# WCAG Understandable

Understandable is the third WCAG principle. Users can follow what the content means and can predict how the interface will behave. AI strains this in two ways. The output is written by a model and can be dense or strange, and the behavior is non-deterministic, so the same action can give a different result next time.

## Key points

- Understandable means users can follow what the output means and predict how the interface behaves.
- Offer plain wording by default and let users ask for a simpler version.
- Keep controls consistent even though the generated content keeps changing.
- WCAG has no criterion for uncertainty yet, but if an answer might be wrong, say so in text, not only with faint styling.
- Help users avoid and recover from mistakes, including a way to review and reverse actions an agent takes.

## Readable output

AI output should be plain, under the 3.1 criteria. Long, jargon-heavy answers are hard for many readers, including people with cognitive disabilities, people reading in a second language, and anyone who is tired or rushed. Offer plain wording by default and let users ask for a simpler version. Short sentences and clear structure help everyone, not only disabled users.

## Predictable behavior

Controls should behave consistently even though the content keeps changing, under the 3.2 criteria such as consistent navigation and consistent identification. Do not let a regenerate, a new model version, or a fresh answer quietly move the controls around or rename them. When the content itself is unpredictable, a stable frame around it matters more, not less.

## Signaling uncertainty

WCAG has no success criterion for this yet, but Understandable is where it belongs. If an answer might be wrong, the interface should say so in text, not only with faint or grayed styling that a screen reader will not convey. Confident wrong output is a known and structural trait of language models, not a rare slip. One paper argues that these falsehoods are best understood as [indifference to truth](https://link.springer.com/article/10.1007/s10676-024-09775-5) rather than as occasional errors, and other research has shown model reasoning [breaking down](https://www.theguardian.com/technology/2025/jun/09/apple-artificial-intelligence-ai-study-collapse) as problems get harder. Plan for the wrong answer and make sure every user can tell when to be careful. See [Uncertainty and confidence](https://artificia11y.ds.house/patterns/uncertainty-and-confidence/).

## Explainability

People often need to understand why a system produced a given answer or decision, especially when the output affects them in a real way. This is the field of explainable AI. The US National Institute of Standards and Technology sets out four principles for it, that an explanation is given, that it is meaningful to the person receiving it, that it accurately reflects how the system reached the result, and that the system stays within the limits of what it can reliably explain. Accessible explanations have to reach assistive technology users too. See [Decision outputs](https://artificia11y.ds.house/patterns/decision-outputs/) and [NIST AI RMF](https://artificia11y.ds.house/guidelines/nist-ai-rmf/).

## Input help and error recovery

Help users avoid mistakes and fix them, under the 3.3 criteria. For AI inputs this means clear labels on the prompt field and its controls, error messages that explain what to do next, and a way to undo or correct an action. When an agent acts on a user's behalf, the chance to review and reverse the action is part of input assistance, not a separate feature.

> [!PRODUCT-MANAGER]
>
> Turn uncertainty into acceptance criteria. For any answer that could be wrong in a way that matters, require a visible text signal of confidence or a prompt to verify, and require that the signal reaches screen reader users rather than being carried by color alone. The Air Canada case, where an airline was held responsible for its chatbot's wrong answer, is a useful reminder that your organization owns what the model says.

> [!ACCESSIBILITY-SPECIALIST]
>
> Uncertainty and explainability are the clearest places where WCAG runs out. There is no criterion to cite, so frame findings against Understandable as intent, against the AI Act transparency duties, and against the NIST explainability principles. Check that any confidence cue, source citation, or "this may be wrong" notice is present in the accessibility tree and not implied only by styling.

> [!SELF-ADVOCATE]
>
> Being asked to trust an answer you cannot independently check is uncomfortable for anyone, and it is worse when the only hint that something is shaky is a faint gray font you cannot see or a tone you cannot infer. Output that states plainly when it is unsure, and that explains where an answer came from, is far easier to rely on or to reject.

## Further reading

- Hicks, Humphries, and Slater on [why model falsehoods are more than slips](https://link.springer.com/article/10.1007/s10676-024-09775-5).
- IBM's overview of [explainable AI](https://www.ibm.com/topics/explainable-ai) for the underlying concepts.