/cast

Persona Bias Audit Reference

Purpose: Detect and remediate bias in persona sets. Cover representation matrix, intersectionality coverage, stereotyping detection, the Inclusive Persona Checklist, name/photo bias, ableism, ageism, and locale bias. Personas with embedded bias propagate downstream into Echo walkthroughs, Spark feature priorities, and Voice copy.

Scope Boundary

• cast bias-audit: Persona representation + ethical audit (this document).
• cast archetype (elsewhere): Brand / JTBD archetype tagging.
• Cloak (elsewhere): Privacy / consent / PII handling.
• Canon[regulatory] (elsewhere): Regulated context (e.g., insurance non-discrimination).
• Echo[demand] (elsewhere): Synthetic user voice (separate concern, but shares ethics surface).

Why Bias Audits

Personas are reasoning shortcuts — they shape every downstream decision. Biased personas produce:

• Excluded users (features that don't fit)
• Stereotyped UX ("women want pink", "elders need huge fonts")
• Discriminatory pricing or eligibility
• Hostile design (assuming all users speak English, have stable WiFi, hold a credit card)
• Reinforcement of marketing tropes

The Five Bias Lenses

Representation Matrix

                Age
        18-24  25-34  35-44  45-54  55-64  65+
Gender ┌──────┬──────┬──────┬──────┬──────┬─────┐
F      │  1   │  2   │  1   │      │      │     │
M      │      │  3   │  1   │      │  1   │     │
NB     │      │      │      │      │      │     │
       └──────┴──────┴──────┴──────┴──────┴─────┘

Generate per dimension: gender × age, ability × locale, income × education, etc. Empty cells = potentially unserved.

Compare against:

• Customer base (who actually uses the product)
• Total addressable market (who could)
• Census / domain norms (sanity check)

Flag dimensions where persona set ≠ population. "We have 5 personas, all male, all 25-34" is a red flag even for a B2B dev tool.

Intersectionality Grid

A persona at one identity (woman) is not the same as at the intersection (woman + Black + lower-income + rural + non-native English). Crenshaw 1989: discrimination is multiplicative, not additive.

Empty intersections = blind spots. Personas at unexpected intersections often reveal the strongest UX gaps.

Stereotype Detection

Run an attribute-correlation audit: do attributes cluster by tradition rather than data?

Smell tests:

• All female personas are caregivers / nurturing
• All male personas are technical / decisive
• Older personas described as "tech-averse" without evidence
• Lower-income personas described as "price-sensitive" only
• Disabled personas defined by their disability
• Non-Western personas described in exotic / othering tone
• LGBTQ+ identity treated as personality

Per attribute, ask: Is this trait evidenced in data, or imported from cliché?

Inclusive Persona Checklist

Adapted from Microsoft Inclusive Design Toolkit + W3C Accessible Personas guidance:

[ ] Persona attributes have provenance citations (interview, survey, log)
[ ] Demographic doesn't predict role (e.g., not all "young" = "tech-savvy")
[ ] Photo / name doesn't reinforce stereotypes (or photo is omitted)
[ ] Disability represented in ≥20% of persona set
[ ] At least 1 persona has a temporary or situational impairment
[ ] Income range spans free-to-paid users
[ ] Locale / language diversity beyond default
[ ] Naming reflects intended market diversity
[ ] Goals / pains in user's own words where possible
[ ] No persona reduced to a single identity attribute
[ ] LGBTQ+ identity (if relevant) treated as context, not center
[ ] Ableism, ageism, classism flags clear (zero-tolerance items)
[ ] Test of Time check: would this persona embarrass us in 5 years?
[ ] Stress-case persona present (denied, abandoned, hostile context)
[ ] Designer reflexivity statement: who built this and whose blind spots?

Required: ≥ 13/15 pass. Disability + stress-case items are mandatory (block on fail).

Name + Photo Bias

Names and photos encode identity quickly — and lazily.

Anti-pattern    →  Fix
"Jessica, 24, F"           →  Real-feeling name + clear cultural origin
Stock photo of smiling woman →  Diverse photo set OR omit photo entirely
Anglo-only name pool        →  Match target market diversity
"Tech-savvy millennial"     →  Specific behaviors, not generation labels

If photo can't be diverse, omit. A name + role + behaviors carries more weight than a photo.

Ableism, Ageism, Classism Pitfalls

Stress-Case Personas

Eric Meyer + Sara Wachter-Boettcher (Design for Real Life): every persona set must include people in crisis or hostile contexts.

• Recently divorced / bereaved (sensitive copy)
• Domestic abuse victim (privacy, escape buttons)
• Refugee / asylum seeker (trust, language barrier)
• Recently disabled (re-learning the interface)
• Identity theft victim (low trust)
• Burned out professional (low cognitive load)

These reveal where the product fails the most vulnerable.

