Write Inclusive, People-First Language About AI - Alexandra Klepper  🔗 **Source:** [Write Inclusive, People-First Language About AI - Alexandra Klepper](https://www.youtube.com/watch?v=wH3H8ko43gM) > [!note]+ Document Notes > tags: [accessibility, people-first-language, anthropomorphism, ai-documentation, technical-writing, ai-agents] > author: Alexandra Klepper > source: Write the Docs (YouTube) > reference: https://www.youtube.com/watch?v=wH3H8ko43gM > date created: 2026-09-23 10:27 > > > [!info] Generated analysis — not the author's words. Sections below are machine-written > > from the captured source. Only text inside `> [!quote]` blocks is verbatim from the author. > > Do not quote or attribute any other line to the named author. > > # 🏷️ Title: Write inclusive, people-first language about AI - Alexandra Klepper > <!-- deep-read:sections start --> ## 🧭 Table of contents - [📝 Summary](#-summary) - [📖 Main content](#-main-content) - [🗒️ Notes](#️-notes) - [📚 References](#-references) - [🗣️ Paraphrase](#️-paraphrase) - [📰 Newsletter](#-newsletter) - [✅ Action items](#-action-items) - [🧪 Hypotheses](#-hypotheses) - [📊 Methodology](#-methodology) - [❓ Questions](#-questions) - [🔮 Future research](#-future-research) - [💡 Implications](#-implications) - [🧠 Learning styles assessment](#-learning-styles-assessment) - [📊 Accessibility assessment](#-accessibility-assessment) - [🌍 DEIB assessment](#-deib-assessment) - [✅ Logical validity](#-logical-validity) - [🛡️ DARVO analysis](#️-darvo-analysis) - [🎯 Conclusion](#-conclusion) - [🔄 Refactor appendix](#-refactor-appendix--write-inclusive-people-first-language-about-ai---alexandra-klepper) ## 📝 Summary - **What happened.** Write the Docs published a 35:09 recording of Alexandra Klepper's talk from Write the Docs Berlin 2026, uploaded 2026-09-22. Klepper is a senior technical writer on Google Chrome. The auto-transcript renders her name as "Alexander Clipper" throughout; quotes below preserve transcript spelling. - **The claim.** Accessibility, as the W3C Web Accessibility Initiative defines it, is a practice for people with disabilities. AI systems and agents benefit from accessible markup because they read the same semantics, but they don't need accessibility and can't have disabilities. Writing "AI needs accessibility" or "AI will solve accessibility" detaches the word from the people it exists for. - **The second claim.** Anthropomorphic language in AI writing (biology, neurons, feeding, "AI voice," "the model thinks") trains readers to treat probabilistic software as a person. Klepper proposes people-first language for AI docs: center the human user, describe the machine as a tool, and swap biological metaphors for construction metaphors where the field hasn't already locked the term in. - **Why it matters.** Klepper cites 16% of the world living with a disability and 95.9% of homepages failing an automated accessibility check. If accessibility gets reframed as a machine-readability feature, the budget and attention follow the machines. The description names the stake plainly: the people most likely to be left behind in the shift to AI are people with disabilities. - **Key stakeholders.** Technical writers and docs teams; Google Chrome's Modern Web Guidance project (called out by name as her own team's work); Anthropic, whose paper title "On the Biology of a Large Language Model" is her lead example; Google Cloud's AI Studio docs; PMs and engineers newly interested in accessibility because agents use the accessibility tree; the W3C WAI; WebAIM. - **Risks and tradeoffs.** Klepper admits the risk herself at 00:12:26: insisting that accessibility is for humans may alienate PMs who are excited about it for the first time because AI needs it. The talk also rests on a definitional argument (accessibility is not a problem, so it can't be solved) that an executive can sidestep by saying "remediation" instead. Several statistics arrive without a named source in the transcript. - **Bottom line.** The core distinction holds: "benefits from" is not "needs." Docs teams can act on the three concrete edits she demonstrates (biology to construction, drop "AI voice," state guidance rather than guarantee) without waiting for a framework. The bigger people-first-language-for-AI framework is a proposal, not a published standard. ## 📖 Main content **The argument in one line.** Accessibility is defined by who it serves, not by who reads the output, so an agent parsing the accessibility tree is a beneficiary and not a stakeholder. **Definitions she anchors on.** At 00:06:31 Klepper takes the W3C's framing: accessibility is tools and technologies so that people with disabilities can use the web. She pulls two consequences from it. First, accessibility is not a solution and not a checklist, because it is a practice rather than a problem with an end state (00:06:31 to 00:07:17, restated at 00:20:33). Second, the access is specifically for people with disabilities, which is the premise every later claim depends on. For AI, at 00:13:10, she uses the textbook definition (a field of computer science building machines that perform tasks that typically require human intelligence) and concedes the acronym now covers fields, features, platforms, and machines indiscriminately, so readers have to infer which sense applies. **Determinism, probabilism, agents.** At 00:13:52 she splits classic code (deterministic, one answer, "algebra") from large language models (probabilistic, many possible outputs, "calculus"). The split does two jobs later. It explains why coding agents can't guarantee accessible output every time (00:27:44 to 00:28:29), and it sorts accessibility testing into automatable checks (color contrast has exact math, 00:19:05) and judgment calls (complex language) where probabilistic tools can help. Her agent definition at 00:15:21 has four parts: a model for reasoning and language, rules for consistency, memory for context, and tools for taking action. Action is what separates an agent from a chat interface, and she frames tool use as a way to add determinism back into a probabilistic system. **The intersection, and where she draws the line.** At 00:16:06 she grants the fact that motivates the whole "accessibility for AI" trend: when websites are accessible, agents understand them better. Then she refuses the inference. AI benefits from accessible design because accessibility is a practice designed for people with disabilities. The benefit is real; the need is not. The Mona Lisa passage at 00:16:51 to 00:18:20 makes the same point by analogy: a replica, however skilled, doesn't carry the original's value, and an AI that replicates human behavior isn't a human with human needs. **Where AI does help.** She lists concrete uses at 00:18:20 to 00:19:48: automated compliance testing for deterministic criteria, probabilistic evaluation of complex language, and generative support for executive functioning (task prioritization, information architecture, distraction reduction). This is the section that keeps the talk from being a blanket rejection, and she says so ("I promise I'm not just a hater"). **Anthropomorphism, with three examples.** Defined at 00:21:15 as attributing human actions to inanimate sources. Example one, 00:21:58: Anthropic's paper "On the Biology of a Large Language Model," about the Claude 3.5 Haiku model. Her objection is that biology is the science of life and AI simulates life without having it. Her proposed fix at 00:24:08 is a one-word swap to "construction." She also concedes at 00:24:08 that neurons and neural networks are locked in after decades of use and she won't try to rename them. Example two, 00:24:51: a Google Cloud tutorial titled "How to build a custom text-to-speech AI voice generator in AI Studio." Removing "AI voice" shortens the heading and stops implying the machine has a voice rather than a replication of one. Example three, 00:27:02 to 00:28:29: her own team's Modern Web Guidance project, which ships skills for coding agents. She reports Chrome engineer Jeffrey Yasskin's framing (the transcript spells it "Yaskin") that the project is harm reduction rather than a solution, then shows corrections to the docs so they say "guidance," not "guarantee." **The people-first proposal.** At 00:26:19 she borrows the term from disability language, where it puts the person before the disability, and repurposes it: in AI docs, put the human before the non-human user (agent, model, bot, scraper). The operating rules she gives at 00:28:29 to 00:29:58 are: use the W3C definition when asked what accessibility is; don't equate humans, AI, and agents one to one; read your text and ask whether it sounds written for humans; optimize docs for the human users of agents rather than for agents; write semantic HTML for assistive technology and let agents ride along. **Contested or unverified in this source.