Decision outputs

Some AI output is a decision or a recommendation that affects the user, such as a score, a classification, a ranking, or an approval. This shows up in hiring, lending, benefits, content moderation, and many public services. These uses carry two distinct accessibility concerns at once, and both matter.

Key points

An AI decision or recommendation carries two accessibility concerns, a usable interface and a fair outcome.
The interface must present the score, reasons, and next steps in text and structure, not color alone.
The decision itself can discriminate against disabled people, since disability data is often missing or mishandled.
Explain the decision in a way that reaches everyone, including screen reader users.
Provide a real, accessible path to contest or appeal, operated by a human.

Two concerns, not one

The ordinary one. The interface that presents the decision has to be perceivable and operable, so the score, the reasons, and the next steps cannot be conveyed by color alone, and any controls have to work with a keyboard and a screen reader.
Specific to decisions. The decision itself can discriminate against disabled people, regardless of how accessible the screen is. An accessible page that shows an unfair result is still a serious harm.

Bias against disabled people

Decision systems learn from data, and disability is often missing or mishandled in that data. Researchers at Harvard have argued that disability is so varied it gets treated as outlier data and left out, and Penn State found that language models score disability-related terms more negatively. A model that screens job applicants or moderates content can carry these patterns into decisions, penalizing a disabled person for a gap in their history, an atypical phrasing, or an assistive-technology fingerprint. Legal scholars describe this as a form of algorithmic disability discrimination. See Bias and representation.

Explain the decision accessibly

People affected by a decision often have a right to understand it, and that explanation has to reach everyone. The US explainability principles say an explanation should be given, should be meaningful to the person, should reflect how the system actually reached the result, and should respect the limits of what it can explain. An explanation that a screen reader user cannot perceive is not meaningful to them. See WCAG Understandable.

Provide a real path to contest

A decision that affects someone needs a way to question or appeal it, operated by a human. Under the EU AI Act, many such uses are high-risk and require human oversight, as covered in EU AI Act. Make the appeal route as accessible as the decision itself, so the person who most needs to contest an unfair result is not blocked by an inaccessible form.

Present scores and criteria clearly

Where you show a score, a ranking, or the criteria behind a decision, present it as structured text that assistive technology can read, not as a color-coded chart with no text equivalent. The reasons, and the actions a person can take next, are the most important parts to get right.

Audience: Product Manager

Work out early whether a decision feature is high-risk under the EU AI Act, because that brings duties around human oversight, transparency, and accessibility. Beyond compliance, commission a check for disability bias in the decisions, not only an accessibility check of the screen, since a perfectly accessible page can still deliver a discriminatory result. Make sure the appeal path is staffed by a person and is itself accessible.

Audience: Accessibility Specialist

Audit two things here.

The conformance of the decision interface, including how scores, reasons, and appeal controls are exposed.
Whether disabled users are disadvantaged by the decision itself, which is easy to forget and calls for outcome testing across disability-related cases rather than a markup review.

Pair findings with the explainability and oversight expectations from law and frameworks.

Audience: Self-advocate

Being judged by a system that never accounted for people like you is its own harm, separate from whether the screen is readable. The decisions that respect people explain themselves in plain, readable terms, show the reasons, and give a real way to push back that does not dead-end in an inaccessible form. A fast no with no explanation and no appeal is the pattern to call out.

Audience: Engineer

Expose the decision, its reasons, and the appeal controls as structured, named, keyboard-operable content, not as a color-coded widget. If the system can produce an explanation or the main factors, render them as real text associated with the result. Make the contest or appeal action a first-class, focusable control, and never gate it behind a step that only works with a mouse.

Carry the result, the reasons, and the appeal in text and structure, not in color alone.

<h2>Application result</h2>
<p>Your application was not approved.</p>

<h3>Main factors in this decision</h3>
<ul>
  <li>Reported income was below the required threshold</li>
  <li>Employment history was incomplete</li>
</ul>

<a href="/appeal/12345">Request a human review of this decision</a>