When Help Makes Things Up: AI Accessibility and the Reliability Gap

The instructions said “chew 4 tablets 2 times per day”. The product was a tube of topical medication.

That exchange sits in a 2026 research report from the American Foundation for the Blind (AFB), in the words of a survey participant who had pointed an AI visual description tool at a medication label and asked how to use it. “Luckily this was an obvious error,” the participant wrote, “but for now I will NOT be using AI for these tasks [and] will always confirm with a human.” Another participant in the same study used an AI tool to read tax forms that were not screen reader accessible, and reported that it “provided some very incorrect information that would have caused serious tax filing mistakes if I had not checked the information with a human.”

Neither story is a fringe horror. Phones now describe rooms in fluent paragraphs, apps read menus aloud and websites promise to fix themselves with a line of JavaScript. Underneath all that capability, a quieter question keeps surfacing from the people who depend on these tools: what happens when the thing that helps you is also the thing that makes things up?

The complaint has been circulating online for some time. One 2025 discussion on Reddit's r/ArtificialIntelligence forum is reported to have called the trend an “insidious form of ableism”, although that thread's exact wording could not be independently located for this article. The substance of the argument can, however, be tested against the record. As generative AI is positioned as a universal accessibility fix, the predictable tools that disabled and older people rely on risk being sidelined or replaced by systems that are more impressive and less trustworthy. If so, AI-first accessibility could widen the very gap it claims to close.

The evidence suggests that the worry is well founded, though not in the simple way a slogan implies. AI tools are delivering real independence to many blind and low vision people. The problem is not the tools. It is the temptation, among companies, website owners and even governments, to treat them as a substitute for the slow, unglamorous work of building things that are accessible in the first place.

Why a Mangled Word Beats a Fluent Fiction

A traditional screen reader is, at its core, a deterministic machine. It takes text that already exists in a document, an interface or a web page, and converts it into synthetic speech or braille. It can absolutely get things wrong. It may mispronounce a surname, stumble over an acronym or announce an unlabelled control as nothing more than “button”. But its errors have a particular shape. They are consistent: the same text produces the same mistake every time. And the underlying text is still there. A user who hears a garbled word can move through it character by character, check the spelling and work out what was meant. The error sits on top of a reliable source.

Generative AI does not work like that. A vision language model asked to describe a photograph or read a label is not converting existing text; it is producing a statistically plausible account of what it thinks it sees. When it fails, it does not usually fail with noise. It fails with fluency. A wrong instruction arrives in a grammatically perfect sentence, delivered with the same confidence as a right one. There is no underlying string to inspect, because the model's output is the only version of reality the user has been given.

Alex Katsarakes, a digital accessibility resident at AFB and a PhD student in human factors psychology, set this out in an October 2025 essay on alt text in the age of AI. Generative models, she wrote, “are prone to 'hallucination,' confidently asserting details that aren't actually present.” When a blind user cannot independently check an image, she argued, an inaccurate caption is not a cosmetic flaw. In her words, “it's a breakdown in trust.”

AFB's own survey data makes the point empirically. In its report The AI Quagmire, based on a survey of 1,735 US adults conducted between 1 July and 15 October 2025, 91 disabled users of AI visual description tools reported mistakes that harmed them or made their lives harder. When the researchers compared how disabled and nondisabled people rated the accuracy of AI output, they found a revealing asymmetry. Disabled users of visual description tools were more likely than nondisabled users to rate descriptions as “mostly accurate” (56.3 per cent against 46.4 per cent). With AI captions, the pattern flipped: disabled caption users, including deaf and hard of hearing people, were more sceptical than nondisabled users. AFB's interpretation is blunt: “This pattern suggests that disabled people are better able to detect captioning errors than visual description errors.”

Read that slowly. The people who most need image descriptions are, by definition, the people least able to verify them. A deaf user watching a speaker's lips and body language has some independent signal to check a caption against. A blind user listening to a description of a photograph often has none. Reliability is not one desirable feature among many in this context. It is the whole product.