** The 16% disability figure (00:08:00) matches the WHO's published estimate but the transcript names no source. The 95.9% homepage failure rate (00:08:00) is attributed to WebAIM; WebAIM publishes a yearly Million report, and the exact year isn't stated. The survey of 1,600 professionals (00:09:31), with 37% EAA compliance, 75% saying accessibility affects revenue growth, and 86% expecting AI to solve it, has no named publisher in the transcript. The $11 trillion spending-power figure is attributed to the World Economic Forum and the $4.75 trillion German GDP for 2025 to no one (00:10:14 to 00:10:57). The closing quote at 00:29:58 ("Where a human judges, a model correlates...") is read aloud as "one last quote" without attribution. The claim that "better data... does not make a better model" (00:23:25) overstates a real point about data quality being one factor among several. And the "people have started to care less about other people" claim at 00:18:20 is a causal assertion with no evidence offered. ## 🗒️ Notes #### ⏱️ [00:00:00](https://www.youtube.com/watch?v=wH3H8ko43gM&t=0) Host introduction and why words matter - The host introduces Klepper as a senior technical writer for Google Chrome, joining from New York, returning to the Write the Docs stage after mentoring in the program. The transcript renders her name "Alexander Clipper." - Opening thesis at 00:00:42: language uplifts or harms, and grandiose product-marketing claims about generative AI leak into documentation. - Two claims she promises to dissect, 00:01:26: "AI is going to solve accessibility" and "AI needs accessibility." - 00:01:26 to 00:02:54: Orwell's *1984* as the frame. A controlled, simplified language designed to prevent critical thinking. When words built to support marginalized people lose their meaning, the people lose protection. > [!quote] [00:00:42] > The words that we choose matters. > [!quote] [00:02:09] > When the words that we create to support marginalized people are taken over and detached from the original meaning like the word accessibility, we lose some of our humanity. #### ⏱️ [00:03:38](https://www.youtube.com/watch?v=wH3H8ko43gM&t=218) Klepper's path: burnout, Chrome, a glossary, a proposal - Three years ago she returned from burnout leave to the Chrome team as it began talking about large language models, and took the AI documentation role partly to question the language being used (00:03:38 to 00:04:22). - Scale argument at 00:04:22: millions of people read what Google's writers publish, so word choice at that scale is a front-line responsibility. - Six years ago she gave a Write the Docs Portland talk on accessible documentation; thesis then was small technical changes by writers support people. The same move now applies to supporting humanity over machines (00:04:22 to 00:05:05). - Steps taken, 00:05:05 to 00:05:47: a glossary for web AI to fix exact terms; then a proposal titled "People-first accessibility language for AI and agents," written after "enough frustration with AI accessibility bros." - Next steps, 00:05:47: speak at Write the Docs (done), gather feedback, convene a group to build a larger people-first framework. She references Kate Wilcox's earlier talk on a "craft methodology for moonshot projects." #### ⏱️ [00:06:31](https://www.youtube.com/watch?v=wH3H8ko43gM&t=391) Defining accessibility; 16% and 95.9% - W3C definition, 00:06:31: tools and technologies so that people with disabilities can use them. Two consequences: accessibility is not a solution and not a checklist (00:06:31 to 00:07:17), and it exists for people with disabilities. - Audience poll: what percentage of the world has a disability? Someone shouts "five." Answer at 00:08:00: 16%, up 1% since 2018, which she ties to the pandemic. Source not named in the transcript; the figure matches the WHO's published estimate. - WebAIM (transcript: "Web AIM") has tested the top 1 million homepages for 27 years per her framing. 95.9% of homepages failed an accessibility test (00:08:00 to 00:08:45). - WebAIM's stated takeaway: shift to simpler systems and refine best practices. She argues third-party frameworks and AI-assisted "vibe coding" push the other way, toward larger, more complex pages (00:08:45). > [!quote] [00:06:31] > They define accessibility as tools and technologies so that people with disabilities can use them. > [!quote] [00:08:00] > Okay, 16% of the world has a disability. This is an increase of 1% since 2018. > [!quote] [00:08:45] > We're moving in the wrong direction. #### ⏱️ [00:09:31](https://www.youtube.com/watch?v=wH3H8ko43gM&t=571) The business case and ableism - 00:09:31: she states empathy doesn't move executives and says she's unconvinced the business argument works either, then gives it anyway "in case I'm wrong." - Survey of 1,600 professionals (publisher not named): 37% report compliance with the European Accessibility Act; 75% say accessibility directly affects revenue growth; 86% think AI will solve it (00:09:31 to 00:10:14). - World Economic Forum estimate: $11 trillion global spending power for people with disabilities and their families. Comparison: Germany, third-largest GDP, $4.75 trillion in 2025. The disability figure is about double (00:10:14 to 00:10:57). - 00:11:42: the belief that AI will solve accessibility is itself the problem, because it forgets accessibility is for people, AI is built by people, and people are imperfect. - Ableism, 00:11:42 to 00:12:26: internalized bias about disability is built into machines because people build machines. She names the tradeoff at 00:12:26: prioritizing humans may alienate PMs who are excited about accessibility for the first time because AI needs it. > [!quote] [00:10:14] > Really developers and executives want to make accessibility somebody's else's problem. So somebody else can solve it. > [!quote] [00:11:42] > The bias is built into machines because machines are built by people. #### ⏱️ [00:13:10](https://www.youtube.com/watch?v=wH3H8ko43gM&t=790) Defining AI: determinism, probabilism, agents - AI defined as a field of computer science building machines that perform tasks typically requiring human intelligence. The acronym now names the field, features, platforms, and machines, so it's "essentially meaningless" without context (00:13:10 to 00:13:52). - Determinism vs probabilism at 00:13:52: classic code always returns one answer; an LLM picks among many possible outputs based on context and can answer differently tomorrow. That flexibility is the power and also the source of a nonzero error rate with consequences (00:14:35). - Agent anatomy, 00:14:35 to 00:15:21: model (reasoning, language), rules (consistency), memory (context), tools (action on the user's behalf). Action is the differentiator. An agent can return an exact URL, which she frames as adding determinism to probabilism. - 00:15:21 to 00:16:06: agents work best when they understand the contents and context of resources, and accessible websites help them do that. She anticipates the objection this sets up. > [!quote] [00:13:52] > Determinism is classic coding. It's algebra. 2 + 2 equals 4. There's always one answer. Probabilism is large language models. It's calculus. > [!quote] [00:15:21] > That action is what separates general AI from an agent. #### ⏱️ [00:16:06](https://www.youtube.com/watch?v=wH3H8ko43gM&t=966) The intersection, the line, and the Mona Lisa - 00:16:06: "Didn't you just say agents need websites to be accessible?" Not exactly. AI benefits from accessible design because accessibility is a practice designed for people with disabilities. Benefit is not need. - 00:16:51: the tech world describes AI as a reflection or replication of humanity, then talks about it as parallel to humanity with human needs and thought processes. - Mona Lisa analogy, 00:16:51 to 00:18:20: Leonardo's students made copies immediately; you can own a print or a lifelike repaint, but the original is in the Louvre (transcript: "Louv"). A replication doesn't duplicate the original's value. - 00:18:20: her diagnosis is that people have started to care less about other people in service of a "greater version of us." Unsupported as stated; see Logical validity. - 00:18:20 to 00:19:48: where AI helps. Automated testing for deterministic criteria (color contrast has exact math), probabilistic evaluation of complex language, generative support for executive functioning: task prioritization, information architecture, distraction reduction. > [!quote] [00:16:06] > When I say that AI benefits from accessible design for humans, I mean it benefits because accessibility is a practice designed for people with disabilities. > [!quote] [00:17:36] > A replication can't duplicate the value of the original or otherwise you'd be fabulously wealthy. > [!quote] [00:19:05] > Color contrast has an exact math. #### ⏱️ [00:19:48](https://www.youtube.com/watch?v=wH3H8ko43gM&t=1188) AI has no needs; anthropomorphism and the "biology" paper - 00:19:48: she cites "dozens of articles" claiming bots are people too and need accessibility. Her response, repeated for emphasis: AI is not human, cannot have human needs, so does not need accessibility. - 00:20:33: AI cannot have disabilities; a model can imitate one, and that's all it is. Accessibility can't be solved because neither accessibility nor disability is a problem to solve. It's how we support and uplift people with disabilities. - 00:20:33 to 00:21:15: people who don't understand accessibility as a framework, or don't respect people with disabilities, can't build accessible experiences, however capable the generative tool. - 00:21:15: "get on board or get left behind" pressure; AI therapy startups as the example of writing that wants you to believe you're talking to a human. Anthropomorphism defined as attributing human actions to inanimate sources. - 00:21:58 to 00:22:40: Anthropic's paper "On the Biology of a Large Language Model," about the Claude 3.5 Haiku model. Biology is the science of life; AI simulates life and doesn't have it. She guesses the word was chosen to make an unfamiliar system feel familiar. > [!quote] [00:19:48] > AI is not human. AI cannot have human needs. > [!quote] [00:20:33] > Instead, accessibility is how we support and uplift people with disabilities to improve their experience and access. > [!quote] [00:21:58] > Biology