Look for Unequal Effects and Access Barriers

Work through the explanation, apply it to the example, and complete the quick check before continuing.

Start with affected people

Bias can enter through data, labels, measurement, task design, model behaviour, interface design, and human use. A result may work well on average while failing a smaller group.

Fairness is context-dependent. Asking a model Are you fair? does not test the process or its effects.

Text description of the visual

The figure begins by showing that many users may do fine while a smaller group fails and has no clear challenge route. It then lists five impact questions: who receives the benefit, burden, decision, or error; who may be missing from the evidence; whether performance differs for relevant groups or situations; whether affected people can correct, challenge, or get human help; and who monitors outcomes and changes the process. A workflow accessibility panel lists keyboard access, focus, alt text, captions, reflow, and human help. The example shows that speech-only course support may exclude learners who cannot hear audio, speak commands, or use the interface, so an equivalent text, keyboard, and human-support route is needed. A final note says high-impact decisions require approved fairness methods and specialist review.

Ask five impact questions

1. Who receives the benefit, burden, decision, or error?
2. Who may be missing or poorly represented in the evidence?
3. Does performance differ for relevant groups or situations?
4. Can affected people understand, correct, challenge, or obtain human help?
5. Who monitors outcomes and changes the process?

High-impact decisions need approved fairness methods and specialist review, not an improvised checklist alone.

Accessibility is part of the workflow

An output can be accurate yet inaccessible.

Check the complete process, including:

• Keyboard use and visible focus.
• Clear headings, labels, instructions, and errors.
• Text alternatives for meaningful images.
• Captions and transcripts for media.
• Colour-independent meaning and sufficient contrast.
• Reflow and readable content at small widths.
• Accessible authentication and human support routes.

Automated checks help, but they do not cover every barrier or every person. Combine technical testing with human evaluation.

Example

A chatbot provides course support only through speech. Accurate answers do not make the service accessible to learners who cannot hear the audio, speak commands, or use the interface.

Provide equivalent text, keyboard operation, and a practical human route. Test with people who use relevant access methods.

Remember

• Average performance can hide unequal harm.
• Consider who is missing, burdened, or unable to challenge.
• Accessibility covers the complete process, not only the final text.
• Use technical and human testing.

Next, we will identify rights questions that need evidence and escalation.