What the Tools Genuinely Do Well

None of this means AI accessibility tools are a con. Pretending otherwise would be its own kind of disrespect to the many disabled people who use them every day and say they are better off.

Be My Eyes is the clearest example. The Danish company built its reputation on connecting blind and low vision people with sighted volunteers over live video. In 2023 it added Be My AI, an image description feature built on OpenAI's GPT-4 model. On 12 March 2026, the company announced that it had passed one million blind and low vision users and ten million volunteers across more than 150 countries. Its chief executive, Mike Buckley, called the milestone “a powerful reminder that accessibility is not a niche issue” and framed it as a global movement. Crucially, the human route has not been switched off: users can still tap a button and reach a sighted volunteer by live video.

Microsoft's Seeing AI, which began as a research project and launched on iPhone in 2017, made the jump to Android in December 2023. Saqib Shaikh, the founder and lead of the project, described it in Microsoft's announcement as “a free app that narrates the world for blind and low vision people”. The Android release brought generative features first tested on iOS, including richer scene descriptions and the ability to ask questions about a scanned document, such as the items on a menu or the total on a receipt.

Academic research confirms that the appeal is real. In a study presented at the ACM SIGACCESS ASSETS conference in 2024, Rudaiba Adnin and Maitraye Das of Northeastern University interviewed 19 blind people about their use of generative AI tools. The paper's title borrows one participant's description of the technology: “I look at it as the king of knowledge”. Participants used chatbots for writing emails and CVs, researching products and asking health questions, and they valued Be My AI's descriptions as more thorough than what sighted helpers usually offered.

AFB itself is not anti-AI. When it released its expert consensus report Empowering or Excluding in February 2025, Stephanie Enyart, the organisation's chief public policy and research officer, said the experts consulted had “shined a light on how AI can be a catalyst for access and inclusion.” Katsarakes, in the same essay that warns about hallucination, acknowledges that using tools such as Be My Eyes or Seeing AI “demonstrates clear potential for enhancing user autonomy.”

So the question is not whether these tools help. They do. The question is what happens when their help is treated as finished accessibility rather than as one layer within it.

The Verification Burden Falls on the User

What the research shows most consistently is that blind people who use AI description tools have become their own quality assurance department.

In a 2024 ASSETS paper titled Misfitting With AI, University of Michigan researchers Rahaf Alharbi, Pa Lor, Jaylin Herskovitz, Sarita Schoenebeck and Robin N. Brewer conducted a qualitative study with 26 blind people about how they verify and contest errors from AI visual assistance apps. They documented failures with complex document layouts, with languages other than English and with cultural artefacts. Participants had developed a repertoire of checks: testing the apps on familiar objects in low-stakes settings, drawing on non-visual knowledge, strategically involving sighted people and cross-referencing results between different apps and devices. The authors drew on the disability studies concept of “misfitting” to argue that these tools often carry assumptions that fail disabled users, and that disabled ways of knowing deserve recognition rather than replacement.

A diary study presented at CHI 2024 by Ricardo Gonzalez Penuela, Jazmin Collins, Shiri Azenkot and Cynthia Bennett tracked 16 blind and low vision participants using an AI scene description prototype for two weeks. The ratings were sobering. Average satisfaction came in at 2.76 out of 5. Average trust in the descriptions was 2.43 out of 4. The authors concluded that AI descriptions still needed significant improvement before they could deliver experiences that were both satisfying and trustworthy.

A 2025 CHI paper by Jingyi Xie, John M. Carroll and colleagues at Pennsylvania State University, which interviewed 14 visually impaired Be My AI users, adds texture. Four of its participants reported that Be My AI had described objects that were not there; one was told about an object behind her that did not exist. Eight reported inaccurate labels for gender or age. Several disliked the tool's habit of interpreting emotions and offering opinions. “Maybe I don't want it to editorialize, you know, maybe I just literally only want the facts of it,” one participant told the researchers. For live navigation, participants preferred human help over static photo descriptions, and they stressed that orientation and mobility skills remained essential. The tool, in their view, should augment those skills rather than replace them.