is the science of life. #### ⏱️ [00:23:25](https://www.youtube.com/watch?v=wH3H8ko43gM&t=1405) Feeding, construction, and "AI voice" - 00:22:40 to 00:23:25: biological language is everywhere in AI: neurons, neural networks, feeding. Feeding is a life function that also names how a model is given data. - 00:23:25: better food doesn't make a human healthy on its own; better data is one piece of how a model works and doesn't by itself make a better model. This overstates the case; data quality is a major lever. - 00:24:08: she concedes the field won't rename neurons or neural networks after decades of use. But one word in one title can change: "On the Construction of a Large Language Model." Small technical changes separate machines from humans. - 00:24:51: technology should serve users, not the other way around. Documentation example: Google Cloud tutorial "How to build a custom text-to-speech AI voice generator in AI Studio." She sets aside the missing hyphens and targets "AI voice." - 00:25:35: voices are human; Google Assistant offers many "voices" varying in language, accent, and tone, all replications of human voices. Cutting "AI voice" shortens the heading and removes the overlap with "AI language." - 00:26:19: machines reflect their builders' biases; they're collaboration tools, not replacements; they can't care or feel about people, however convincing they sound. > [!quote] [00:23:25] > Better data is just one tiny piece of how a model works. It does not make a better model. > [!quote] [00:24:08] > the chances of creating new words for neurons and neural networks, zero. Not going to happen. Not going to try. > [!quote] [00:24:51] > And users should expect that technology serves their needs, not the other way around. > [!quote] [00:25:35] > These are replication of human voices. > [!quote] [00:26:19] > They are collaboration tools, not human replacements. #### ⏱️ [00:26:19](https://www.youtube.com/watch?v=wH3H8ko43gM&t=1579) People-first language and Modern Web Guidance - 00:26:19 to 00:27:02: people-first language comes from disability contexts, where personhood is primary and distinct from disability. Repurposed here: humans over non-human users (agents, models, bots, scrapers). - 00:27:02: side note that "user" implies human and she'd like a different word. Both people with disabilities and non-disabled people have had their humanity sidelined in AI language. - 00:27:02 to 00:27:44: she calls out her own team. Chrome's Modern Web Guidance (MWG) ships skills that give coding agents current browser best practices so developers build better websites. It raised the question of what makes a website accessible, which she says is hard to answer. - 00:27:44: Chrome engineer Jeffrey Yasskin (transcript: "Yaskin") frames MWG as harm reduction, not a solution: if people will use AI to generate websites, help them do it better. - 00:27:44 to 00:28:29: LLMs are probabilistic, so coding agents can't produce accessible output 100% of the time. She wishes the docs said so and shows corrections she'll share with the writer, whom she names as Jeremy. > [!quote] [00:27:02] > Side note, the word user kind of implies human. Personally, I don't like it. > [!quote] [00:27:44] > Jeffrey Yaskin, who is a Chrome engineer, explained that MWG is harm reduction rather than solutionoriented. > [!quote] [00:28:29] > We should s say that we are providing the guidance but that we are not guaranteeing it. We must choose our words carefully. #### ⏱️ [00:29:14](https://www.youtube.com/watch?v=wH3H8ko43gM&t=1754) Recommendations, the pipe, and the closing quote - Rules, 00:28:29 to 00:29:58: answer "what is accessibility" with the W3C definition; don't treat humans, AI, and agents as a one-to-one equivalence; some swaps are easy (biology to construction), others take thought; read the text and ask whether it sounds written for humans. - Docs rule: understand what users need, support their assistive technology, and maybe agents. Don't optimize docs for agents; optimize for the agents' human users. Semantic HTML serves assistive tech and agents both. - 00:29:58: Magritte's pipe painting as the closing image (not named in the transcript; "This is not a pipe. This is a painting of a pipe."). An AI model is a replication of people, built on its builders' interpretations and flaws. - Closing quote, read without attribution: judges vs correlates, evaluates vs predicts, engages with the world vs engages with the distribution of words. - 00:30:42: thanks to Darcy Lima, Justine Owens, and the documentation accessibility group; seven years of work on accessible docs. Slides via QR code on screen; no slide link in the description. > [!quote] [00:29:14] > Don't optimize your docs for agents. Optimize the docs for the human users of the agents. > [!quote] [00:29:14] > When you center humans, agent can better support them too. > [!quote] [00:29:58] > An AI model is not a person. It is a replication of people based on the interpretation of the humans that built it and all of their flaws. > [!quote] [00:29:58] > Where a human judges, a model correlates. Where a human evaluates, a model predicts. Where a human engages with the world, a model engages with the distribution of words. #### ⏱️ [00:30:42](https://www.youtube.com/watch?v=wH3H8ko43gM&t=1842) Q&A: the word "agent," honesty about AI use, customization - Host connects the talk to an earlier speaker's (transcript: "Kel") call to decenter machines and recenter humans as active actors, and praises the nomenclature work (00:30:42 to 00:31:25). - Emily, a support agent, asks what human agents should call themselves now that "agent" means software. Klepper defers: "let me get back to you" (00:31:25 to 00:32:09). - Deepak asks about AI messaging that promises determinism and probabilism at once. Answer at 00:32:51: be honest and transparent whenever AI is used, say what kind; most users don't care, so scale the detail to the audience. - Elizabeth asks whether automation tools and agents can support individualized needs. Klepper asks to read the question herself ("I do better with visual"), then answers: it depends; big models learn you over time until an update wipes context; building your own is heavy; sharing your own context up front helps assistants help you (00:33:34 to 00:34:18). - 00:34:18: returning to Emily's question, her recommendation for human agents is "customer advocate." The recording ends on applause. > [!quote] [00:32:51] > Think about your audience. > [!quote] [00:34:18] > Advocate, customer advocate. That is my recommendation. ## 📚 References The transcript itself contains no links; the following come from the video description or are primary sources for named entities. - [Write inclusive, people-first language about AI - Alexandra Klepper](https://www.youtube.com/watch?v=wH3H8ko43gM): the source video, 35:09, uploaded 2026-09-22 by Write the Docs. - [Write the Docs Berlin 2026](https://www.writethedocs.org/conf/berlin/2026/): the conference, linked from the video description. - [W3C WAI, Introduction to Web Accessibility](https://www.w3.org/WAI/fundamentals/accessibility-intro/): the definition Klepper anchors on at 00:06:31 and 00:28:29. - [WebAIM Million](https://webaim.org/projects/million/): the annual top-1-million homepage evaluation behind the 95.9% figure at 00:08:00. Confirm which year's report she used. - [WHO, Disability fact sheet](https://www.who.int/news-room/fact-sheets/detail/disability-and-health): the likely source for "16% of the world has a disability," which the transcript doesn't name. - [Anthropic, On the Biology of a Large Language Model](https://transformer-circuits.pub/2025/attribution-graphs/biology.html): the paper criticized at 00:21:58; it studies the Claude 3.5 Haiku model. - [European Accessibility Act](https://commission.europa.eu/strategy-and-policy/policies/justice-and-fundamental-rights/disability/union-equality-strategy-rights-persons-disabilities-2021-2030/european-accessibility-act_en): the law the 37% compliance figure at 00:09:31 refers to. - Modern Web Guidance (Chrome): the project discussed at 00:27:44. No URL is given in the transcript or description; look under the GoogleChrome GitHub organization before citing. - George Orwell, *Nineteen Eighty-Four*: the Newspeak frame at 00:01:26. - René Magritte, *The Treachery of Images*: the "this is not a pipe" image at 00:29:58; the artist isn't named in the transcript. - Klepper's earlier Write the Docs Portland talk on accessible documentation, referenced at 00:04:22. Search the Write the Docs YouTube archive; no link is given. - The survey of 1,600 professionals (00:09:31), the World Economic Forum $11 trillion estimate, and the $4.75 trillion German GDP figure (00:10:14 to 00:10:57) have no named publisher in the transcript. Find the primary reports before repeating any of the three. ## 🗣️ Paraphrase Klepper's case starts from a definition and refuses to let go of it. The W3C says accessibility is the tools and technologies that let people with disabilities use the web. That sentence has a subject, and the subject isn't software. So when a PM says "agents need accessibility" or a vendor says "AI will solve accessibility," both are borrowing a word that was built to protect a specific group and pointing it at something else. She grants the observation behind the trend: an agent that reads the accessibility tree does better on a page with real headings, labels, and landmarks. Her point is that this is a side effect, and if you let the side effect become the reason, the work follows the machines and the people it was for get left out. Her numbers for the stakes: 16% of the world, 95.9% of homepages failing. The second half turns from accessibility to anthropomorphism, and here she's practical. Nobody is going to rename neurons. But you can retitle a paper from "biology" to "construction," cut "AI voice" from a tutorial heading, and rewrite your agent-skills docs to say they offer guidance and not a guarantee, because a probabilistic system can't promise accessible output on every run. She does this last one to her own team's project in public. Her closing rule for docs writers is short: optimize for the humans who use the agents, not for the agents, and semantic HTML will serve both. ## 📰 Newsletter **Who is accessibility for? Same answer as before.