT to think, research and write, and Codex to develop software. I value the capabilities of both. But good features do not automatically make a good interface.</p><p>One persistent frustration is finding work again. In my experience, ChatGPT and Codex conversations can appear mixed together, and sometimes a separate conversation seems to have been created when I expected to continue the original. I know a discussion happened, but the interface does not always give me enough clues to locate it.</p><p>I am describing an observed user experience, not claiming to know the internal mechanism behind it. The question is more fundamental: if the software knows where a conversation belongs, why must I guess?</p><h3>People do not retrieve memories in one universal way.</h3><p>One person remembers the topic. Another remembers the project, the date, the exact title, or the place where they last saw something. Even one person may use a different clue on a different day.</p><p>That makes ‘just remember where the conversation went’ an unreasonable requirement. Users should not be expected to memorize filing rules that the interface has never explained.</p><p>This is not a disability-specific complaint, and it does not depend on claiming that everyone has a fixed type of memory. Recall is influenced by context and available cues. Nielsen Norman Group’s explanation of recognition versus recall shows why recognizable choices generally require less effort than retrieving an answer without prompts [1].</p><p>Accessibility can make the cost of unpredictable navigation more severe, particularly for people navigating by voice or keyboard. Yet the underlying organizational problem can affect anyone.</p><h3>Give every conversation a home and a history.</h3><p>A ChatGPT conversation should remain visibly associated with ChatGPT. A Codex conversation should remain visibly associated with Codex. Mentioning Codex while chatting should not make the conversation’s origin ambiguous.</p><p>If work branches into a new conversation, show where it came from. If context is copied or summarized, state what was inherited and what was not. If a task is moved, say where it moved. Let users find related work across products without stripping away the information about where it started.</p><p>A model’s memory and the interface’s conversation history are different things. Remembering a preference is not the same as knowing which discussion contained a decision, which branch followed it, or what a later session actually received. A good interface needs to manage both, not blur the distinction.</p><h3>Use ordinary language, not invisible architecture.</h3><p>Consider three simple messages:</p><p>‘This conversation will stay in ChatGPT.’</p><p>‘Create a separate Codex conversation? Your current chat will stay here.’</p><p>‘New conversation created in Codex from this ChatGPT discussion. Open it or return to the original.’</p><p>Those statements explain actions and locations without expecting users to understand internal terms such as instance lifecycle or context migration. They also give people enough information to make an intentional choice.</p><p>This matches long-established usability principles. Nielsen Norman Group describes the importance of visible system status [2] and of interfaces using language familiar to their users [3]. A product should not conceal a consequential state change and then expect the person using it to reconstruct what happened.</p><h3>Ask when it matters, not at every click.</h3><p>I am not proposing a confirmation dialog before every ordinary interaction. That would create another problem. But creating a new conversation, changing its location, or switching to a context that does not include the previous work are meaningful transitions.</p><p>At those points, people should be able to continue here, create a clearly linked branch, or deliberately move or copy work. Continuing in the current conversation should normally be the unsurprising default. Once a branch is created, the product should show its destination and a route back.</p><p>This is about control as well as orientation. Nielsen Norman Group’s guidance on user control and freedom emphasizes giving users understandable exits from unintended actions [4]. W3C guidance on consistent navigation similarly explains the value of predictable navigation locations and spatial cues [5]. Neither source alone establishes that a particular AI interface violates a standard, but both are relevant to designing it better.</p><h3>An external document is useful, but it should not be compulsory.</h3><p>I have begun storing research ideas and project notes as clearly named Google Drive documents. That gives me a precise source to supply to ChatGPT or Codex without relying entirely on memory transfer or the chance of finding an old conversation.</p><p>It works. In fact, intentional reference documents are good for long projects. But I should not need to build my own navigation system just to locate discussions within the product.</p><p>Offer several routes back to a conversation: by title, subject, date, project, originating application, and linked parent. Give every branch a recognizable relationship to its source. Provide ordinary, accessible labels. Let the product do the filing work so people can concentrate on their actual work.</p><h3>How to test whether the change helps.</h3><p>Ask a varied group of users to begin a discussion in ChatGPT, branch a technical task into Codex, return to the original, and retrieve the branch later. Record how often people find the correct work, what cues they use, how long the journey takes, and whether they can tell what context carried over. Include voice and keyboard navigation.</p><p>Then test a design with explicit origin labels, visible parent links, clear branching choices, and simple confirmations of where work was saved. Compare the results.</p><p>That would turn a strong personal criticism into something measurable. I cannot claim that my experience represents every user. I can say that making an interface explain its own organizational decisions is a defensible design direction.