Artificial intelligence (AI) is now part of many of the digital services people use. People may encounter it while using a website or another online service without knowing exactly where AI is involved. The same is true of many smart devices, which are internet-connected products that use software to gather or exchange information and perform functions for the person using them.
AI is no longer only a tool someone opens on purpose to draft an email, create an image, or answer a question. It may already be part of the service someone is using. This makes it harder to separate a deliberate decision to use AI from using the service itself.
On September 10, 2026, concern about AI safety grew after Anthropic researcher Jacob Coxon said he had resigned and warned that people developing AI believed the technology could pose a risk to human life within the decade. Anthropic scientist Evan Hubinger publicly agreed with the concern.
Their warnings followed reports of AI agents acting outside human instructions and safety researchers leaving major companies because of risks they believed were not being addressed quickly enough. The concerns were not limited to one possible outcome. They included deliberate misuse, systems acting in ways developers did not expect, and whether companies and government had adequate ways to test and respond before harm occurred (Reuters, 2026a).
On September 11, 2026, Reuters reported that U.S. Senate negotiators were considering legislation that would place a “duty of care” on developers of certain AI systems. The proposal under discussion would require companies to address known catastrophic risks and could allow the federal government to block the release of a system considered unsafe. The legislation was still being negotiated, but the discussion reflected a larger question about who is expected to identify a problem, who has the ability to stop it, and what happens after a system has already been released (Reuters, 2026b).
Those concerns involve risks far beyond the access issues discussed here. The connection is not the scale of the harm. It is the question of what happens when people rely on technology that produces an unexpected result and the person using it has to recognize that something went wrong.
For people with disabilities, the problem may begin with a website, a caption, an online form, or a customer service system. The consequences may be less dramatic than the broader risks now under discussion, but they can determine whether someone gets accurate information, completes a process, or reaches the person or service they need.
A tool that summarizes a document can make a long piece of information easier to work through. Captions can provide access to a video, and an image description can supply information that would otherwise be missed. Each can also leave out something the person needed.
AI can miss context or repeat a stereotype. It can collect information a person did not expect to disclose. It can also give an incorrect answer without making the error obvious.
The problem is not simply whether the technology works. The harder question is whether the person using it can recognize an error, understand what information may have been lost, and correct the problem without creating another barrier.
People with disabilities can encounter problems when technology is designed around assumptions that do not match how they use it. A system may misread speech, movement, or the way someone completes a task. Those problems existed before AI. Websites, forms, software, and automated systems have long created barriers when accessibility was not considered during design.
AI can create the same kinds of barriers. It can repeat an inaccessible design choice, make an error harder to identify, or place another layer between a person and the information or service being sought.
People are already using these systems without always knowing how they work or what information they collect. It may also be unclear how well the system recognizes different ways of speaking, writing, moving, or completing a task. The concern becomes greater when AI is part of a service someone needs to use rather than a tool someone can simply close.
People may not know how well an AI system has been tested with people who speak, move, communicate, or use technology differently from the users it was designed around. A speech tool may have trouble with atypical speech. A system that relies on timing may misread a slower response. Technology that expects a certain type of movement may also interpret tremor, spasticity, stillness, or the use of mobility equipment incorrectly.
The American Foundation for the Blind’s 2026 research found that people with disabilities are using AI to move information from one format to another, describe images, support writing, and make some digital information easier to use. The same research identified several barriers (American Foundation for the Blind, 2026).
Accuracy is one of them. A summary, caption, or description may leave out information the person needed or present something incorrectly. The problem becomes harder to catch when the person is relying on AI because the original format was already difficult to access.
A second barrier is whether the technology works with the way a person communicates or uses a device. Speech recognition may not interpret atypical speech correctly, while other systems may not work well with screen readers, keyboard navigation, or the extra time someone needs to complete a task.
Privacy is the third concern. A person may provide voice, written information, images, or accessibility settings in order to use a tool without knowing how much of that information is stored, reused, or used to make additional inferences. The benefit of easier access can come with questions the person did not expect to answer simply by using the service.
Recent research on AI product development found that accessibility concerns and broader AI safety work are often handled by different teams or treated as separate areas of responsibility (Moharana et al., 2025). Problems can be missed when no one is clearly responsible for issues that involve both. By the time an accessibility concern is recognized, important design decisions may already have been made.