** Alexandra Klepper, senior technical writer on Google Chrome, spent 35 minutes at Write the Docs Berlin 2026 on a question the WordPress project is about to hit hard: now that coding agents and browser agents read the accessibility tree, does accessibility become a machine feature? Her answer is no, and the reasoning is worth having on hand for the next Trac ticket that justifies an ARIA fix "so agents can parse it." The W3C defines accessibility as tools and technologies so people with disabilities can use the web. Agents benefit. Agents don't need. Keep the justification pointed at people and let the agents ride along. Two numbers she uses: 16% of the world lives with a disability, and 95.9% of homepages fail an automated check, per WebAIM. She argues vibe coding and framework bloat push that number the wrong way, which is a direct argument for the Accessibility team's review of AI-generated block markup and for the Documentation team's style guide to state, in so many words, that AI-assisted code needs the same accessibility review as any other patch. The second half is a copy-editing clinic. Anthropic's "On the Biology of a Large Language Model" becomes "On the Construction of..." with one word. Google Cloud's "AI voice generator" loses "AI voice." Chrome's own Modern Web Guidance docs get rewritten to say they provide guidance, not a guarantee, because a probabilistic model can't promise accessible output every time. If you write about the Abilities API, AI Building Blocks, or any WordPress AI feature, that last edit is the one to steal. Links: [the talk](https://www.youtube.com/watch?v=wH3H8ko43gM), [W3C accessibility intro](https://www.w3.org/WAI/fundamentals/accessibility-intro/), [WebAIM Million](https://webaim.org/projects/million/). ## ✅ Action items #### DevRel - Developer advocacy: build a 10-minute talk segment that walks one real docs page through Klepper's three edits (biology to construction, drop "AI voice," guidance not guarantee) and shows the diff. - Education: add a "benefits from vs needs" exercise to any AI workshop: give attendees five sentences from real AI docs and have them mark which ones attribute needs or feelings to software. - Developer experience: audit quickstarts and READMEs for AI features for anthropomorphic verbs ("the model understands," "the agent decides," "the AI thinks") and rewrite to mechanism ("returns," "ranks," "selects"). - Developer marketing: never write "AI-ready accessibility" or "accessibility for agents" in campaign copy. If agent compatibility is a selling point, say "semantic HTML that assistive technology and agents both parse." - Community management: add a one-line channel norm that accessibility asks get justified by user impact, not by agent compatibility. - Internal/external/cross-project: send Klepper's people-first proposal, when she publishes it, to the WordPress Accessibility and Documentation teams for comment; she asks for feedback at 00:05:47. #### OSPO - Policy and governance: write the rule that AI-generated contributions pass the same accessibility review as human-written ones, and that the PR template says so. - License compliance: no direct action from this source; nearest relevance is that agent-skill files like Modern Web Guidance carry licenses and provenance that should be checked before adoption. - Security management: none from this source. - Software inventory: if agent skill packs enter the toolchain, inventory them like any dependency, with the "guidance, not guarantee" caveat recorded alongside. - Developer support: publish a short internal glossary (accessibility, agent, model, user) with the W3C definition first, mirroring Klepper's web AI glossary at 00:05:05. - Community engagement: none from this source. - Strategic alignment: if leadership reports include "AI will improve accessibility," require the metric to be a human-outcome metric (task completion by assistive-tech users), not a tree-parsing metric. #### Community - Maintainers: add an "impact on people with disabilities" field to accessibility issue templates so the justification stays on people even when an agent surfaced the bug. - Contributors: pair new contributors on one accessibility fix and have them write the commit message with the affected user, not the tool, as the subject. - Sponsors: if a sponsor funds AI-assisted remediation, require outcome reporting against WCAG success criteria and a manual assistive-tech test, not only an automated scan delta. - Ecosystem: run one cross-project review of how AI features are described in docs, scoring for anthropomorphism; publish the rubric. #### WordPress - Core: no direct action from this source; nearest relevance is that the Accessibility Coding Standards already point at people, and any AI-generated markup entering core inherits them. - Core Performance: none. Klepper's claim that AI-assisted code inflates page complexity (00:08:45) is a hypothesis worth measuring but not a perf action. - Core AI: write the responsible-AI guardrail language now, in these words: outputs are probabilistic; generated markup is guidance and not a guarantee of WCAG conformance; every generated block or pattern still requires accessibility review. Build an evaluation set of generated block markup scored against WCAG 2.2 AA (headings, labels, alt text, contrast tokens, focus order) and publish pass rates per release. Describe the Abilities API and AI Building Blocks in terms of what they do for site owners and end users, never in terms of what "the AI needs." - **Core Program Team:** put "AI-generated changes get the same a11y review" in the cross-team release checklist and confirm it in the retrospective. - Mobile: none. - Playground: build a blueprint that loads a vibe-coded page next to a hand-written one and runs an automated a11y check on both, as a teaching demo for the 95.9% claim. - Community: offer Klepper's talk as a WordCamp speaker-mentorship reference for anyone proposing an "AI and accessibility" session, with the benefits-vs-needs distinction as the required framing. - Training: draft one Learn WordPress lesson, "Writing about AI features without anthropomorphism," using the biology, AI voice, and guidance examples. - Marketing: strike "AI-powered accessibility" from release messaging unless a human-outcome result backs it. - Support: add a HelpHub note that AI-generated content and code in a site may fail accessibility checks and that the site owner remains responsible. - Design: annotate design-system components with which semantics serve assistive technology, so agent compatibility is never the stated reason a component has a label. - Documentation: three concrete edits. (1) Add glossary entries for accessibility (W3C definition verbatim), agent (model plus rules plus memory plus tools, per 00:15:21), and model. (2) Add a style-guide rule against anthropomorphic verbs and against "AI voice," "the AI thinks," "the model understands," with mechanism-based replacements. (3) Add a standard disclaimer block for any AI-feature doc: "This feature produces probabilistic output. It provides guidance and does not guarantee results, including accessibility conformance." - Photos: none from this source. - TV: caption quality is Klepper's own exposure here (see Accessibility assessment); apply human-corrected captions to WordPress.tv AI and accessibility talks first. - Polyglots: when translating AI docs, flag anthropomorphic source strings for rewrite rather than translating them faithfully; add "guidance, not guarantee" to the glossary. - Accessibility: (1) Add a handbook statement that agent compatibility is a byproduct of accessible markup and not a justification; link the W3C definition. (2) Review open Trac tickets tagged accessibility for "for AI" or "for agents" rationale and reword to the affected users. (3) Publish a short position on AI-generated markup: automated checks cover deterministic criteria (contrast, missing alt, missing labels); manual assistive-tech testing stays mandatory. (4) Invite Klepper to an Accessibility team meeting to review the people-first proposal when it's public. - Meta: none from this source. - CLI: none. - Hosting: none. - Tide: none from this source; nearest relevance is that automated scoring covers only the deterministic subset Klepper describes at 00:19:05. - Openverse: none. - Themes: add "AI-generated theme code passes the same accessibility-ready review" to the review handbook. - Plugins: same as Themes, for plugin review guidelines and the readme's accessibility claims. #### The WPOCC - Programs: if The WP Open Community Collective funds accessibility fellowships, write the call in people-first terms and exclude "agent readiness" as a stated goal. - Projects: a documentation sprint on AI-feature docs, using Klepper's three edits as the rubric, is a bounded, fundable project. - Funding experiments: none from this source. ### FAIR - Technical independence: none from this source. - Policies and covenants: if FAIR publishes AI-related guidance, adopt the "guidance, not guarantee" language and the W3C accessibility definition in the covenant text. - Productized governance artifacts: add an "anthropomorphism check" row to any docs scorecard covering AI features. ## 🧪 Hypotheses 1. In WordPress Trac and GitHub accessibility issues opened since 2025, the share whose rationale cites agents, LLMs, or "AI-readiness" is rising year over year, and those issues get triaged faster than issues whose rationale cites screen-reader users only. 2. Adding a "guidance, not a guarantee" disclaimer to an AI coding-tool doc page lowers the share of users who believe generated code is WCAG conformant by at least 15 percentage points, measured by a post-read comprehension survey. 