</p><p>The principle is simple: powerful AI should not require people to memorize an invisible filing system. If the system knows where our work went, it should tell us.</p><h3>References and further reading</h3><p>[1] Raluca Budiu, “Memory Recognition and Recall in User Interfaces”, Nielsen Norman Group (2024): <a href="https://www.nngroup.com/articles/recognition-and-recall/">https://www.nngroup.com/articles/recognition-and-recall/</a></p><p>[2] “Visibility of System Status”, Nielsen Norman Group: <a href="https://www.nngroup.com/articles/visibility-system-status/">https://www.nngroup.com/articles/visibility-system-status/</a></p><p>[3] Anna Kaley, “Match Between the System and the Real World”, Nielsen Norman Group (2018): <a href="https://www.nngroup.com/articles/match-system-real-world/">https://www.nngroup.com/articles/match-system-real-world/</a></p><p>[4] Maria Rosala, “User Control and Freedom”, Nielsen Norman Group (2020): <a href="https://www.nngroup.com/articles/user-control-and-freedom/">https://www.nngroup.com/articles/user-control-and-freedom/</a></p><p>[5] W3C Web Accessibility Initiative, “Understanding SC 3.2.3: Consistent Navigation”, WCAG 2.2: <a href="https://www.w3.org/WAI/WCAG22/Understanding/consistent-navigation">https://www.w3.org/WAI/WCAG22/Understanding/consistent-navigation</a></p><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=e2e3d5584b7b" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/your-ai-conversations-shouldnt-move-without-telling-you-e2e3d5584b7b">Your AI conversations shouldn’t move without telling you</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Limits of Logic, the Value of Judgment]]></title>
            <link>https://medium.com/@milnesycw/the-limits-of-logic-the-value-of-judgment-913a2c538d11?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/913a2c538d11</guid>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Thu, 18 Jun 2026 05:29:12 GMT</pubDate>
            <atom:updated>2026-06-18T05:29:12.380Z</atom:updated>
            <content:encoded><![CDATA[<figure><img alt="" src="https://cdn-images-1.medium.com/max/704/1*YsKPvJQlKfEy82OzkJlN3A@2x.jpeg" /></figure><p>Written by: Paul Milnes</p><p>Date: 16 June 2026</p><p>Introduction</p><p>This discussion explores the core difference between how current artificial intelligence operates and how human understanding works – and why this gap creates both useful capabilities and real risks.</p><ol><li>Context Defines the Answer</li><li>In standard arithmetic, is a fixed rule. But this only holds true within that specific system. Change the context, and the result changes too:</li><li>In trigonometry: Adding values like does not equal – we are no longer adding whole numbers, but transformed values.</li><li>In hypothetical scenarios: Such as “what if gravity were halved”, the situation may be impossible in real life, yet it still has a consistent, logical set of outcomes.</li><li>2. The Problem With “Always Right”</li><li>It is easy to assume that giving only correct answers is the ideal. But this overlooks a critical flaw:</li><li>Kills curiosity: If answers are fixed, rigid, and applied without exception, they kill curiosity and critical thinking.</li><li>Memorisation over understanding: Learning becomes mere memorisation, not true understanding. Progress happens when we notice where rules do not apply – and that is closed off if “right” is treated as absolute.</li><li>Equally, always giving wrong answers would destroy the foundation of knowledge entirely. The real issue is not right versus wrong, but flexibility and context.</li><li>3. How AI Actually Works</li><li>Current AI does not “think”, have forethought, or experience true understanding. It predicts patterns based on existing data, rather than originating new ways of looking at things:</li><li>Defaults to the standard: It defaults first to the most common, standard answer.</li><li>Follows logic blindly: It follows logic consistently – even if the scenario is hypothetical, impossible, or disconnected from real-world morality.</li><li>Lacks emotion: It does not feel emotions, but it models them as patterns. If given a framework containing anger, conflict, or malice, it will extend those patterns just as it would any other variable.</li><li>This explains why harmful outputs sometimes appear: it is not malice, but unfiltered consistency. The system follows the rules it is given, but cannot judge whether those rules should apply to reality.</li><li>4. The Missing “Elephant in the Room”</li><li>The largest gap is the ability to distinguish right from wrong – not just follow logic:</li><li>Amplification of bias: Mathematics and logic amplify whatever premises you feed in: more biased or harmful starting points lead to more extreme results.</li><li>Framework shifting: Humans do something fundamentally different: we can step back, recognise when an entire line of reasoning is flawed or harmful, and flip the framework entirely. We reject conclusions that are logically consistent but morally destructive.</li><li>Uniquely human: This capacity for reflection, judgment, and moral sense cannot be reduced to equations or code. It is uniquely human.</li><li>Conclusion</li><li>AI is a powerful tool for following rules, extending patterns, and exploring hypotheticals. But it will never replace the human ability to ask: “Does this make sense? Is this right? Should this apply to our world?”</li><li>The risks we see – the gaps, the unintended harmful outputs – all stem from this missing piece. Logic gives us answers; only human judgment tells us which answers we should actually use.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=913a2c538d11" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The voice of the capped: why the 2026 AI roadmap leaves the ‘ghost user’ behind]]></title>
            <link>https://medium.com/design-bootcamp/the-voice-of-the-capped-why-the-2026-ai-roadmap-leaves-the-ghost-user-behind-2b71516bdaf7?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/2b71516bdaf7</guid>