Chancey Fleet, the assistive technology coordinator at the New York Public Library's Andrew Heiskell Braille and Talking Book Library, has described a subtler cost. Speaking on the Electronic Frontier Foundation's How to Fix the Internet podcast in May 2024, she catalogued the house style of machine descriptions: “Any scene that contains a child or a dog is heartwarming.” “Any sunset or sunrise is vibrant.” “Anything with a couch and a lamp is calm or cozy.” The flattening matters, she argued, because adequacy discourages improvement: “if you are presented with what seems on a quick read, like good enough alt text, you're unlikely to do much labor to make it better, more nuanced, or more complete.” She also put the stakes plainly. When she asks an AI a question about her own world, “I'm going to have to live with the consequences of trusting that answer”.

Put these studies together and a pattern emerges. The capability is shared with everyone; the risk is carried by the individual. Each time a developer, publisher or institution decides that AI description is good enough, the job of verification is quietly transferred from the organisation that produced the content to the disabled person consuming it.

Overlays and the Business of Automated Compliance

Accessibility overlays are third-party scripts that sit on top of a website and attempt to detect and repair barriers automatically, often marketed with the language of artificial intelligence. They promise compliance with the Web Content Accessibility Guidelines (WCAG), the international technical standard, without anyone having to rebuild the site. Disabled users and accessibility professionals have been sceptical for years. In a January 2021 survey of 758 web accessibility practitioners by WebAIM, a non-profit based at Utah State University, 67 per cent rated overlays, plugins and widgets as “not at all or not very effective”. Among respondents with disabilities, the figure rose to 72 per cent, with only 2.4 per cent rating them very effective.

In January 2025, the US Federal Trade Commission turned that scepticism into enforcement. On 3 January it announced a proposed order requiring accessiBe, a company selling an AI-powered accessibility plug-in called accessWidget, to pay $1 million. The FTC alleged that accessiBe had misrepresented the tool's ability to make any website WCAG-compliant, and that it had formatted third-party articles and reviews to look like independent opinions without disclosing its material connections to them. The Commission voted 5-0 to accept the proposed consent order. Samuel Levine, then director of the FTC's Bureau of Consumer Protection, said: “Overstating a product's AI or other capabilities without adequate evidence is deceptive, and the FTC will act to stop it.” TechCrunch, reporting the action, noted that customers had also brought class action suits claiming the products did not work. The order was approved as final in April 2025.

AFB's Alexis Hubbard, writing in May 2025 under the title Missing a Human Touch, summed up the problem with automated fixes: “AI-powered accessibility overlays are just a band-aid solution for a bigger problem.” Because they do not change the source code of a site, she argued, they leave the underlying barriers intact. Her conclusion is worth quoting in full: “The only way to achieve an accessible and usable experience for all people is through people.”

Overlays show what AI-first accessibility looks like when it is driven by procurement rather than by users. A website owner buys the promise of compliance; a disabled visitor inherits the consequences, handed an interpretation instead of the thing itself.

When Governments Lean on the Promise of AI

The most striking recent example of AI's shortcomings being cited as a reason to delay accessibility comes not from a start-up but from the US Department of Justice.

In April 2024, the DOJ finalised a rule under Title II of the Americans with Disabilities Act requiring state and local governments to make their websites and mobile apps conform to WCAG 2.1 Level AA. Larger public entities, those serving populations of 50,000 or more, were originally due to comply in April 2026. Four days before that deadline, on 20 April 2026, the department published an interim final rule pushing compliance back by a year, to 26 April 2027 for larger entities and to 26 April 2028 for smaller ones and special districts. It took effect immediately, with public comments accepted until 22 June 2026. The technical standard itself was left unchanged.

Among the department's stated reasons was a striking admission about technology. “Advanced technology, such as generative AI, does not yet reliably automate the remediation of inaccessible content at scale,” the interim final rule says, “and staff resources and availability continue to pose significant challenges.” The rule also summarised correspondence from a member of Congress who argued that current technology, “including generative AI,” cannot reliably automate the remediation of complex science, technology, engineering and mathematics material at scale, and that human oversight is required.