3. Rewriting AI-feature docs to remove anthropomorphic verbs lowers users' overtrust, measured as the gap between stated confidence in the tool's output and its measured accuracy. 4. Semantic HTML fixes that raise screen-reader task completion also raise browser-agent task completion on the same pages, with a correlation above 0.6, supporting Klepper's "center humans, agents follow" claim at 00:29:14. 5. Pages generated with AI coding assistants in 2026 fail WebAIM-style automated checks at a rate at or above the 95.9% baseline Klepper cites, and carry more DOM nodes per page than hand-written comparison pages. 6. Among disabled readers of technical documentation, preference splits between people-first ("people with disabilities") and identity-first ("disabled people") phrasing, so a style guide that mandates one form alone will be rejected by a meaningful minority. ## 📊 Methodology - **H1.** Pull all issues with an accessibility label from WordPress Trac and the Gutenberg repo, 2023 to present. Code the rationale text for agent/AI justification with two annotators and report inter-rater agreement. Compare time-to-first-triage and time-to-close across groups with survival analysis. - **H2.** Two-arm randomized test on one real doc page (a WordPress AI feature or the Abilities API). Arm A carries the disclaimer block; arm B doesn't. After reading, ask three comprehension questions including "Does code from this tool meet WCAG 2.2 AA?" Pre-register the 15-point threshold. Recruit 200 developers. - **H3.** Give participants two versions of the same docs page (anthropomorphic vs mechanism verbs), then have them predict the tool's accuracy on 10 tasks and run the tasks. Overtrust is predicted minus observed accuracy. Compare arms with a t-test; report effect size. - **H4.** Take 30 pages before and after semantic fixes. Measure task completion with three screen-reader users per page and with two browser agents on scripted tasks. Compute the correlation between the two deltas. - **H5.** Collect 200 public pages self-described as built with AI assistants and 200 matched hand-coded pages. Run axe-core and count DOM nodes. Report failure rate and node count with confidence intervals. - **H6.** Survey disabled readers recruited through WordPress Accessibility team channels and at least two disability organizations. Ask phrasing preference and acceptability of each form in docs. Report by disability type and region; a minority above 20% counts as meaningful. - **Controls throughout.** Blind annotators to hypothesis. Fix page content across arms except for the manipulated language. Publish instruments before collecting data. ## ❓ Questions - Where is the line between a definitional argument and a semantic dodge? If an executive says "AI will remediate WCAG failures" instead of "solve accessibility," does Klepper's objection still apply? - Who publishes the "1,600 professionals" survey cited at 00:09:31, and does its 86% figure ask about solving accessibility or about assisting with it? - Which year's WebAIM Million report yields 95.9%? WebAIM publishes annually and the figure moves. - Klepper concedes neurons and neural networks are locked in. What's the rule for deciding which biological metaphors are worth fighting? Age of the term? Whether it appears in a title? Whether a plain mechanism word exists? - The disability community includes many who prefer identity-first language. Does repurposing "people-first" for AI docs risk importing a contested term into a new domain without that debate? - If "user" implies human (00:27:02), what word should docs use for the agent that calls an API on a human's behalf, and does the answer differ for scrapers? - How should docs disclose AI use to audiences who "don't know, don't care" (00:32:51) without either hiding it or burying the page in caveats? - Emily's question stands: what do human support agents call themselves now? Klepper offered "customer advocate." Does that survive contact with job titles and org charts? ## 🔮 Future research - **A measured anthropomorphism index for docs.** Nobody has a validated lexicon and scoring rubric for anthropomorphic language in technical documentation. Building one would make the H3 overtrust study replicable and let style guides set a threshold. - **Agent compatibility as a11y proxy.** If H4 holds, browser-agent task completion becomes a cheap, continuous proxy for some assistive-tech outcomes. If it fails, that's a warning against exactly the conflation Klepper attacks. Either result is useful. - **Vibe-coding and page complexity.** Klepper asserts at 00:08:45 that AI-assisted coding moves accessibility the wrong way. A longitudinal sample of AI-generated sites against the WebAIM Million baseline would test it. - **People-first vs identity-first in tech writing.** The disability-language debate is well documented outside tech; how it plays in developer docs, and whether a mixed style is acceptable, isn't. - **The "guidance, not guarantee" pattern.** Track whether AI-tool vendors adopt explicit non-guarantee language in docs after 2026 and whether it correlates with fewer conformance complaints. - **Attribution of the closing quote.** The judges/correlates passage at 00:29:58 is read as a quote without a source. Find the origin before reusing it. ## 💡 Implications - **Practice.** Three edits you can make to any AI doc today: replace biological metaphors in titles where a construction word fits, cut "AI voice" and similar attributions, and add a guidance-not-guarantee statement to any probabilistic feature. Justify every accessibility fix by the people it serves, and mention agent compatibility, if at all, as a side effect. - **Policy.** Style guides need a rule, not a preference. The WordPress Documentation style guide's inclusive-language section can carry the W3C definition and an anthropomorphism rule in under 200 words. Governance for AI-generated contributions needs the same accessibility review gate as human contributions, written down. - **Security.** Indirect. Docs that overstate what a probabilistic tool guarantees create liability under the European Accessibility Act for site owners who trusted them. "Guidance, not guarantee" is a disclosure control as much as an editorial choice. - **Education.** The determinism/probabilism split at 00:13:52 and the four-part agent definition at 00:15:21 are ready-made lesson material. The benefits-vs-needs distinction at 00:16:06 is the single idea most worth teaching to PMs. - **Funding models.** Accessibility funding pitched as "AI readiness" will be redirected the moment agents can parse messy DOMs. Funding pitched on the 16% and the $11 trillion (once sourced) survives that shift. Where the money is justified determines where it goes when priorities move. ## 🧠 Learning styles assessment - **Reading level estimate.** Flesch-Kincaid Grade Level on the transcript body: about grade 8. Flesch Reading Ease near 68. Short sentences and spoken rhythm keep it low; terms like probabilism, anthropomorphism, and accessibility tree are each defined on first use, which is the talk's own method. #### 👀 Visual learners - Present: slides with statistics, paper titles, a tutorial heading, and marked-up corrections; the Mona Lisa and Magritte's pipe as images; a QR code for slides. - Improve: the corrections at 00:28:29 are shown, not read, so the transcript records "I've made some corrections" without the text. A before/after table in the description or a slide link would fix it. No diagram of the four-part agent model. #### 🔊 Aural learners - Present: full narration with repetition for emphasis ("AI is not human. AI cannot have human needs," 00:19:48), audience call-and-response on the statistics, and a read-aloud Q&A. - Improve: the QR code, paper titles, and tutorial heading are read once, fast. A 3-minute audio summary with the three edits would serve podcast listeners. #### 🗣️ Verbal learners - Present: definitions for accessibility, AI, determinism, probabilism, agent, and anthropomorphism, each stated in one or two sentences before use. Two analogies (algebra vs calculus, Mona Lisa replica). - Improve: the closing quote is unattributed. "Modern Web Guidance" and "MWG" appear without a one-line description of what a skill file is. #### 🧩 Kinesthetic learners - Present: three worked edits on real text (biology to construction, AI voice removal, guidance not guarantee); a "read your language and ask if it sounds written for humans" self-test at 00:29:14. - Improve: no checklist to take away, no downloadable proposal. A five-item audit card would convert the talk into a task. #### 🧠 Logical learners - Present: a definitional argument with premises stated in order (W3C definition, AI has