A patient may be able to read every word on an AI interface and still be unable to use the service. They might struggle to express the problem in the expected format, understand a question, correct a mistaken summary or reach a person. Accessibility should therefore be assessed across the whole task, not reduced to the appearance of the screen.
Start with the person and the task
Consider a fictional adult who communicates using a text-based aid. They can describe their concern clearly in writing, but a service requires an uninterrupted spoken conversation with its assistant before a clinician will review the request. The problem is not a lack of clinical information. It is the requirement to provide it through one particular mode.
Now consider another patient for whom sustained typing is difficult but speech is practical. Replacing the entire service with a written form would not solve both problems. A design that works well for one person can remain inaccessible to another.
The GMC's patient partnership and communication standards, reviewed on 10 October 2026, emphasise responding to individual needs and asking what support is required. An organisation should not assume a preferred format from a diagnostic label or treat one adjustment as a universal solution.
Distinguish technical access from successful use
W3C's Web Content Accessibility Guidelines 2.2, reviewed on 10 October 2026, cover areas including perceivable information, operable controls, understandable content and compatibility with assistive technologies. These are important technical foundations, not evidence that a patient can necessarily complete a particular clinical interaction safely.
A button may be accessible to a screen reader while its label remains ambiguous. A form may support keyboard navigation while forcing an answer the patient does not understand. A summary may be readable while giving no clear way to correct an inaccurate statement before submission.
The service should therefore test a meaningful endpoint. Can the person communicate the concern, understand what happens next and obtain help when the expected interaction does not work? Technical conformance and task success answer related but different questions.
This is not an argument against standards. It is an argument for applying them within a wider evaluation of the actual care pathway.
Make uncertainty an acceptable response
A proposed intake design should allow the patient to indicate uncertainty without treating it as an obstacle to be overcome through guessing. "I do not know", "I cannot describe it that way" and "that is not what I meant" can contain important information about the limits of the interaction.
For the fictional user of a communication aid, the system might need more time or a different sequence of questions. The goal should be an accurate account, not completion at the speed assumed by the interface.
W3C's supplemental guidance on content for people with cognitive and learning disabilities addresses understandable tasks, support and user involvement in design. As reviewed on 10 October 2026, it is a Working Group Note, not a separate certification of a healthcare product.
An organisation can use those principles to ask better questions without claiming that a particular interface has been validated for every user. Simplifying the interaction must also preserve clinically important distinctions rather than remove them because they are harder to express.
Correction is part of access
A patient-facing AI may summarise the information before passing it to a clinician. A proposed accessibility assessment should examine whether the patient can meaningfully review that summary, not merely click a button labelled "confirm".
For some users, a long block of text may be difficult to check. Others may need an alternative presentation or help understanding the meaning of a clinical term. Any adaptation should make the original information easier to inspect without creating a different clinical account.
A useful fictional test is to insert a misunderstanding into the summary and ask whether the user can identify and correct it. The evaluation should examine the correction route and whether the final record reflects the change. It should not blame the patient when the interface makes the task unclear.
Professional review remains necessary where required. Patient confirmation can add valuable information, but it should not become the service's excuse for transferring responsibility for generated errors to the user.
Human help must be reachable before the process fails
An alternative route is of limited value if it appears only after the patient completes the inaccessible steps. A proposed design should make support available when the barrier is encountered, with an explanation of what information will be carried across.
The handover should preserve the patient's effort. Requiring them to repeat the entire account can create additional burden, although a clinician may still need to clarify or verify important details. The distinction between appropriate clinical questioning and avoidable repetition should be assessed from both sides of the interaction.
The route must also exist operationally. A visible "contact us" button is not evidence that someone receives, prioritises and responds to the concern. The organisation should test the service behind the control, including when the original pathway is abandoned.
These are proposed service standards, not a claim that every AI product currently has the same barriers or that a particular communication channel is always preferable.
Evaluate with affected users, not only simulated personas
Fictional scenarios can expose obvious design assumptions, but they do not replace participation by people with relevant lived experience. A proposed evaluation should include users with different communication, sensory, cognitive and motor needs, without treating any individual as representative of everyone sharing a label.
Observe where the task becomes difficult, what assistance is requested and whether the final information accurately represents the user's concern. Include people who stop, switch channels or need support. A completion rate calculated only among successful users cannot reveal the barrier faced by those excluded earlier.
Healthwatch's research on AI scribes published on 16 July 2026 also illustrates why patient perspectives require direct investigation. Its polling and qualitative accounts should not be collapsed into a universal preference or a population-wide error estimate.
The proposed objective is neither maximum digital uptake nor minimum support use. A patient reaching an appropriate person through an alternative route can be a successful outcome.
Apply the same standard to learning tools
Clinicians and students also differ in how they access educational material. As described by iatroX in October 2026, its simulations offer voice and text, while its Tutor supports guided discussion of an attempted question. Those formats may suit different preferences, but availability of a format is not proof of accessibility or educational benefit for every learner.
This article is published by iatroX, and the same task-based assessment should apply to its products. Can the learner start the intended activity, understand the feedback, challenge an error and return after interruption? These questions are more useful than declaring an interface universally "easy".
For healthcare services, the corresponding standard is practical: the patient can communicate, understand, correct and obtain help. A larger font can contribute, but it cannot carry that responsibility alone.
Frequently asked questions
Does an accessible website guarantee an accessible AI care pathway?
No. Technical access must be considered alongside the interaction, clinical communication, correction process and availability of appropriate support.
Is voice always more accessible than text?
No. The suitable format depends on the individual, the task and the circumstances, so services should avoid assigning a single preference to an entire group.
Should asking for human help count as a failed AI interaction?
Not necessarily. Recognising that another route is needed and reaching appropriate help can be the safest and most useful outcome.
Record a learning need about accessible clinical communication →