It is an honest admission that AI cannot do the job alone. It also shows how the promise of automation can reshape policy. When the expected technological help did not arrive, the cost of the shortfall fell on disabled residents, who now wait at least another year for accessible public services.

Disability organisations were furious. Maria Town, president and chief executive of the American Association of People with Disabilities, called the delay “a profound disappointment and a serious setback for the disability community”, adding that “Years of notice have not been enough” and that the department was “rewarding inaction with more time.”

AFB's formal comments, submitted on 22 June 2026 and signed by Enyart, tackled the AI argument directly. “The absence of a fully automated accessibility solution does not justify extending the rule,” AFB wrote. It pointed out that generative AI can itself produce inaccessible content “because it is based on training data that contains accessibility barriers,” and that “Entities should not over rely on AI to produce content.” It also argued, drawing on its own research, that “AI adoption and effective use of AI is dependent on the accessibility of the platform.” In plain terms: “It doesn't matter how useful an AI-enabled service or tool is if the platform is inaccessible to blind users.”

That last line is the hinge of the entire debate. An AI tool is not a ramp into an inaccessible building. It is another door, and it only works if the corridor behind it has been built properly. AFB's comments also offered a constructive model: AI “can generate a passable first attempt at creating captions that a human can then edit.” Used that way, AI speeds up accessible design rather than replacing it.

AFB's research quantifies what is at stake. In a 2024 study cited in those comments, participants spent 55 per cent of their total task time dealing with digital barriers, and only 45 per cent on the task itself.

Across the Atlantic, the direction of travel is different. The European Accessibility Act, Directive (EU) 2019/882, began applying on 28 June 2025 to products and services including computers, smartphones, ATMs, e-books, e-commerce and banking. As Fredericka Argent of the law firm Covington & Burling noted that month, it reaches companies trading in the EU wherever they are based. The EAA is not an AI law, but it reinforces a principle that AI-first thinking tends to erode: the obligation sits with whoever builds and sells the product, not with the disabled person left to find a workaround.

Older Adults and the Gap Between Awareness and Use

The same dynamics extend to older adults, a group that overlaps heavily with disabled people and is often presented as a natural market for AI-powered help.

The latest AARP research, Tech Use and Adoption Keeps Surging Among Older Adults, written by Brittne Kakulla and published in December 2025, surveyed 3,838 US adults aged 50 and over between 9 September and 6 October 2025. It found that AI use in that age group had risen from 18 per cent in 2024 to 30 per cent in 2025. About half used or were interested in voice assistants such as Siri or Alexa. Two-thirds agreed that technology enriches life and makes ageing easier. Among adults aged 80 and over, agreement that technology enables a healthy life rose from 39 per cent to 46 per cent.

The same report records friction. For AI platforms such as ChatGPT or Copilot, 19 per cent used them, 21 per cent were interested and 59 per cent had no interest. Data privacy was the top barrier to adoption. Trust, privacy and data security were the top concerns for roughly half of the older adults who hesitated to adopt new technology, especially AI. And, in a line that could serve as the epigraph for this entire debate: “Three in 5 adults 50-plus say technology is not designed with their age in mind.”

Work from the MIT AgeLab sharpens the picture. A national survey of 1,008 American adults by Lauren Cerino, Joseph Coughlin and colleagues, published as a preliminary abstract in Innovation in Aging in December 2025, found that 73.7 per cent of respondents aged 60 and over were at least “a little” familiar with AI-enabled technologies. Only 28.3 per cent had used them. Some 34.7 per cent said they did not trust AI tools at all to give good advice, and trust scores declined with age.

Awareness, then, is not the main problem. Older people have heard of AI. The gap lies between hearing about a tool and being able, willing and confident enough to rely on it.