no disabilities, therefore AI needs no accessibility); the deterministic/probabilistic partition of testable criteria at 00:19:05. - Improve: statistics lack sources and years; the "better data doesn't make a better model" claim at 00:23:25 needs qualification. Show the WebAIM year and the survey publisher. #### 🧑‍🤝‍🧑 Social learners - Present: a live audience, shout-out polls at 00:07:17 and 00:08:00, three audience questions, an explicit request for feedback and a call to form a working group at 00:05:47. - Improve: no channel named for the feedback. A Write the Docs Slack channel or a form link in the description would close the loop. #### 🧘‍♀️ Solitary learners - Present: the self-test question at 00:29:14 ("Does it sound like this is written for humans?"); a self-contained argument that rewards a second viewing. - Improve: a worksheet with five sentences to rewrite would give solo learners a finish line. #### 🌿 Naturalistic learners - Present: biological metaphor is the subject of the middle third, with feeding, neurons, and life discussed at 00:22:40 to 00:23:25. - Improve: the talk critiques biological framing without offering an ecological lens on AI's real footprint. One line on the environmental cost of the systems being anthropomorphized would land with this group. ## 📊 Accessibility assessment #### Screen reader semantics - Present: YouTube supplies player landmarks; the description is plain text with one link; the auto-transcript is machine-readable. - Improve: slide content (statistics, the two titles, the doc corrections) is visual only at several points, notably 00:28:29. WCAG 2.2 AA 1.2.5 (Audio Description, Prerecorded) isn't met for those slides. Speaking the corrected text aloud would have fixed it at no cost. #### Keyboard navigation - Present: the YouTube player is keyboard operable with visible focus; no in-video interactions exist. - Improve: nothing in the content adds barriers. 2.1.1 (Keyboard) and 2.4.7 (Focus Visible) pass via the platform. #### Contrast and readability - Present: unverifiable from a transcript. Conference slide decks on a projector are the likely presentation. - Improve: the QR code at 00:30:42 is the only path to slides and can't be checked for contrast or size from the transcript. Put the slide URL in the description as text; 1.4.3 (Contrast Minimum) then stops mattering for that element. #### Text resizing and reflow - Present: captions resize independently in the player. - Improve: on-screen slide text is burned into the video and doesn't reflow, so 1.4.4 (Resize Text) and 1.4.10 (Reflow) fail for slide content. A linked deck resolves it. #### Media and captions - Present: auto-generated captions exist, which is how this analysis was produced. The Q&A host reads written questions aloud, which is good practice. - Improve: the captions garble the speaker's own name ("Alexander Clipper"), Jeffrey Yasskin ("Yaskin"), WebAIM ("Web AIM"), the Louvre ("Louv"), and "text-to-speech." On a 35-minute accessibility talk, uncorrected captions are a 1.2.2 (Captions, Prerecorded) failure in substance. No chapters exist, so 2.4.10-style navigation by section isn't available. Human-corrected captions and a chapter list are the two highest-value fixes. #### Forms and interactions - Present: none in the video. - Improve: the QR code is a visual-only interaction. 1.1.1 (Non-text Content) requires a text alternative; the description should carry the URL. #### Standards reference - Primary gaps against WCAG 2.2 AA: **1.2.2** (captions inaccurate on proper nouns), **1.2.5** (slide content not described in audio at 00:28:29 and elsewhere), **1.1.1** (QR code without a text alternative), **1.4.4 / 1.4.10** (burned-in slide text). The publisher, not the speaker, controls captions and chapters, and Write the Docs can fix both after the fact. The irony of an accessibility talk shipped with auto-captions is the most useful thing to say about it, and Klepper's own rule at 00:29:14 (optimize for the humans) applies to the recording. ## 🌍 DEIB assessment #### Cultural breadth - Present: a European audience and European law (the EAA), a German GDP comparison, an American speaker, the Louvre and the Mona Lisa as shared references. - Improve: every example is Western. The W3C definition is global, but accessibility law, assistive-tech availability, and language-model coverage vary by region, and none of that appears. One non-Western example of AI language use would widen the case. #### Inclusive language - Present: consistent "people with disabilities"; the talk is itself an argument for careful language; "user" is flagged as a term worth rethinking (00:27:02); she notes "AI accessibility bros" as her own frustration, not as analysis. - Improve: the talk adopts people-first language without acknowledging that many disabled people prefer identity-first phrasing, and it repurposes the term for a new use without engaging that debate. "Able-bodied people" (00:27:02) is a term many in the community replace with "non-disabled." "You guys" appears at 00:05:05. "AI accessibility bros" (00:05:05) is gendered labeling of opponents. Each is a one-word swap. #### Neurodiversity support - Present: the talk is chunked by definition, with each concept defined before use; repetition marks the key claims; Klepper discloses burnout leave (00:03:38) and asks to read a question herself because she processes visually (00:33:34), modeling an accommodation request in public. Executive-functioning support is named as a legitimate AI use at 00:19:05. - Improve: no chapters on the recording; the pace accelerates in the last five minutes. A chapter list and a linked deck would help viewers who need to pause and re-read. #### Representation - Present: a woman presenting technical content at a docs conference; credits to Darcy Lima and Justine Owens; questions from Emily, Deepak, and Elizabeth; a disability-community concept (people-first language) treated as a source of expertise rather than a footnote. - Improve: no disabled person's voice appears directly; the talk speaks about people with disabilities as the group accessibility serves, with statistics rather than testimony. One quoted assistive-tech user would change the register. #### Psychological safety - Present: the speaker calls out her own employer (Google Cloud, 00:24:51) and her own team (Chrome MWG, 00:27:02) before anyone else's, apologizes to Anthropic in the room, invites correction in Q&A ("I can handle corrections," 00:13:52), and admits when she has no answer (00:32:09). The host thanks a questioner for a hard question. - Improve: "Ew" at the biology title (00:21:58) and "bros" are mockery, mild but present. They lower the bar for anyone in the audience who works on those papers to raise a counterpoint. ## ✅ Logical validity #### Strong reasoning - **Deductive, 00:06:31 to 00:07:17 and 00:19:48 to 00:20:33.** Premise: accessibility is defined by the W3C as tools and technologies so that people with disabilities can use the web. Premise: AI systems are not people and cannot have disabilities. Conclusion: AI does not need accessibility. Valid given the premises. To clarify further: state that the conclusion is about need and not about benefit, which she does at 00:16:06. - **Distinction between benefit and need, 00:16:06.** She grants the strongest fact for the other side (agents understand accessible pages better) and refuses only the inference from "benefits from" to "needs." This is the load-bearing move in the talk and it holds. - **Causal chain, 00:13:52 to 00:28:29.** LLMs are probabilistic; therefore coding agents can't produce accessible output every time; therefore docs must say guidance rather than guarantee. Each step follows. Strengthen it by citing a measured failure rate for generated markup. - **Partition of testable criteria, 00:19:05.** Color contrast is deterministic and automatable; complex language is a judgment and suits probabilistic tools. A correct and useful partition that also explains why automated scans catch only part of the 95.9%. - **Concessions against interest, 00:18:20 to 00:19:48 and 00:24:08.** Naming real uses of AI for accessibility, and admitting neurons and neural networks won't be renamed, both narrow her claim to what she can defend. - **Self-application, 00:27:02 to 00:28:29.** Applying the critique to her own team's docs before anyone else's is both rhetorically and logically sound; it shows the rule has no exemption for the speaker. #### Fallacies - **Definitional stipulation, 00:06:31 and 00:20:33.** "AI cannot solve accessibility because accessibility is not a problem to solve" wins by definition. Someone who says "AI will fix the WCAG failures on this site" is making a claim the definition doesn't touch. *Fix:* argue the empirical point directly: automated remediation covers the deterministic subset and misses the rest, so "solve" overstates what any tool can do. - **Bare assertion, 00:09:31.** "Empathy doesn't work" and "I'm unconvinced the business argument works" arrive with no evidence, and she then presents business numbers anyway. *Fix:* cite a study on which arguments move accessibility budgets, or drop the aside. - **Unsourced statistics, 00:08:00 to 00:10:57.** 16%, 95.9%, the 1,600-person survey, $11 trillion, $4.75 trillion. None has a named publisher or year in the transcript. *Fix:* name WHO, the WebAIM Million year, the survey publisher, and the WEF report on the slide and aloud. - **Weak analogy, 00:16:51 to 00:18:20.