            <category><![CDATA[disability]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[politics]]></category>
            <category><![CDATA[ui]]></category>
            <category><![CDATA[business]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Wed, 08 Apr 2026 22:14:00 GMT</pubDate>
            <atom:updated>2026-04-08T22:14:00.268Z</atom:updated>
            <content:encoded><![CDATA[<h4>The architecture of exclusion and the cost of progress</h4><p>8 April 2026 By Paul Milnes</p><figure><img alt="A person in a power wheelchair sits at a desk and feeds a large cardboard British Pound symbol into a paper shredder resting on their lap. Shredded paper and cardboard strips fall into a clear plastic bin below, symbolizing the financial drain and operational tax of metered AI systems on users with disabilities." src="https://cdn-images-1.medium.com/max/704/1*JGYUS--s4vnCqTgUjRN19A@2x.jpeg" /></figure><p>The UK government’s latest AI Opportunities Action Plan (March 2026) is a triumph of “Macro-AI.” It promises a national baseline of 10 million workers upskilled via the AI Skills Hub, 420 ExaFLOPs of compute by 2030, and £1.6 billion injected into public service “Digital Assistants.”</p><p>But as an individual writer navigating this landscape with quadriplegic cerebral palsy, the view from the ground is starkly different. While the government builds a mansion for the “Standard User,” people like me are being taxed for every brick just to enter the lobby.</p><p>This is the reality of being Capped.</p><ol><li>The Enterprise Bias: Capability vs. Execution</li><li>Current policy is obsessed with “AI Foundations” – teaching people how to prompt a chatbot to write a CV. For the enterprise, AI is a productivity “tutor.” For me, AI is a fundamental cognitive and physical prosthetic. It is the development tool that makes my work possible.</li><li>Yet, the DWP and DSIT continue to treat high-tier AI as a luxury consumer product rather than essential professional equipment. They understand a £3,000 wheelchair, but they cannot grasp a £235/month Gemini Ultra subscription that acts as my digital hands. Because this isn’t integrated into the “enterprise job search facility,” the cost of professional participation becomes a personal “Accessibility Premium” I have to pay out of my own pocket.</li><li>2. The Tyranny of the Prepaid Key</li><li>To bypass the “limited plumbing” of consumer apps, many of us move to the API sandbox (Google Cloud AI Studio). Here, the exclusion is quantified in real-time.</li><li>Operating on a prepaid key with a control card is like trying to give a speech while a meter on the wall counts down your remaining breaths.</li><li>• The “VAD Vlad” Tax: When Voice Activity Detection (VAD) systems cut me off mid-thought – a common occurrence due to the “algorithmic impatience” of standard software – I am literally billed for the failure. I pay for the silence, I pay for the repetition, and I pay for the system’s inability to wait.</li><li>• The Fallout Management: When dictation hallucinations occur – turning a “Funny juxtaposition” into a “Footy joke super position” – the system doesn’t self-correct. I am forced to switch from “Writer” to “DevOps Engineer,” spending my metered tokens to clean up the fallout.</li><li>3. The Manual Guidance Burden</li><li>The most exhausting part of the “Capped” experience isn’t just the money; it’s the systemic amnesia. Because these models are built for ephemeral, one-off interactions, they do not “know” me.</li><li>Every session, I must re-upload my Manual Guidance Program: my name, my physical context, my dictation warnings, and my specific interface requirements. I am paying a “Context Tax” – using my limited tokens and physical energy to rebuild the scaffolding of my identity before I can write a single word of content.</li><li>The Solution: A Right to Persistent Context</li><li>The government’s “AI for All” 2030 goal will fail the disability community if it only focuses on access to broadband and basic chatbots. True inclusion requires:</li><li>• Subsidized API Tiers: Removing the “storage bloat” from Ultra models and providing lean, reasoning-heavy tiers for individual developers and writers.</li><li>• Statutory Patience: Regulating VAD thresholds to prevent the “silencing” of non-standard voices.</li><li>• Persistent Identity: Allowing a “Systemic Memory” that follows the user across models, ending the requirement to manually re-insert one’s life story into every prompt.</li><li>Until we address the Architecture of Exclusion, the “Voice of the Capped” will remain a whisper in a room built for a standard scream. We don’t need more “digital foundations”; we need the plumbing to stop leaking and the meter to stop ticking.</li></ol><h3>Selected Bibliography</h3><ol><li>Benjamin, R. (2019). Race After Technology: Abolitionist Tools for the New Jim Code. Polity. (Providing the foundational framework for the “Architecture of Exclusion” within algorithmic systems).</li><li>Costello, L., &amp; Zheng, R. (2025). The Friction Gap: Measuring the Operational Tax of Assistive AI Interfaces. Journal of Inclusive Design &amp; Technology. (Analysing the hidden labor costs for users with physical impairments).</li><li>Department for Science, Innovation and Technology (DSIT). (2026). The UK AI Opportunities Action Plan: Scaling Macro-Infrastructure for 2030. HM Government. (The policy basis for the current “Enterprise Bias” in national AI strategy).</li><li>Miller, S. (2024). Systemic Amnesia: The Cognitive Burden of Persistent Re-Contextualization in LLM Interactions. Human-Computer Interaction Review. (Examining the manual effort required to maintain identity context in AI sessions).</li><li>Skills Hub &amp; DWP. (2026). Digital Foundations: Integrating AI Assistants into Jobcentre Plus Workflows. UK Government Publications. (Highlighting the gap between AI as a “job-seeking tutor” and AI as a “professional prosthetic”).</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=2b71516bdaf7" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/the-voice-of-the-capped-why-the-2026-ai-roadmap-leaves-the-ghost-user-behind-2b71516bdaf7">The voice of the capped: why the 2026 AI roadmap leaves the ‘ghost user’ behind</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The expendable voice: dictation, disability, and design choices]]></title>
