As someone who relies on assistive technology, I know the value of tools that remove unnecessary effort. Good technology can turn a frustrating series of steps into something quick, private and independent. Artificial intelligence (AI) promises more of that by describing images, summarising complicated information and helping people navigate systems that were never designed with them in mind.

But there is an important difference between a tool helping me and a system quietly taking control. That difference is agency.

What are we at risk of losing?

Brad Frost recently wrote about our response to AI through the lens of grief. He explores how people may be grieving the loss of certainty, professional identity, stability and familiar ways of working. One line particularly struck me:

“It’s just an acknowledgment that AI is here; it’s a reality.”

Brad Frost, independent designer and developer

Frost is not suggesting that acceptance means approving of everything being done with AI. It means recognising what has changed and deciding how to respond. That prompted me to think about a particular kind of loss that matters in accessibility: the loss of control over how we complete a task, make a choice or challenge a decision.

AI can increase independence, but it can also make decisions and take actions that are difficult to understand, correct or reverse. The distinction is not always obvious to the person using it.

Assistance can quietly become authority

An assistant expands your ability to act. An authority decides what happens next. That distinction becomes blurred when an AI system moves from suggesting an action to completing it.

A tool might help someone fill in a form, but does it clearly show what information it has entered before submitting it? An assistant might summarise several options, but does it explain what was omitted or why one option was recommended? These questions become more serious when AI is used to interpret eligibility, prioritise customer enquiries, assess applications or personalise the information someone receives.

In each case, the interface may still look helpful. It may even be technically accessible. Yet the person using it can gradually lose the ability to understand the process and influence the outcome.

The problem is not automation by itself. It is automation without meaningful control.

Accessibility cannot stop at operability

The Web Content Accessibility Guidelines, usually known as WCAG, provide an essential foundation for digital accessibility. They address requirements such as keyboard access, programmatic names for controls, status messages and opportunities to review or correct important submissions. Those requirements remain just as important when an interface includes AI.

We still need to ask whether someone can reach and operate the controls with a keyboard. We need to know whether changes are announced by a screen reader, whether instructions are understandable and whether errors can be identified and corrected. However, an AI system can satisfy all those requirements and still leave a user without agency.

A screen reader user might be able to operate an AI assistant perfectly while remaining unable to discover what information it used, inspect a proposed action or understand why it produced a particular result. That is an accessible interaction in the narrowest technical sense, but it is not necessarily an equitable experience.

Accessibility should mean that people with disabilities can perceive, understand, navigate and interact with technology. It should also mean that they can make informed choices, contribute on equal terms and recover when something goes wrong.

The risks are not shared equally

When an AI assistant makes a mistake, one person may experience a minor inconvenience. For someone relying on that assistant to interpret visual information or navigate an otherwise inaccessible process, the same mistake can remove their independence entirely.

People with disabilities can then face a poor choice: use an imperfect AI tool or return to a process that was already inaccessible. That is not meaningful choice.

There is also a risk that organisations see AI as a convenient replacement for human support. A chatbot may reduce the number of routine enquiries reaching a contact centre, but it should not become a barrier between a customer and the person who can resolve their problem. Making the chatbot keyboard accessible does not solve that wider issue.

If an organisation encourages customers to depend on AI, it assumes responsibility for accessible alternatives, clear escalation and recoverable mistakes. Otherwise, it has simply transferred effort and risk from the organisation to the customer. That affects more than accessibility. It affects trust.

Five rights for accessible AI

The National Institute of Standards and Technology describes its AI Risk Management Framework as rights-preserving and focused on managing risks to individuals, organisations and society. We can bring that thinking into product requirements through five practical rights.

1. The right to know

People should be told when AI is generating content, interpreting information or influencing an outcome. That explanation needs to be available in an accessible format and written in language the intended audience can understand. A vague reference hidden in lengthy terms and conditions is not enough.

2. The right to choose

People should be able to decide whether they want AI assistance where a genuine choice is possible. Declining AI should not force someone into a less accessible route, nor should people be expected to surrender unnecessary personal information in exchange for an accessible experience.

3. The right to inspect

Users should be able to review important information before an AI system submits it or acts upon it. That includes the opportunity to check generated text, entered data and proposed actions. For consequential decisions, organisations should also explain the relevant factors in a way that a person can understand and challenge.

4. The right to correct

AI will make mistakes, so accessible design must assume that rather than treating errors as unusual exceptions. People should be able to amend information without starting again, while important actions should be confirmed before completion and reversible wherever possible. Error recovery must work with assistive technology just as reliably as the main journey.

5. The right to reach a person

Human support should be a designed part of the service, not a hidden emergency exit. The route to a person should be easy to find, accessible and appropriate to the urgency of the task. Users should not have to repeatedly prove that the AI cannot help them before they are allowed to move on.

Test the outcome, not just the interface

These five rights can become practical requirements, but they also need to be tested with people with disabilities. Do not test only whether someone can send a prompt and receive an answer. Ask whether they noticed that AI was involved, understood what it proposed, could change its output and knew how to reach a person.

Most importantly, find out whether the system left them more independent or simply more dependent on automation they could not examine. Standards and audits give us a necessary baseline, but real user testing tells us whether the experience preserves dignity, choice and control.

Acceptance should lead to commitment

Brad Frost is right that accepting AI does not require us to approve of all its uses or ignore its risks. Acceptance can be the point at which we decide what we are prepared to protect. For accessible AI, that must include agency.

The best AI tools will not simply complete more tasks for people. They will help people understand their choices, remain in control and act independently. If a system is easy to use but difficult to question, correct or escape, it is not genuinely accessible.

Sources

• Brad Frost, “Grief in the AI Age”, 18 May 2026.
• World Wide Web Consortium, Web Content Accessibility Guidelines 2.2.
• W3C Web Accessibility Initiative, Accessibility, Usability and Inclusion.
• W3C Web Accessibility Initiative, Artificial Intelligence and Accessibility Research Symposium.
• National Institute of Standards and Technology, Artificial Intelligence Risk Management Framework 1.0.