** The Mona Lisa's value comes from scarcity and provenance; a tool's value comes from function. A perfect replica of a hammer is a hammer. The analogy supports "a replica isn't the original" but not "a replica can't do what the original does." *Fix:* keep the analogy for identity and use the probabilism argument for capability. - **Overstatement, 00:23:25.** "Better data... does not make a better model." Data quality is among the largest drivers of model quality. *Fix:* "better data is necessary and not sufficient." - **Unsupported causal claim, 00:18:20.** "People have started to care less about other people in service of this greater version of us, AI." No evidence is offered for the trend or its cause. *Fix:* cite a measured shift in accessibility investment or user research, or frame it as a concern rather than a finding. - **Labeling and appeal to ridicule, 00:05:05 and 00:21:58.** "AI accessibility bros" and "Ew" characterize opponents and a paper title rather than their arguments. *Fix:* quote the position being opposed and answer it. - **Unattributed quotation, 00:29:58.** "One last quote" with no author leaves the audience unable to check it. *Fix:* name the source on the slide and aloud. #### Tone of voice - Warm, self-deprecating, and direct, in conference-talk register. Contractions and second person throughout ("You all can do better," 00:26:19). Deliberate repetition as emphasis ("AI is not human. AI cannot have human needs," 00:19:48). Classroom energy: shout-out polls, "Great job, everyone" after landing a point (00:19:48). The tone hardens into imperatives in the recommendations ("Don't optimize your docs for agents," 00:29:14) and softens into candor in Q&A ("let me get back to you," 00:32:09). The mockery is light and aimed at institutions (Anthropic, Google Cloud, her own team) more than people, and she apologizes in the same breath both times. ## 🛡️ DARVO analysis Not applicable to this source. DARVO (Deny, Attack, Reverse Victim and Offender) describes a defensive response to accusation. Klepper is the critic here, not the accused; nobody in the transcript charges her with anything, and the Q&A questions are requests for guidance. For completeness across the three prongs: **Deny**: no denial or minimization appears; she concedes limits of her own position at 00:12:26 (may alienate PMs), 00:18:20 (AI does help), and 00:24:08 (won't rename neurons). **Attack**: she criticizes word choices by Anthropic, Google Cloud, and her own team, with proposed fixes attached; the closest thing to an attack on persons is "AI accessibility bros" at 00:05:05, which labels a group rather than answering it (see Fallacies). **Reverse victim/offender**: no such move; she never casts herself as wronged, and she positions people with disabilities, not herself or writers, as the party at risk. ## 🎯 Conclusion The durable idea is a one-line test for any sentence about AI and accessibility: does this justify the work by the people it serves, or by the machine that reads the output? Klepper's argument at 00:16:06 grants that agents parse accessible pages better and refuses only the slide from "benefits from" to "needs." That distinction holds on the W3C's own definition and it's the sentence to keep. The second idea is a copy-editing method, demonstrated three times on real text: swap a biological metaphor for a construction one where the field hasn't locked the term in (00:24:08), cut attributions of human capacities like "voice" (00:24:51 to 00:25:35), and state that a probabilistic tool provides guidance and not a guarantee (00:28:29). She applies the last one to her own team's docs in public, which is the strongest evidence in the talk that the method is practical. What doesn't survive: the statistics are unsourced in the transcript, the "can't solve accessibility because it isn't a problem" line is a definitional move an executive can route around, and the Mona Lisa analogy supports identity but not capability. The recording itself ships with auto-captions that misspell the speaker's name and no chapters, which is the fastest fix Write the Docs could make. **Next steps for readers:** pull one AI-feature doc you own and apply the three edits. Then check your accessibility backlog for any ticket justified by agent compatibility and rewrite the rationale to name the affected users. Source the 16%, 95.9%, and $11 trillion figures before repeating them. And watch for Klepper's people-first proposal; she asked for feedback at 00:05:47, and the WordPress Accessibility and Documentation teams are the right reviewers. ## 🔄 Refactor Appendix — Write inclusive, people-first language about AI - Alexandra Klepper > Density: 2–5 Literature atoms (H3 pad), 0–3 Permanent candidates (H4 pad). > Emitted: 5 Literature, 2 Permanent. Resonance filter applied — the unsourced > statistics and the Mona Lisa analogy were dropped as not load-bearing. > > Konik claim-header rules: present tense, declarative, ≤10 words, searchable, unique. > > Atomicity: one declarative sentence per atom. "and"/"also" → split. > > Tag pool: #bmf · #addendum · #xref · #articleseed · #storystem · #share · #aphorisms ### 📚 Literature Split Pad #### [H1] Accessibility exists for humans, so AI needs none > [!quote] [00:19:48] > AI does not need accessibility because accessibility exists for human beings. **Context & use**: - 00:19:48, the conclusion of the definitional argument that starts with the W3C definition at 00:06:31 and passes through the benefit-vs-need distinction at 00:16:06. - The reusable part is the test, not the slogan: justify accessibility work by the people it serves; agent compatibility is a byproduct you may mention, never the reason. **Tags**: #xref #share #addendum <!-- atomicity: 1 idea ✓ --> #### [H2] Accessibility is a practice, not a problem to solve > [!quote] [00:20:33] > AI cannot solve accessibility because accessibility is not a problem to solve and neither are disabilities. **Context & use**: - 00:20:33, restating the "not a solution, not a checklist" point from 00:06:31 against the "86% think AI will solve it" survey figure at 00:10:14. - Record the weakness alongside: this wins by definition. Pair it with the empirical version (automated tools cover the deterministic subset, 00:19:05) when using it against "AI will fix our WCAG failures." **Tags**: #xref #aphorisms <!-- atomicity: 1 idea ✓ --> #### [H3] Optimize docs for the agent's human users > [!quote] [00:29:14] > Don't optimize your docs for agents. Optimize the docs for the human users of the agents. **Context & use**: - 00:29:14, the operating rule in the recommendations, followed by "write semantic HTML" and "when you center humans, agent can better support them too." - Directly applicable to llms.txt debates, agent-skill files, and any "AI-readable docs" initiative: the human on the other end of the agent is the audience. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> #### [H4] Probabilistic tools warrant guidance, never guarantees > [!quote] [00:28:29] > We should s say that we are providing the guidance but that we are not guaranteeing it. We must choose our words carefully. **Context & use**: - 00:28:29, showing her corrections to Chrome's Modern Web Guidance docs after establishing at 00:27:44 that LLM-based coding agents can't produce accessible output 100% of the time. - The transcript stutter ("s say") is preserved. The pattern is a disclosure control as much as a style choice; it belongs in any WordPress AI-feature doc. **Tags**: #xref #share <!-- atomicity: 1 idea ✓ --> #### [H5] Retitle "biology" as "construction" for model papers > [!quote] [00:24:08] > What if instead of biology the title was on the construction of a large language model like a little better maybe? **Context & use**: - 00:24:08, the proposed fix for Anthropic's "On the Biology of a Large Language Model" (00:21:58), offered right after conceding that neurons and neural networks won't be renamed. - The generalizable rule: fight the metaphor where a plain mechanism word exists and the term isn't yet locked in; titles and headings are the highest-value targets. **Tags**: #xref #articleseed <!-- atomicity: 1 idea ✓ --> ### 🧠 Atomic Split Pad #### [P1] Agent compatibility is a byproduct of accessibility, never its justification **Claim** (one sentence, Courtney's voice): I justify accessibility work by the people with disabilities it serves, and I treat agent compatibility as a side effect I can mention but never as the reason. **Evidence pointer**: [[#H1]] · [[#H3]] **Scope & conditions**: Holds for any accessibility fix, ticket rationale, handbook page, or funding pitch. The risk it guards against is budget and attention following the machines once agents can parse messy DOMs. It does not require hiding the agent benefit; it requires ordering the reasons. Counterexample: a pure machine-readability feature (a sitemap, an API schema) isn't accessibility work and shouldn't be labeled as such either. **Links**: - [[WordPress accessibility advocacy]] — because applies-in ticket rationale and handbook language - [[Semantic HTML serves assistive technology]] — because extends - [[llms.txt and AI-readable documentation]] — because contrasts on who the audience is <!