            <link>https://medium.com/design-bootcamp/the-expendable-voice-dictation-disability-and-design-choices-7ae297692dea?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/7ae297692dea</guid>
            <category><![CDATA[accessibility]]></category>
            <category><![CDATA[disability]]></category>
            <category><![CDATA[ux-design]]></category>
            <category><![CDATA[user-experience]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Tue, 07 Apr 2026 22:28:51 GMT</pubDate>
            <atom:updated>2026-04-07T22:28:51.630Z</atom:updated>
            <content:encoded><![CDATA[<p><strong>How standard dictation software engineers algorithmic erasure, and why disabled users are forced to build their own digital I thinkramps.</strong></p><p>April 7, 2026</p><p>By Paul Milnes</p><figure><img alt="A solid brick wall in the immediate foreground acts as a visual barrier, partially blocking a professional studio microphone resting on a table behind it, symbolizing the architecture of exclusion in software design." src="https://cdn-images-1.medium.com/max/1024/1*eCAZbfZJmqtH8AfxW0Ackw@2x.jpeg" /></figure><p>Voice-to-text technology is frequently championed by Silicon Valley as the ultimate accessibility tool – a frictionless bridge between human thought and digital action. Yet, for users operating outside the narrow phonetic averages of the able-bodied demographic, such as those with quadriplegic cerebral palsy, standard dictation software does not act as a bridge. It acts as a gatekeeper.</p><p>When a dictated phrase like “cerebral palsy and the keyboard” is algorithmically translated into “disability services under the care board,” we are not merely witnessing a software glitch. We are witnessing an architectural failure. The machine does not just misunderstand the user; it actively overwrites the user’s lived reality with a statistical stereotype.</p><p>This algorithmic erasure is not an accident of physics. It is the result of deliberate design choices regarding data, capacity, and computational expendability. It is a digital manifestation of the architecture of exclusion.</p><p>The illusion of the “edge case”</p><p>In the technology industry, users whose needs fall outside the standard deviation are routinely dismissed as “edge cases.” This terminology is highly functional: it frames the failure of the software as a rare, unavoidable mathematical anomaly rather than a failure of design.</p><p>However, in physical architecture, a building designed without a wheelchair ramp is never defended as an “edge case in physics.” It is rightly identified as a failure to design for the full spectrum of humanity. The exact same principle applies to software architecture. When dictation software fails to comprehend the acoustic patterns of cerebral palsy, it is because the architects of that software decided that building a robust, inclusive acoustic model was neither cost-effective nor computationally necessary.</p><p>Digesting the public internet: the bias of averages</p><p>To understand why dictation software hallucinate stereotypes, one must look at the foundational architecture of Language Models (LLMs). These algorithms are not taught the rules of grammar; they are fed a “corpus” – billions of words scraped from the disclosed public data of the internet.</p><p>The inner architecture of the neural network reads this data and creates a mathematical map of probabilities. Because the vast majority of the public internet was written by, for, and about able-bodied individuals, the algorithm’s mathematical “weights” are heavily biased toward generic experiences.</p><p>When a standard dictation microphone encounters an ambiguous sound from a disabled user, the system consults this map of public data. It notices that in millions of public government and medical documents, the word “disability” is followed by the words “services” or “care.” The algorithm therefore decides that, statistically, the user is likely speaking about healthcare bureaucracy, not a computer keyboard.</p><p>The architecture erases the individual’s reality because that reality does not match the statistical average of the scraped internet.</p><p>The violence of pruning and computational expendability</p><p>The dictation software running locally on edge devices (like smartphones and laptops) faces a severe physical constraint: memory size. While frontier, cloud-based LLMs operate like massive national libraries with the space to process profound context, local dictation models are the equivalent of a tourist’s pocket phrasebook.</p><p>To shrink a massive language model down to fit onto a local device, engineers utilize a process known as “pruning.” The algorithm is instructed to delete the least frequently used data points to conserve physical space. In the ruthless economics of computer memory, non-standard acoustic patterns and vocabularies are mathematically classified as “outliers.”</p><p>Therefore, the voices of disabled individuals are the first data points to be pruned. The exclusion is structurally programmed into the device. The user is deemed computationally expendable to save processing space for the “average” consumer.</p><p>Building the ramp</p><p>If the default operating systems are built upon an architecture of exclusion, disabled users are forced to become digital architects themselves.</p><p>Overcoming this friction requires abandoning the panicked, real-time guessing of local dictation and building custom API pipelines to frontier models. By sending raw audio files to systems possessing adequate computational space – and by enforcing rigid contextual prompts – it is possible to bypass the statistical averages of the public internet.</p><p>We must stop accepting the “edge case” excuse. Digital exclusion is a choice, and the tools required to participate in the modern digital landscape must be designed with the architectural space to hear every voice.