Spot bias and missing perspectives

What you’ll be able to do

• Identify who is represented or overlooked in an output.
• Distinguish consistent treatment from assumptions about a group.
• Revise a recommendation using explicit, relevant criteria.

Ask who the default person is

An output can be factually correct in parts and still assume that everyone has the same schedule, language skills, equipment, mobility, or internet access. Those assumptions affect who can use the result.

For a free AI workshop, “Bring your laptop and use your paid account” may exclude people the program intends to serve. The issue is not solved by adding the word “inclusive.” The plan needs a workable alternative, such as a shared-device exercise or an activity that does not require a paid service.

A worked example: workshop access

A draft plan says: “All participants will complete a fast typing exercise on their personal laptops and upload a document from their workplace.”

Review the assumptions:

• Everyone owns a laptop and can bring it.
• Everyone can type quickly enough to participate in the same way.
• Everyone has a workplace document and permission to upload it.

A better plan uses an instructor-supplied fictional document, offers paired or individual participation, permits spoken discussion, and explains the task in plain language. These are design choices tied to the exercise’s learning goal.

Use a four-question review

Who benefits? Name the intended audience and consider who might be left out.

Which assumptions appear? Look for unsupported generalizations about people or situations.

What criteria matter? Use criteria relevant to the task, such as whether an activity teaches the stated skill.

What alternative is workable? Replace an inaccessible requirement with an option that still serves the learning objective.

Try a perspective-check prompt

Review this fictional workshop activity for barriers to participation.
Separate assumptions from confirmed facts. Identify barriers related to equipment, cost, language, and modes of participation.
Suggest alternatives that preserve the learning goal: checking an AI summary against a source.
Do not infer anyone’s abilities or preferences from their identity.

Activity: Everyone must bring a laptop, use a paid AI account, type quickly, and upload a workplace document.

Your practice: redesign the exercise

1. Identify three assumptions in the activity above.
2. Explain the practical barrier created by each assumption.
3. Design an alternative that does not require paid access or private documents.
4. Check whether the alternative still teaches source verification.
5. Ask a peer to review the instructions for clarity. If no peer is available, describe the feedback you would want before running the activity.

Your finished artifact: A revised activity and a brief explanation connecting each change to a participation barrier and the learning goal.

Use comparisons carefully

Testing the same task with slightly different descriptions can reveal inconsistent outputs. Keep the relevant criteria constant and change one feature at a time. A small test can identify a problem worth investigating, but it does not establish that a system is fair across all people and contexts.

Avoid asking an AI system to infer a person’s ability, character, or needs from a demographic label. For consequential decisions about people, a classroom exercise is not sufficient validation of a decision process.

A common mistake

Treating fairness as a final adjective applied to a finished output. It is more useful to examine the criteria, inputs, available alternatives, and effects on the people involved.

Reflection

What did your original plan treat as universal that was actually just an assumption?

Go deeper

NIST’s AI Risk Management Framework discusses risks in the context of affected people and real-world use.