-- atomicity: 1 idea ✓ --> #### [P2] Docs for probabilistic tools state guidance, not guarantees **Claim** (one sentence, Courtney's voice): When a feature's output is probabilistic, the documentation says what it helps with and refuses to promise the outcome, including accessibility conformance. **Evidence pointer**: [[#H4]] **Scope & conditions**: Applies to AI coding assistants, generated block markup, agent skill packs, and any WordPress AI feature. Does not apply to deterministic checks (color contrast, missing alt text), which can state pass/fail outright. The tell that a doc violates it is any form of "ensures," "guarantees," or "makes your site accessible" attached to generated output. **Links**: - [[Every claim carries a fact]] — because applies-in writing standards - [[WordPress Core AI guardrails]] — because applies-in guardrail language - [[Determinism versus probabilism in AI tooling]] — because evidence-for the boundary between what can be promised and what can't <!-- atomicity: 1 idea ✓ --> ### ✏️ Fleeting Queue - [F1] Source the four statistics before reusing any: 16% (WHO?), 95.9% (which WebAIM Million year?), the 1,600-professional survey (publisher?), $11 trillion (WEF report?) — https://www.youtube.com/watch?v=wH3H8ko43gM at 00:08:00 to 00:10:57 #fleeting #inbox - [F2] Find the origin of the closing quote "Where a human judges, a model correlates..." read at 00:29:58 without attribution #fleeting #inbox - [F3] Open question: does repurposing "people-first language" for AI docs import the people-first vs identity-first debate without engaging it? Worth a paragraph if I write about this talk #fleeting #inbox - [F4] Article seed: "Benefits from is not needs" as the one-line test for AI-and-accessibility sentences, with the WordPress Trac rationale audit as the worked example #fleeting #inbox #articleseed - [F5] Watch for Klepper's "People-first accessibility language for AI and agents" proposal going public; she asked for feedback at 00:05:47 and named a working group as step three #fleeting #inbox - [F6] Check whether Chrome's Modern Web Guidance docs now carry the guidance-not-guarantee language she showed at 00:28:29; the repo is under the GoogleChrome GitHub org #fleeting #inbox ### → Things 3 (max 2 actions) - [ ] Audit open WordPress accessibility tickets and the Accessibility handbook for "for AI" or "for agents" rationale and reword to the affected users — WP a11y advocacy, 45 min #things - [ ] Add an anthropomorphism and guidance-not-guarantee check to my AI-article editing checklist (biology, neurons, feeding, voice, thinks, understands, ensures) — writing process, 15 min #things <!-- deep-read:sections end --> 📄 **Full text:** [[Full Document Contents/Articles/Write Inclusive, People-First Language About AI - Alexandra Klepper.md|Write Inclusive, People-First Language About AI - Alexandra Klepper by Write the Docs]] ## Highlights - [00:19:05] is deterministic. Color contrast has an exact math. You can test if yes, this is accessible or no, it's not. But then also there are things that you can evaluate with the uh complexity of probabilism like complex language. Generative AI can also be great at supporting executive functioning skills such as task prioritization, clearer information architecture, and distraction reduction. All things that can be helpful in your docs process. If you built a task management app or wrote the docs for it, you could use AI to prioritize your work based on additional context such as calendar events and [00:19:48] emails. So yes, AI can be a tool for accessibility. Meanwhile, there are dozens of articles and research written by folks that say bots are people, too. Bots need accessibility. AI is increasingly integrated into our lives. It's a reality we need to face. AI is not human. AI cannot have human needs. I'm going to keep saying that because I really want to make sure it is heard and internalized. Thus, AI does not need accessibility because accessibility exists for human beings. Great job, everyone. Great job. You heard. [laughter] AI cannot have disabilities. A model or tool may be able to imitate certain ([View Highlight](https://read.readwise.io/read/01m37b8g06vg0zdt4d0zgdt13y)) - Note: [H1] Accessibility exists for humans, so AI needs none - Tags: atom/literature/h1 - [00:20:33] types of disabilities, but that's all it ever is, an imitation. AI cannot solve accessibility because accessibility is not a problem to solve and neither are disabilities. Instead, accessibility is how we support and uplift people with disabilities to improve their experience and access. Even though generative AI can improve some people's capabilities to create work, it's critical to remember that people build this technology. If people don't understand what accessibility is, that it's a framework, or worse, if they don't respect people with disabilities, they can't create accessible [00:21:15] experiences. Even though most researchers and engineers have yet to assert that a machine is like truly equal to humans, there is a push to say, "We're almost there, though. Get on board or get left behind." When writers say that a machine can think, reason, and discuss with you, they want you to believe it's just like you're talking with another human. There are plenty of um AI therapy startups, for example. This is anthropomorphism. Uh I practiced saying that so many times and I was so nervous I'd get it wrong. Thank you. Thank you. The anthropomorphism is the attribution of human actions to ([View Highlight](https://read.readwise.io/read/01m37b8j968awx53065p14ms72)) - Note: [H2] Accessibility is a practice, not a problem to solve - Tags: atom/literature/h2 - [00:23:25] alive. It's also the word used to describe how a model is given data to process and improve its output. The better the b the data, the better the machine can predict outputs maybe. But just like humans, better food doesn't make you healthy. You know what is healthy anyway? It is just a tiny piece of the puzzle. Better data is just one tiny piece of how a model works. It does not make a better model. We are complex. We being humans and machines also are complex. How can I say anthropic is wrong to use the word biology when there are so many similarities. When the field has been [00:24:08] using language like this for decades, the chances of creating new words for neurons and neural networks, zero. Not going to happen. Not going to try. But just because you can't change some things doesn't mean you can't you just have to give up on this whole premise. One change makes a huge difference at least to me in this paper. What if instead of biology the title was on the construction of a large language model like a little better maybe? We can ask for these changes. We can try to make these changes that divorce the idea of machines from humanity. This sort of small technical language change is that separation between humans and ([View Highlight](https://read.readwise.io/read/01m37b8s063bpf348mzy95x9yd)) - Note: [H5] Retitle "biology" as "construction" for model papers - Tags: atom/literature/h3 - [00:27:44] called modern web guidance. And the premise is to offer a set of skills for coding agents so that they have the best practices modern browser information so developers can build better websites. Of course, this led to a lot of conversation about okay, what makes a website accessible? It's really hard. It's really hard. The answer is really complicated. Jeffrey Yaskin, who is a Chrome engineer, explained that MWG is harm reduction rather than solutionoriented. If people are going to use AI to generate websites, how can we help them do it better and help them make a more inclusive experience? As you know, LLMs are probabilistic. Therefore, AI coding agents with LLMs [00:28:29] cannot create accessible outputs 100% of the time. Honestly, I wish our docs were more explicit about this. So, here we go. I've made some corrections. Again, I'll be sharing this with you, Jeremy. Thank you, Jeremy. >> [laughter] >> We should s say that we are providing the guidance but that we are not guaranteeing it. We must choose our words carefully. When asked what is accessibility, think people first. Accessibility, let's use the W3C definition. Accessibility is a set of tools and technologies used to support people with disabilities. Do not equivocate humans, AI, and agents. It is not always a one:1 ratio. Like ([View Highlight](https://read.readwise.io/read/01m37b8pnp8jvwm23prpqrfdn4)) - Note: [H4] Probabilistic tools warrant guidance, never guarantees - Tags: atom/literature/h3 - [00:29:14] some things feel easy. The construction biology thing to me that was easy. Sometimes you have to really think about it. Come in and read your language. Does it sound like this is written for humans? Then you you might be on the right path. So here I suggest improving your documentation accessibility by understanding what your users need and support their assistive technology and maybe agents. Don't optimize your docs for agents. Optimize the docs for the human users of the agents. Write semantic HTML to support humans with assist of technology and agents to correctly navigate your website. When you center humans, agent can better support them too. [00:29:58] And remember to focus on the real thing. This is not a pipe. This is a painting of a pipe. An AI model is not a person. It is a replication of people based on the interpretation of the humans that built it and all of their flaws. I'll leave you with one last quote. Where a human judges, a model correlates. Where a human evaluates, a model predicts. Where a human engages with the world, a model engages with the distribution of words. Their architecture makes them extremely good at reproducing patterns to be found in text. It does not give them access to the world in which these words refer. And before I finish, special thanks to Darcy Lima, Justine Owens, and my ([View Highlight](https://read.readwise.io/read/01m37b8mekjnck0pqtfj98y17z)) - Note: [H3] Optimize docs for the agent's human users - Tags: atom/literature/h3