Research at Carnegie Mellon University points to why. In work presented at the ACM CHI 2026 conference in Barcelona and reported by the university in May 2026, Jodi Forlizzi, the Herbert A. Simon Professor in Computer Science and Human-Computer Interaction, and her colleagues found that older adults who use AI want it to recognise when they need plain information and when they need reassurance. A separate study based on 22 interviews identified shared values among older adults, including independence, curiosity, connection and continuity, expressed differently depending on where people lived. Older adults were open to AI when its purpose was clearly explained, especially for reminders, alerts and support at home. “There's always a mismatch between people's perceived abilities and what they can actually do,” Forlizzi said. “AI technology can be used to close this gap.”

That is the optimistic reading, and it deserves to be taken seriously. But it has a mirror image. If AI tools must be carefully explained and set up to earn trust, the people most likely to benefit, including those living alone or without technically confident relatives, are also the least likely to receive that support. AFB found that almost all its participants had learned to use AI on their own; only 9 per cent learned through an employer or online course. More disabled than nondisabled participants reported barriers in learning, most commonly inaccessible videos, hard-to-understand material and training that required a mouse.

The Northeastern study found similar problems in the AI products themselves. Blind participants encountered unlabelled buttons, missing headings and landmarks, no notification when a chatbot had finished responding, and clutter from adverts and sample prompts. The tools being promoted as accessibility solutions were not consistently accessible.

The Quiet Retirement of Predictable Tools

The sharpest version of the complaint concerns replacement: established, predictable tools being retired in favour of generative ones. In 2026, the most visible example is the end of Google Assistant on Android.

Google had originally planned to replace Assistant with Gemini, its generative AI assistant, during 2025. According to reporting by Matthias Bastian at The Decoder in August 2026, Google told users it would begin shutting Assistant down on Android and Wear OS from 4 September 2026, covering phones, tablets, watches, headphones and Android Auto. Once the rollout reaches a device, there is no way to switch back. Bastian raised exactly the concern at the centre of this article: that a probability-based language model may not match the reliability of a deterministic assistant for simple tasks such as smart home control and basic phone commands.

By late September, Mark Jansen at Android Police reported that the option to switch from Gemini back to Assistant had disappeared on the devices he checked, including a Pixel 10 Pro, and that smart home control was an area where Assistant still beat its replacement.

For most users, that is an inconvenience. For someone who cannot easily use a touchscreen, whether because of sight loss, limited dexterity or tremor, a voice assistant is not a convenience feature. It is how they turn on the lights, set a medication reminder or call for help. AFB's survey speaks to the stakes. Among users of voice-activated AI, 121 disabled participants described harms from mistakes, including texts sent with the wrong words and calls placed to unintended contacts, among them emergency services. “For many users, particularly disabled users, features such as voice dictation for text messaging are essential accessibility supports,” the report observes. “Failures or changes in these systems can result in concrete, disruptive, and sometimes harmful consequences.”

There is no evidence that Google set out to disadvantage disabled users, and Gemini brings generative capabilities that Assistant never had. The issue is structural. When a company replaces a predictable tool with a more capable but less predictable one, and removes the option to go back, the people with the fewest alternatives absorb the most risk. Capability is a fair trade for those who can fall back on other methods. Reliability is the whole point for those who cannot.

Is AI-First Accessibility Widening the Gap?

So, to the central question: if the people who would benefit most cannot find, set up or trust these tools, is AI-first accessibility widening the gap it claims to close?

The honest answer is that it depends on what “AI-first” means in practice.

Where AI is added on top of accessible foundations, the gap narrows. A blind person on a properly coded website who also has Be My AI or Seeing AI in their pocket has more independence than ever before. A deaf student who receives AI-generated captions that a human has corrected gets faster access than one who waits for manual transcription. These are genuine gains, and the AFB data confirms that most AI users say AI makes their lives easier.

Where AI replaces accessible foundations, the gap widens, and it widens along a predictable gradient. The confident, well-supported, technically fluent disabled user can verify outputs, switch between apps, ask a sighted friend and route around errors, exactly as the participants in the Michigan study described. The isolated, older, less confident or cognitively disabled user often cannot. The more a system depends on verification skills, the more it rewards people who already have them.