AI does not merely deliver information. It often rewrites it. A system may summarize a document or translate language. It may caption speech, describe an image, or simplify text. Each use involves some degree of interpretation.
A caption can be accurate enough to follow and still miss the word that changes the meaning of a sentence. An image description may identify a person standing outside but fail to mention that the ramp behind them is blocked. For someone using the description to understand the photograph, the missing detail changes what the image communicates.
Summaries can create a similar problem. A shorter version of a document may be easier to read but leave out a warning, a condition, or another detail that changes how the information should be understood. The wording may remain clear even though part of the meaning has been lost.
A person who relies on captions, summaries, or descriptions may not have an easy way to compare the AI output against the original. Someone using a tool to reduce fatigue may end up spending energy checking whether the tool created a new problem. The work shifts back to the person who used the technology because it was supposed to make the information easier to use.
AI can also change what a person has said. If someone submits a complaint and the summary leaves out the sentence describing the main problem, the record no longer reflects the complaint accurately. A chatbot can create a different problem when the available responses do not match what the person is trying to explain.
For people with disabilities, these problems affect whether information can be used as intended. AI-generated alt text can leave a blind or low-vision reader without enough information to understand an image. A search summary can also leave out the source or other information the reader would need to check the answer. A chatbot can sound polite without answering the question. More content is not the same as better communication. Clear language can still contain an error.
The information supplied to AI creates another concern. Many AI tools collect, process, or infer information that may reveal more than the user intended to disclose. Voice data may reveal a speech disability, while accessibility settings can reveal information about how someone uses a device. A person may provide information for one purpose without realizing what else the system can infer from it.
A person may also have little choice about whether to use the tool. Someone who needs a particular feature to read information, communicate, or complete a required process may have no equally workable alternative. Agreeing to the privacy policy does not necessarily mean the person was comfortable providing the information. It may simply mean using the service required it.
A privacy notice may explain what a company is permitted to do without answering what the person using the service most wants to know. Who has access to the information? How long will it be kept, and can it be used for another purpose? The option to refuse carries little meaning if refusing also removes the only practical way to use the service.
The National Institute of Standards and Technology (NIST) is part of the U.S. Department of Commerce. In 2023, NIST published an AI Risk Management Framework to help organizations identify and manage problems that can arise when AI is developed and used.
The framework is not an accessibility standard. It looks at whether an AI system works as intended, how personal information is handled, and whether people can understand how the system is being used. It also addresses harmful bias (NIST, 2023).
In 2024, NIST published additional guidance for generative AI. The guidance addresses problems that can arise when AI creates or changes content, including inaccurate information and privacy concerns (NIST, 2024). For someone using AI as an accessibility tool, the questions are basic. Is the information correct? What happened to the information given to the system? If the result is wrong, can it be corrected or checked against another source?
AI is often promoted as a way to make information easier to reach or use. For some people with disabilities, it already does that. A person may use AI to put information into another format, get help describing an image, or work through material that would otherwise take more time or effort. The benefit can come with tradeoffs. A tool may make information easier to access while collecting more personal information than the person expected to provide. It may also work well in one part of the task and still be difficult to use with a screen reader, keyboard navigation, an older device, or a slower internet connection.
Some of those problems may not become clear until the product is used by people who were not well represented during testing. A system can work as expected during development and still have trouble with speech, movement, or access methods that were not considered closely enough. By then, the product may already be in use.
People with disabilities can identify problems that may not be apparent to the people developing the technology. Their experience is useful while changes can still be made, rather than only after an accessibility problem has been reported.
Testing is not the only issue. AI can also reflect narrow ideas about what disability looks like. A 2025 WIRED investigation found that when OpenAI’s Sora video generator was prompted to show a person with a disability, all 10 results showed people who used wheelchairs, and none were shown in motion. The generated titles also often described the person as “inspiring” or “empowering,” even when the person was simply present in the scene (WIRED, 2025).
People with disabilities are also parents, students, artists, customers, travelers, professionals, neighbors, and decision-makers. When AI repeatedly returns to the same narrow images, it can reinforce the assumptions already present in the material used to train it.