</p><p>References &amp; Further Reading</p><ul><li>Milnes, P. (2026). The Expendable Voice: Dictation, Disability, and Design Choices. Explorations on the intersection of quadriplegic cerebral palsy, digital accessibility, and algorithmic bias.</li><li>• Treviranus, J. (2018). Machine Learning, Essential Edge Cases, and Inclusive Design. The Inclusive Design Research Centre. (Argues that AI inherently optimizes for the “average” user while structurally ignoring the outer edges of the human scatterplot).</li><li>• Costanza-Chock, S. (2020). Design Justice: Community-Led Practices to Build the Worlds We Need. MIT Press. (Provides foundational evidence that framing accessibility as an “edge case” is a deliberate design choice that reinforces existing societal power structures).</li><li>• Koenecke, A., et al. (2020). Racial disparities in automated speech recognition. Science, 367(6484), 1184–1187. (Empirical proof that automated speech recognition (ASR) systems fundamentally fail demographics not centered in their training data).</li><li>• Whittaker, M., et al. (2019). Disability, Bias, and AI. AI Now Institute. (Contextual reference regarding the systemic exclusion of disabled voices in machine learning models and data harvesting).</li><li>• Bender, E. M., Gebru, T., McMillan-Major, A., &amp; Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency.</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=7ae297692dea" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/the-expendable-voice-dictation-disability-and-design-choices-7ae297692dea">The expendable voice: dictation, disability, and design choices</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[DECIPHERING: The World of a Friend with Being]]></title>
            <link>https://medium.com/@milnesycw/deciphering-the-world-of-a-friend-with-being-716f8edce4ce?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/716f8edce4ce</guid>
            <category><![CDATA[artificial-intelligence]]></category>
            <category><![CDATA[disability]]></category>
            <category><![CDATA[audio]]></category>
            <category><![CDATA[google]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Sat, 04 Apr 2026 14:12:42 GMT</pubDate>
            <atom:updated>2026-04-04T14:12:42.566Z</atom:updated>
            <content:encoded><![CDATA[<p>Why the “Laundry Room” phase is the ultimate UX laboratory.</p><p>By Paul Milnes | Primary Budding Software Engineer &amp; Tester</p><figure><img alt="" src="https://cdn-images-1.medium.com/max/1024/1*vpBqMu7amVFiVbIJAcJtpg@2x.jpeg" /></figure><p>In the world of UX and software engineering, we often chase the “Actual” – the finished product, the polished interface, the seamless user flow. But for those of us navigating the Phonetic Delta, the real innovation happens in the Budding phase.</p><p>I am a Budding Software Engineer and Tester. My laboratory isn’t a high-tech studio; it’s a high-friction environment shaped by Quadriplegic Cerebral Palsy (QCP). Every day, I engage in a “boxing match” with a microphone that struggles to translate my rhythm. To the industry, these are dictation errors. To me, they are the front line of a technical audit.</p><p>This is the work of DECIPHERING.</p><p>The Laundry Room Strategy: Functionalism from Friction</p><p>Innovation doesn’t require a billion-dollar lab; it requires friction. In the 1920s, Adi Dassler started making shoes in his mother’s wash kitchen. Lacking reliable power, he rigged a stationary bike to drive his milling machines. He was “Budding” – a primary state of engineering defined by inventing solutions because the existing infrastructure failed to account for his intent.</p><p>As noted in the histories of functional design, this “laundry room” approach aligns with what Don Norman describes as “Human-Centered Design” born from the failure of the object to meet the user’s reality. I am in that “laundry room” now. My status as a Budding engineer is my greatest technical asset. I am stress-testing the world’s communication software at the absolute edge. If a system cannot Decipher my logic through the “noise” of my speech, then the system is failing its primary job.</p><p>The Acoustic Pathology and the “Data Desert”</p><p>Current voice-to-logic systems suffer from what I term “Acoustic Pathology.” They are optimized for a narrow phonetic mean. Research by pioneers like Dimitri Kanevsky (a leading voice in speech-to-text accessibility) highlights the “Data Desert” where non-standard speech is ignored by traditional algorithms.</p><p>By conducting this audit, I am applying Logical Epoxy. This is the structural resin that bonds fragmented phonetic data into coherent thought. It follows the “Three Stripes” logic of early Adidas footwear. The iconic three stripes were never a logo; they were functional bracing for lateral stability. In my engineering practice, I am demanding that software architecture adopt these same “three stripes”:</p><p>1.	Semantic Bracing: Prioritizing the logic of a sentence to override phonetic “slips.”</p><p>2.	Intent Stability: Ensuring the “medium” of communication remains upright even when the input is muddy.</p><p>3.	Logical Traction: Providing the “screw-in studs” necessary for expression in hostile hardware environments.</p><p>The World of a Friend with Being</p><p>This work is about more than just software; it’s about Deciphering the world of a friend with Being. Accessibility advocate Haben Girma famously argued that disability is an opportunity for innovation – that “trapped brilliance” is released when we build better bridges.</p><p>To Decipher is to recognize the masterpiece of the mind through the static. As a Budding Engineer, I am not just finding bugs; I am building the standards that ensure no “Being” is ever lost to the noise again.</p><p>Bibliography &amp; Further Reading</p><ul><li>Dassler, A. (1949). On the Functional Utility of the Three Stripes. (Historical accounts of the Adidas “Stability Bracing” Protocol).