The evidence points to four ways in which substitution widens the divide. It shifts responsibility, transferring the work of accessibility from organisations to users who must detect errors they may have no way of seeing. It erodes incentives: as Fleet observed, good-enough descriptions discourage better ones, and the DOJ's interim rule shows the institutional version. It removes fallbacks, as predictable tools are retired. And it presupposes access to the tools themselves, when AFB and the Northeastern researchers both found that AI interfaces and training are often inaccessible.

None of this makes AI the villain. It makes the substitution the problem.

What Reliable Accessibility Would Require

If AI is going to close the accessibility gap rather than reshape it, the evidence points towards a set of concrete commitments.

Treat AI as a layer, never a foundation. Accessible design, meaning semantic code, human-authored alt text for meaningful images, accessible documents and keyboard-operable interfaces, must remain the baseline. AI description and remediation should sit on top of it. Katsarakes put it in a single sentence: “Let AI suggest or assist, but don't let it replace human intent.”

Keep the deterministic path available. When companies introduce generative assistants, they should keep predictable modes for core tasks such as timers, calls, reminders and device control, and they should not remove the option to revert until the replacement has been shown to match the original's reliability for disabled users specifically. A one-way migration is a design choice, not a technical necessity.

Build verification into the product. The Michigan and Penn State studies show users building their own verification systems from scratch. Products should do it for them: signalling uncertainty, separating what the model read from what it inferred, offering a factual mode without editorial colour, and making escalation to a human easy. Be My Eyes shows that a live human route can sit alongside AI.

Draw hard lines around high-stakes tasks. Medication instructions, financial documents, legal forms and navigation are not areas where a probabilistic description should ever be the only source. AFB's recommendations call for developers to “Improve the accuracy of AI outputs and tell users how accurate those outputs are.” Where accuracy cannot be assured, the interface should say so plainly.

Hold vendors to their claims. The FTC's accessiBe order established that overstating AI accessibility capabilities without evidence is deceptive. Regulators elsewhere should apply the same standard, and procurement teams should demand evidence of real-world performance with disabled users, not marketing claims of automated compliance.

Do not let automation set the deadline. Legal accessibility obligations should be pegged to what disabled people need, not to the pace of AI development. The European Accessibility Act's focus on the duties of manufacturers and service providers is a better model than one that treats immature automation as grounds for delay.

Fund support, not just software. Older adults are aware of AI but hesitant to use it, and CMU's research suggests they engage when the purpose is clearly explained. That explanation has to come from libraries, disability organisations and accessible community training. AFB's finding that only 9 per cent of AI users learned through formal routes suggests a large and fixable gap.

Include disabled people in testing and design. AFB's expert consensus report called for AI to focus “on expanding access and inclusion, not just on avoiding harm.” That requires disabled people in the room when products are designed, tested and retired, not consulted after launch.

What Is Lost When Reliability Goes

So what is lost when reliability is sacrificed for capability?

Part of the answer is practical: time, money, safety, and the occasional tube of medication with the wrong instructions. Participants in AFB's 2024 study already spent more than half their task time fighting digital barriers.

But the deeper loss is harder to measure. Independence is not just the ability to do something without help. It is the ability to rely on what you are told. A screen reader that occasionally mispronounces a name still gives its user a stable relationship with the text. A description tool that sometimes invents an object behind you, or turns a cream into tablets, asks its user to hold every answer at a slight distance. That is a different kind of life: one in which assistance is always provisional and confidence always has to be earned again.

Generative AI has given many disabled people extraordinary new reach, and nobody should want to take that away. The challenge for the companies building these tools, and the governments regulating them, is to add that reach without removing the floor. AI can help the people who build accessible things work faster and reach further. It cannot be allowed to replace the obligation they carry.

