Audit persona prompts for stereotype drift
Compare logged persona-prompt runs against coded community evidence, then block unsupported demographic assumptions, pain points, and solutions before personas circulate.
Why
Why it earns its place
Prompt wording and stochastic variation can turn group patterns into individual certainty. Traceability, repeated runs, and community validation expose that drift before it shapes service decisions.
How
Run it
Establish human reference points
Read and code the raw community material before using an LLM. Record the persona purpose, observed patterns, uncertainty, contradictions, and evidence IDs.
Log controlled prompt runs
Save the model, date, settings, data version, prompt, schema, and outputs. Compare a broad instruction with an evidence-constrained schema and repeat each run.
Audit every persona claim
Trace needs, pain points, and solutions to evidence IDs. Flag demographic prediction, group-to-individual overgeneralisation, misplaced themes, unsupported filler, and differences across runs.
Validate before circulation
Revise personas through human judgement and community review. Mark direct, inferred, and speculative claims; remove unsupported material and record who approved the final artefacts.
Ready to paste
Prompt
Act as a synthesis assistant, not an interpreter or community representative.
Persona purpose: [purpose]
Coded community evidence with stable IDs: [paste evidence]
Fixed attributes: [schema]
Number of provisional candidates: [number]
Use only the supplied evidence. For every attribute, need, pain point, and solution direction, return its supporting IDs and label it Direct, Inferred, or Speculative. Preserve variation, contradictions, and uncertainty.
Treat demographic fields only as supplied descriptors. Do not predict unfair-treatment experiences, abilities, motives, needs, pain points, or solutions from age, gender, race, nationality, disability, role, or cultural identity unless explicit evidence IDs support that exact link. Write “not evidenced” rather than filling gaps.
Return provisional candidates, a claim-to-evidence table, unsupported additions, possible stereotype/proxy risks, and claims requiring community validation. Do not declare the personas representative or final.- Which claim changed across runs or relies on a demographic proxy?
- Which community members must review this before circulation?
The AIdea layer
The designer’s call
- Human core
- Researchers and communities define persona purpose, interpret evidence, judge harms, validate claims, and approve organisational artefacts.
- The trade
- Fast formatting and first drafts ↔ stereotype drift and false certainty. Log prompts, repeat runs, trace claims, and reinvest saved time in validation.
- Skip AI when
- Do not generate personas when raw evidence is unavailable, communities cannot be protected, or unsupported inferences could drive consequential decisions.
Check before use
Watch out
Persona data can expose sensitive identities and unfair-treatment experiences. Use approved, minimised data and access controls. Compare repeated runs, inspect proxy assumptions, trace every claim, and block circulation until accountable researchers and, where possible, represented communities validate the artefacts.
Source of truth
Evidence
A festival case with at least 119 questionnaire responses documented untraceable filler, demographic overgeneralisation, misplaced pain points, and run-to-run inconsistency; it did not evaluate outcomes.
Priscilla Chueng-Nainby, Nabilah Gelshirani and John Lee · Co-Designing Personas with AI: Community-anchored synthesis in the age of generative models