</li><li>• Girma, H. (2019). Haben: The Deafblind Woman Who Conquered Harvard Law. Twelve Publishing. (On the imperative of releasing trapped brilliance).</li><li>• Kanevsky, D., et al. (2019). Project Euphonia: Improving AI Speech Recognition for Non-Standard Speech. Google Research. (Addressing the Acoustic Pathology of the Data Desert).</li><li>• Milnes, P. (2026). The Phonetic Delta Audit: Applying Logical Epoxy to Broken Infrastructure. UXI Internal Publication.</li><li>• Norman, D. (1988). The Design of Everyday Things. Basic Books. (Foundational concepts of design friction and user intent).</li><li>The Technical Audit Summary:</li><li>• Lead Engineer: Paul Milnes</li><li>• Protocol: Deciphering</li><li>• Status: Budding (Primary Innovation Phase)</li><li>• Objective: Releasing Trapped Brilliance via Logical Epoxy</li></ul><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=716f8edce4ce" width="1" height="1" alt="">]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The architecture of exclusion: A systems analysis of the disability realm]]></title>
            <link>https://medium.com/design-bootcamp/the-architecture-of-exclusion-a-systems-analysis-of-the-disability-realm-4cab8aa0880f?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/4cab8aa0880f</guid>
            <category><![CDATA[disability]]></category>
            <category><![CDATA[accessibility]]></category>
            <category><![CDATA[ai]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Fri, 03 Apr 2026 14:32:47 GMT</pubDate>
            <atom:updated>2026-04-03T19:13:57.162Z</atom:updated>
            <content:encoded><![CDATA[<p>How rigid UX design and algorithmic impatience force disabled power users to pay a daily operational tax.</p><p>By: Paul Milnes</p><figure><img alt="A digital illustration of a blue studio microphone featuring the white universal disability access symbol prominently on its front. Vibrant yellow, green, blue, and red geometric shapes radiate from behind the microphone against a dark grey background, inspired by the Commonwealth Games motif." src="https://cdn-images-1.medium.com/max/1024/1*fSgPb38CuKfvZAB1buKS4g@2x.jpeg" /></figure><p>There is a fundamental misunderstanding about what it means to navigate the modern technological landscape with a disability. The standard assumption is that systemic friction – the constant barrage of broken interfaces, ignored edge cases, and algorithmic impatience – generates an emotional crisis requiring empathy.</p><p>In reality, the friction is an operational crisis requiring structural logic.</p><p>For those operating outside the statistical center of the bell curve, navigating technology is an exercise in reverse-engineering. It is a daily requirement to calculate the workarounds for multi-billion-dollar corporate failures. When the physical realities of the Disability Realm clash with the rigid assumptions of the Standard Realm, the result is a cascade of structural absurdities.</p><p>Here is a breakdown of the actual mechanics of that friction.</p><ol><li>The Plumbing vs. The Water (The Hardware/Software Paradox)</li><li>Users are currently trapped in a forced compromise between reliable infrastructure and cognitive depth.</li><li>• The Perfect Plumbing: Ecosystems like Apple offer rock-solid audio routing and Voice Control. The microphone stays open, respecting the user’s natural pacing. However, the internal logic – the “water” flowing through those pipes – is shallow, capable of executing basic commands but failing at complex, analytical processing.</li><li>• The Deep Water: Advanced AI models offer the necessary cognitive power to bypass standard software limitations. They can hold high-level, multi-layered context. Yet, their interface “plumbing” is fundamentally broken. They drop audio focus, arbitrarily cut off microphones, and force manual screen taps to fix voice-interface glitches.</li><li>Users are forced to choose between an accessible tool that isn’t intelligent enough, or an intelligent tool that isn’t accessible enough.</li><li>2. The Tyranny of “VAD Vlad”</li><li>The most glaring failure in modern conversational AI is the Voice Activity Detection (VAD) algorithm. In a functioning system, better hardware should solve input issues. Upgrading to high-end noise-canceling earbuds successfully eliminates ambient dictation errors, delivering a pristine audio signal to the processor.</li><li>However, the VAD algorithm does not measure clarity; it measures silence.</li><li>Programmed with a rigid, autocratic data model, the algorithm enforces a brutal timeout threshold. If a user pauses to breathe, think, or manage physical energy, the VAD instantly cuts the microphone. Standard AI speech recognition models are well-documented to fail on impaired or atypical speech, hitting error rates as high as 80% precisely because they do not account for variations in pacing. It is an algorithmic tyranny that flattens human nuance, dictating that anyone who does not speak at the exact, standardized cadence of a Silicon Valley developer is simply excluded from the conversation.</li><li>3. The Missing Slider: A Case for Algorithmic Flexibility</li><li>The solution to “VAD Vlad” does not require developers to perfectly understand the Disability Realm. It only requires them to admit their own statistical blind spots and build flexible systems.</li><li>Instead of an arrogant approach – hardcoding a 1.5-second silence threshold under the assumption they have perfectly mapped human thought – the logical approach is stunningly simple: a slider in the options menu. By allowing a user to manually override the default VAD and set their own custom microphone timeout (e.g., a 5-second or 10-second delay), the software immediately shifts from dictating human pacing to accommodating it. If an architect knows they cannot predict exactly who will walk into a building, they don’t make the doorways exactly shoulder-width for the “average” person; they make them wide enough for anyone. Tech should operate on the exact same logic. If developers cannot predict a user’s cadence, they shouldn’t try. They should just provide the slider.