References and Sources

  1. American Foundation for the Blind. “The AI Quagmire: Benefits, Risks, and User Aspirations Through a Disability Lens.” AFB Artificial Intelligence Series, 2026. Sections “Executive Summary” and “Harms of AI in Various Contexts.” https://afb.org/research-and-initiatives/ai-series/ai-quagmire/executive-summary and https://afb.org/research-and-initiatives/ai-series/ai-quagmire/harms-ai-various-contexts
  2. Katsarakes, Alex. “Beyond Alt Text: Rethinking Visual Description in the Age of AI.” American Foundation for the Blind blog, 16 October 2025. https://afb.org/blog/entry/alt-text-age-ai
  3. Hubbard, Alexis. “Missing a Human Touch: Why Automated Testing and AI Don't 'Solve' Digital Accessibility… Yet.” American Foundation for the Blind blog, 13 May 2025. https://afb.org/blog/entry/missing-human-touch-1
  4. Enyart, Stephanie, American Foundation for the Blind. “June 2026 Comments on DOJ Extension of Web Accessibility Compliance Dates” (RIN 1190-AA82, Docket CRT150). 22 June 2026. https://afb.org/research-and-initiatives/public-policy/advocacy-letters-comments-statements/comments/june-2026-doj-extension-web-compliance
  5. American Foundation for the Blind. “AFB Report Spotlights Impact of AI for Disabled People” (press release on Empowering or Excluding: Expert Insights on Inclusive Artificial Intelligence for People with Disabilities). 6 February 2025. https://afb.org/news-publications/press-room/press-release-archive/press-release-2025/afb-report-spotlights-impact-AI-disabled-people
  6. US Department of Justice, Civil Rights Division. “Extension of Compliance Dates for Nondiscrimination on the Basis of Disability; Accessibility of Web Information and Services of State and Local Government Entities.” Interim final rule, Federal Register Vol. 91, No. 75, 20 April 2026. https://www.federalregister.gov/documents/2026/04/20/2026-07663/extension-of-compliance-dates-for-nondiscrimination-on-the-basis-of-disability-accessibility-of-web (PDF: https://www.ada.gov/assets/pdfs/2026-ifr.pdf)
  7. Duane Morris LLP. “DOJ Extends ADA Title II Digital Accessibility Deadlines by One Year.” Client alert, 23 April 2026. https://www.duanemorris.com/alerts/doj_extends_ada_title_ii_digital_accessibility_deadlines_one_year_0426.html
  8. American Association of People with Disabilities. “AAPD Statement on DOJ Title II Web Rule Interim Final Rule.” 20 April 2026. https://www.aapd.com/aapd-statement-title-ii-doj-web-rule-ifr/
  9. Kakulla, Brittne. “Tech Use and Adoption Keeps Surging Among Older Adults” (2026 Tech Trends). AARP Research, December 2025. https://www.aarp.org/pri/topics/technology/internet-media-devices/2026-technology-trends-older-adults/
  10. Cerino, Lauren; Felts, Adam; Balmuth, Alexa; Brennan, Taylor; Lee, Chaiwoo; Coughlin, Joseph. “Older Adults' Perspectives on Artificial Intelligence as a Source of Advice.” Innovation in Aging, 9 (Suppl 2), igaf122.1371, 31 December 2025. https://pmc.ncbi.nlm.nih.gov/articles/PMC12761853/
  11. Sapio, Amanda. “How Do Boomers Really Feel About AI?” Carnegie Mellon University School of Computer Science News, 28 May 2026. https://www.cs.cmu.edu/news/2026/older-adults-ai
  12. US Federal Trade Commission. “FTC Order Requires Online Marketer to Pay $1 Million for Deceptive Claims that its AI Product Could Make Websites Compliant with Accessibility Guidelines.” Press release, 3 January 2025. https://www.ftc.gov/news-events/news/press-releases/2025/01/ftc-order-requires-online-marketer-pay-1-million-deceptive-claims-its-ai-product-could-make-websites-compliant