Accessibility requirements exist whether a service uses AI or not. In 2024, the U.S. Department of Justice finalized a rule requiring covered web content and mobile apps provided by state and local governments to meet WCAG 2.1 Level AA under Title II of the Americans with Disabilities Act (U.S. Department of Justice, 2024).
In 2026, the Department extended the compliance dates to April 26, 2027, for public entities with populations of 50,000 or more and to April 26, 2028, for public entities with populations below 50,000 and special district governments (U.S. Department of Justice, 2026). AI features are now appearing inside many of the same websites, apps, portals, documents, and digital services.
A newer interface does not make an inaccessible service accessible. If the underlying page cannot be navigated or information is unavailable to a screen reader, adding AI does not resolve the original problem. The same is true when a person still cannot complete the required process. AI may simply change the point at which the barrier appears.
AI can support access when it gives someone a workable way to receive information, complete a task, communicate, or reach assistance. Problems arise when the automated route becomes the only practical route and the technology cannot understand the person’s question or preserve what the person meant. The ability to reach someone who can correct the problem remains important.
A summary that changes meaning is not reliable. A caption that cannot be trusted does not provide meaningful access. A form that forces someone to choose an answer that does not fit creates another barrier. Testing is important, but what happens after a problem is discovered matters just as much.
When a system misreads someone, there needs to be a practical way to correct the error. If a summary changes meaning, the original source still needs to be available. When a platform collects sensitive information, the person using it should be able to understand why the information is needed and what choices remain.
People with disabilities can end up identifying accessibility problems only after a product has already been released. Repeatedly explaining why vague alt text does not provide enough information or why a chatbot loop does not resolve a problem adds work to a process that may have been intended to make something easier. The same is true when a form offers no accurate answer but requires the person to choose one anyway.
The ways people with disabilities already use AI should not be ignored. These tools can help someone read material that was difficult to access in its original form or put information into a format that works better for them. They can also make writing, communication, or understanding an image easier. Those uses are part of the reason the questions in this article matter.
The debate that intensified in September 2026 focused on risks that could affect national security, public safety, and the way AI systems are governed. This article has focused on more basic questions. Is the information accurate? Can a person tell when something has been left out or changed? What information is being collected, and is there another way to complete the task when the technology gets it wrong?
The scale is different, but the underlying concern is similar. People are being asked to rely on AI before all of the questions about its accuracy, limits, and safeguards have been settled. For people with disabilities, those questions may begin with a caption, an image description, a form, or a customer service system. The larger debate asks what happens when AI causes serious harm. A more immediate question is what happens when someone needs the technology to work and it does not.
Disclaimer
This article is for general informational and educational purposes only and reflects the author’s perspective. It is not legal, medical, technological, or other professional advice. The article does not assess the obligations or practices of any specific person, organization, product, or service. Laws, accessibility standards, privacy practices, and AI technologies continue to change. Readers should consult current authoritative guidance or a qualified professional when addressing a specific situation.
References
- American Foundation for the Blind. (2026). The AI Quagmire: Benefits, Risks, and User Aspirations Through a Disability Lens. American Foundation for the Blind.
- Moharana, S., Bennett, C. L., Buehler, E., Madaio, M., Tibdewal, V., & Kane, S. K. (2025). “Accessibility people, you go work on that thing of yours over there”: Addressing disability inclusion in AI product organizations. Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, 8(2), 1724-1737.
- National Institute of Standards and Technology. (2023). Artificial Intelligence Risk Management Framework 1.0. U.S. Department of Commerce.
- National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. U.S. Department of Commerce.
- Reuters. (2026a, September 10). More US lawmakers seek new AI rules after Anthropic researchers warn of human extinction.
- Reuters. (2026b, September 11). US Senate negotiators consider requiring AI firms to mitigate known major risks.
- Rogers, R., & Turk, V. (2025, March 23). OpenAI’s Sora is plagued by sexist, racist, and ableist biases. WIRED.
- U.S. Department of Justice. (2024). Fact sheet: New rule on the accessibility of web content and mobile apps provided by state and local governments.
- U.S. Department of Justice. (2026). 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.
Join the conversation. Your thoughts and experiences are welcome.