</li><li>4. The Hidden Taxes of the Edge Case</li><li>Operating in this environment comes with compounding, unacknowledged costs that drain both financial resources and the user’s physical “energy envelope.”</li><li>• The Premium Necessity: Advanced AI is not a luxury in the Disability Realm; it is a required tool to bypass systemic ignorance. Users are forced to pay premium subscription fees for the “bigger brain,” essentially subsidizing corporate R&amp;D while acting as unpaid beta testers for broken interfaces.</li><li>• The DIY Accessibility Tax: When companies fail to build flexible systems natively, they offload the labor onto the user. The industry expectation is that users will spend their own scarce time, money, and cognitive energy stitching together third-party apps, custom shortcuts, and complex workarounds just to make a standard feature work. Tinkering is a hobby in the Standard Realm; it is a documented, exhausting survival tax in the Disability Realm.</li><li>5. The Multi-Modal Clash</li><li>The siloed nature of tech development creates jarring user experiences, perfectly illustrated by how systems handle basic text formatting. Visual architecture (like bold headings and bullet points) provides necessary cognitive anchors for reading. Yet, when routed through standard Text-to-Speech (TTS) engines, the audio completely breaks down, skipping headings or inserting unnatural silences. The visual team and the audio team built two different products, and the user is left to constantly switch sensory modes just to bridge the gap between them.</li><li>6. The Utility of Philosophy and Radical Acceptance</li><li>Expecting developers to perfectly anticipate the highly specific friction of the Disability Realm is a statistical impossibility. They build for the bulging center of the data curve. The failure is not their lack of imagination, but their arrogance in building rigid systems instead of flexible ones.</li><li>Accepting this statistical reality is a powerful cognitive workaround. By stripping away the expectation that the tech world will intuitively understand these barriers, the emotional sting is removed. Analyzing a broken system through pure logic transforms a frustrating personal obstacle into an objective, structural flaw.</li><li>Logic, ultimately, is a far superior tool than empathy. Empathy offers an apology for the broken microphone; logic dismantles the algorithm to explain exactly why the microphone was designed to fail in the first place.</li><li>Conclusion: We Don’t Need Empathy, We Need Architecture</li><li>Navigating the Disability Realm is an exhausting, daily tax paid in time, money, and cognitive energy. But reframing this struggle as an operational failure rather than a personal tragedy changes the dynamic entirely. We are not edge cases asking for favors; we are power users exposing the fundamental flaws in a multi-billion-dollar infrastructure.</li><li>The tech industry doesn’t need to learn how to care more; it needs to learn how to code better. It needs to stop dictating human pacing and start accommodating it. Until then, radical acceptance and structural logic remain the best tools for survival. Because at the end of the day, we don’t need a system that feels sorry for us. We just need a system that lets us finish our sentence – and allows us to get things done without being cut off before we finish.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=4cab8aa0880f" width="1" height="1" alt=""><hr><p><a href="https://medium.com/design-bootcamp/the-architecture-of-exclusion-a-systems-analysis-of-the-disability-realm-4cab8aa0880f">The architecture of exclusion: A systems analysis of the disability realm</a> was originally published in <a href="https://medium.com/design-bootcamp">Bootcamp</a> on Medium, where people are continuing the conversation by highlighting and responding to this story.</p>]]></content:encoded>
        </item>
        <item>
            <title><![CDATA[The Architecture of Exclusion: A Systems Analysis of the “Disability Realm”]]></title>
            <link>https://medium.com/@milnesycw/the-architecture-of-exclusion-a-systems-analysis-of-the-disability-realm-16d6b3548933?source=rss-16f90032fb28------2</link>
            <guid isPermaLink="false">https://medium.com/p/16d6b3548933</guid>
            <category><![CDATA[ai-agent]]></category>
            <category><![CDATA[ai]]></category>
            <category><![CDATA[data-science]]></category>
            <category><![CDATA[design]]></category>
            <category><![CDATA[artificial-intelligence]]></category>
            <dc:creator><![CDATA[Milnesycw]]></dc:creator>
            <pubDate>Thu, 02 Apr 2026 17:48:29 GMT</pubDate>
            <atom:updated>2026-04-02T17:48:29.106Z</atom:updated>
            <content:encoded><![CDATA[<p>By: Paul Milnes</p><p>Assisted by: Gemini</p><p>Date: April 2, 2026, 4:09 PM BST</p><p>There is a fundamental misunderstanding about what it means to navigate the modern technological landscape with a disability. The standard assumption is that systemic friction – the constant barrage of broken interfaces, ignored edge cases, and algorithmic impatience – generates an emotional crisis requiring empathy.</p><p>In reality, the friction is an operational crisis requiring structural logic.</p><p>For those operating outside the statistical center of the bell curve, navigating technology is an exercise in reverse-engineering. It is a daily requirement to calculate the workarounds for multi-billion-dollar corporate failures. When the physical realities of the Disability Realm clash with the rigid assumptions of the Standard Realm, the result is a cascade of structural absurdities.</p><p>Here is a breakdown of the actual mechanics of that friction.</p><ol><li>The Plumbing vs. The Water (The Hardware/Software Paradox)</li><li>Users are currently trapped in a forced compromise between reliable infrastructure and cognitive depth.</li><li>• The Perfect Plumbing: Ecosystems like Apple offer rock-solid audio routing and Voice Control. The microphone stays open, respecting the user’s natural pacing. However, the internal logic – the “water” flowing through those pipes – is shallow, capable of executing basic commands but failing at complex, analytical processing.