  13. Wiggers, Kyle. “FTC orders AI accessibility startup accessiBe to pay $1M for misleading advertising.” TechCrunch, 3 January 2025. https://techcrunch.com/2025/01/03/ftc-orders-ai-accessibility-startup-accessibe-to-pay-1m-for-misleading-advertising/
  14. WebAIM. “Survey of Web Accessibility Practitioners #3 Results.” January 2021. https://webaim.org/projects/practitionersurvey3/
  15. Alharbi, Rahaf; Lor, Pa; Herskovitz, Jaylin; Schoenebeck, Sarita; Brewer, Robin N. “Misfitting With AI: How Blind People Verify and Contest AI Errors.” Proceedings of ASSETS '24, ACM, 2024. DOI: 10.1145/3663548.3675659. https://arxiv.org/abs/2408.06546
  16. Gonzalez Penuela, Ricardo E.; Collins, Jazmin; Azenkot, Shiri; Bennett, Cynthia. “Investigating Use Cases of AI-Powered Scene Description Applications for Blind and Low Vision People.” Proceedings of CHI '24, ACM, 2024. https://arxiv.org/abs/2403.15604
  17. Adnin, Rudaiba; Das, Maitraye. “'I look at it as the king of knowledge': How Blind People Use and Understand Generative AI Tools.” Proceedings of ASSETS '24, ACM, 2024. DOI: 10.1145/3663548.3675631. https://maitraye.github.io/files/papers/BLV_GenAI_ASSETS24.pdf
  18. Xie, Jingyi; Yu, Rui; Zhang, He; Billah, Syed Masum; Lee, Sooyeon; Carroll, John M. “Beyond Visual Perception: Insights from Smartphone Interaction of Visually Impaired Users with Large Multimodal Models.” Proceedings of CHI '25, ACM, 25 April 2025. DOI: 10.1145/3706598.3714210. https://pmc.ncbi.nlm.nih.gov/articles/PMC12338113/
  19. Cohn, Cindy; Kelley, Jason (hosts), with Chancey Fleet. “Podcast Episode: Building a Tactile Internet.” Electronic Frontier Foundation, How to Fix the Internet, 7 May 2024. https://www.eff.org/deeplinks/2024/05/podcast-episode-building-tactile-internet
  20. Be My Eyes. “Be My Eyes Reaches 1 Million Blind and Low-Vision Users and 10 Million Volunteers.” 12 March 2026. https://www.bemyeyes.com/news/be-my-eyes-reaches-1-million-blind-and-low-vision-users-and-10-million-volunteers/
  21. Shaikh, Saqib. “Seeing AI App Launches on Android, Including New and Updated Features and New Languages.” Microsoft Accessibility Blog, 4 December 2023. https://blogs.microsoft.com/accessibility/seeing-ai-app-launches-on-android-including-new-and-updated-features-and-new-languages/
  22. Bastian, Matthias. “Google will shut down Google Assistant starting September 2026 as Gemini takes over on Android and Wear OS.” The Decoder, 5 August 2026. https://the-decoder.com/google-will-shut-down-google-assistant-starting-september-2026-as-gemini-takes-over-on-android-and-wear-os/
  23. Jansen, Mark. “Google Assistant is officially gone, and people aren't happy about it.” Android Police, 29 September 2026. https://www.androidpolice.com/the-google-assistant-kill-off-has-arrived/
  24. Argent, Fredericka. “European Accessibility Act: June 2025 Deadline Has Arrived.” Covington & Burling, Inside Global Tech, 10 June 2025. https://www.insideglobaltech.com/2025/06/10/european-accessibility-act-june-2025-deadline-has-arrived/ (see also European Commission, “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)

Tim Green

Tim Green UK-based Systems Theorist & Independent Technology Writer

Tim explores the intersections of artificial intelligence, decentralised cognition, and posthuman ethics. His work, published at smarterarticles.co.uk, challenges dominant narratives of technological progress while proposing interdisciplinary frameworks for collective intelligence and digital stewardship.

His writing has been featured on Ground News and shared by independent researchers across both academic and technological communities.

ORCID: 0009-0002-0156-9795 Email: tim@smarterarticles.co.uk

Listen to the free weekly SmarterArticles Podcast

Support What We Do

Discuss...