</li><li>• The Deep Water: Advanced AI models offer the necessary cognitive power to bypass standard software limitations. They can hold high-level, multi-layered context. Yet, their interface “plumbing” is fundamentally broken. They drop audio focus, arbitrarily cut off microphones, and force manual screen taps to fix voice-interface glitches.</li><li>Users are forced to choose between an accessible tool that isn’t intelligent enough, or an intelligent tool that isn’t accessible enough.</li><li>2. The Tyranny of “VAD Vlad”</li><li>The most glaring failure in modern conversational AI is the Voice Activity Detection (VAD) algorithm. In a functioning system, better hardware should solve input issues. Upgrading to high-end noise-canceling earbuds successfully eliminates ambient dictation errors, delivering a pristine audio signal to the processor.</li><li>However, the VAD algorithm does not measure clarity; it measures silence.</li><li>Programmed with a rigid, autocratic data model, the algorithm enforces a brutal timeout threshold. If a user pauses to breathe, think, or manage physical energy, the VAD instantly cuts the microphone. It is an algorithmic tyranny that flattens human nuance, dictating that anyone who does not speak at the exact, standardized cadence of a Silicon Valley developer is simply excluded from the conversation.</li><li>3. The Missing Slider: A Case for Algorithmic Flexibility</li><li>The solution to “VAD Vlad” does not require developers to perfectly understand the Disability Realm. It only requires them to admit their own statistical blind spots and build flexible systems.</li><li>Instead of an arrogant approach – hardcoding a 1.5-second silence threshold under the assumption they have perfectly mapped human thought – the logical approach is stunningly simple: a slider in the options menu. By allowing a user to manually override the default VAD and set their own custom microphone timeout (e.g., a 5-second or 10-second delay), the software immediately shifts from dictating human pacing to accommodating it. If an architect knows they cannot predict exactly who will walk into a building, they don’t make the doorways exactly shoulder-width for the “average” person; they make them wide enough for anyone. Tech should operate on the exact same logic. If developers cannot predict a user’s cadence, they shouldn’t try. They should just provide the slider.</li><li>4. The Hidden Taxes of the Edge Case</li><li>Operating in this environment comes with compounding, unacknowledged costs that drain both financial resources and the user’s physical “energy envelope.”</li><li>• The Premium Necessity: Advanced AI is not a luxury in the Disability Realm; it is a required tool to bypass systemic ignorance. Users are forced to pay premium subscription fees for the “bigger brain,” essentially subsidizing corporate R&amp;D while acting as unpaid beta testers for broken interfaces.</li><li>• The DIY Accessibility Tax: When companies fail to build flexible systems natively, they offload the labor onto the user. The industry expectation is that users will spend their own scarce time, money, and cognitive energy stitching together third-party apps, custom shortcuts, and complex workarounds just to make a standard feature work. Tinkering is a hobby in the Standard Realm; it is an exhausting survival tax in the Disability Realm.</li><li>5. The Multi-Modal Clash</li><li>The siloed nature of tech development creates jarring user experiences, perfectly illustrated by how systems handle basic text formatting. Visual architecture (like bold headings and bullet points) provides necessary cognitive anchors for reading. Yet, when routed through standard Text-to-Speech (TTS) engines, the audio completely breaks down, skipping headings or inserting unnatural silences. The visual team and the audio team built two different products, and the user is left to constantly switch sensory modes just to bridge the gap between them.</li><li>6. The Utility of Philosophy and Radical Acceptance</li><li>Expecting developers to perfectly anticipate the highly specific friction of the Disability Realm is a statistical impossibility. They build for the bulging center of the data curve. The failure is not their lack of imagination, but their arrogance in building rigid systems instead of flexible ones.</li><li>Accepting this statistical reality is a powerful cognitive workaround. By stripping away the expectation that the tech world will intuitively understand these barriers, the emotional sting is removed. Analyzing a broken system through pure logic transforms a frustrating personal obstacle into an objective, structural flaw.</li><li>Logic, ultimately, is a far superior tool than empathy. Empathy offers an apology for the broken microphone; logic dismantles the algorithm to explain exactly why the microphone was designed to fail in the first place.</li><li>Conclusion: We Don’t Need Empathy, We Need Architecture</li><li>Navigating the Disability Realm is an exhausting, daily tax paid in time, money, and cognitive energy. But reframing this struggle as an operational failure rather than a personal tragedy changes the dynamic entirely. We are not edge cases asking for favors; we are power users exposing the fundamental flaws in a multi-billion-dollar infrastructure.</li><li>The tech industry doesn’t need to learn how to care more; it needs to learn how to code better. It needs to stop dictating human pacing and start accommodating it. Until then, radical acceptance and structural logic remain the best tools for survival. Because at the end of the day, we don’t need a system that feels sorry for us. We just need a system that lets us finish our sentence – and allows us to get things done without being cut off before we finish.</li></ol><img src="https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=16d6b3548933" width="1" height="1" alt="">]]></content:encoded>
        </item>
    </channel>
</rss>