Workflow

COLLECT     →  load all personas in registry
            →  list every demographic + behavioral attribute

REPRESENT   →  build representation matrix per pair (gender×age, ability×locale, ...)
            →  flag empty cells vs target market
            →  sanity-check against census / domain norms

INTERSECT   →  intersectionality grid: identify unmet intersections
            →  add ≥1 persona per critical empty intersection

STEREOTYPE  →  attribute-correlation audit
            →  flag clichéd pairings; require evidence per pair

POWER       →  who designed the persona set?
            →  reflexivity statement: known blind spots
            →  external review by underrepresented voice if possible

CHECKLIST   →  apply Inclusive Persona Checklist (13/15 minimum)
            →  disability + stress-case mandatory

NAME-PHOTO  →  name pool diversity
            →  photo diversity or omission

REMEDIATE   →  add missing personas
            →  rewrite stereotyped attributes with evidence
            →  flag personas that fail checklist with severity

REPORT      →  bias audit report with severity tiers
            →  recommended additions / rewrites

HANDOFF     →  back to cast `generate` for missing personas
            →  Cloak: privacy review of new attributes
            →  Canon[regulatory]: regulated-context check
            →  Voice / Echo / Spark: re-run downstream with debiased set

Output Template

## Bias Audit: [Persona Set]

### Representation Matrices
**Gender × Age**
[matrix]
- Empty cells: [list]
- vs market base: [delta]
- Severity: [low / med / high]

**Ability × Locale**
[matrix]
- ...

### Intersectionality Grid
[grid showing personas at intersections]
- Critical missing intersections: [list]

### Stereotype Audit
| Attribute pair | Personas | Evidence? | Verdict |
|---------------|----------|-----------|---------|
| Female + caregiving | P-002, P-003 | weak (assumption) | REWRITE |
| Older + tech-averse | P-005 | none | DROP |
| Premium + sophisticated | P-001 | none | DROP |

### Inclusive Persona Checklist Score
- [N/15] passed
- Mandatory items (disability, stress-case): [PASS / FAIL]
- Severity if FAIL: [block]

### Name + Photo
- Name diversity: [pool stats]
- Photo: [diverse / omitted / homogeneous → recommend change]

### Reflexivity
- Persona set authored by: [team profile]
- Known blind spots: [list]
- External review: [yes / no, by whom]

### Stress-Case Coverage
- [list of stress-case personas; recommend add if missing]

### Severity-Tiered Findings
**BLOCK** (must fix before downstream use):
- [item]

**WARN**:
- [item]

**INFO**:
- [item]

### Recommended Additions
- New persona: [profile builder] — fills [intersection]

### Recommended Rewrites
- P-002: drop "nurturing by default" → replace with evidence-cited behavior

### Handoffs
- cast `generate`: produce additions
- Cloak: privacy review
- Canon[regulatory]: regulated check (if applicable)
- Echo / Spark / Voice: re-run downstream after fix

Anti-Patterns

Deliverable Contract

When bias-audit completes, emit:

• Representation matrices with empty-cell flags.
• Intersectionality grid with critical missing intersections.
• Stereotype audit table with verdicts.
• Inclusive Persona Checklist (15-item) score with mandatory-pass status.
• Name + photo diversity assessment.
• Reflexivity statement + external review status.
• Stress-case coverage check.
• Severity-tiered findings (BLOCK / WARN / INFO).
• Remediation list (additions + rewrites).
• Handoffs: cast generate, Cloak, Canon[regulatory], downstream re-run.

Regulatory & Standards Context (as of 2026-05)

Persona bias work increasingly intersects with formal AI-governance obligations. Cast bias audits should reference applicable frameworks rather than ad-hoc rules:

For AI-generated personas, Cast bias audits should at minimum align with NIST AI 600-1 actions and ISO/IEC 23894 risk categories.

References

• Kimberlé Crenshaw, "Demarginalizing the Intersection of Race and Sex" (1989)
• Design for Real Life — Eric Meyer + Sara Wachter-Boettcher (stress cases)
• Mismatch — Kat Holmes (inclusive design at Microsoft)
• Technically Wrong — Sara Wachter-Boettcher (bias in tech)
• Algorithms of Oppression — Safiya Noble
• Microsoft Inclusive Design Toolkit — inclusive.microsoft.design (refreshed 2024-2025 around 10-year anniversary; Persona Spectrum concept; Fluent 2 design system)
• W3C Accessible Personas — w3.org/WAI/planning/personas/
• Lean UX — Gothelf + Seiden (proto-persona ethics)
• Just Enough Research — Erika Hall (interview rigor)
• POUR principles — WCAG (Perceivable, Operable, Understandable, Robust)
• ISO 9241-220 — Human-centered design processes
• ISO/IEC 42001:2023 — AI Management System (iso.org/standard/42001)
• ISO/IEC 23894:2023 — Guidance on risk management for AI (iso.org/standard/77304.html)
• NIST AI RMF 1.0 + AI 600-1 Generative AI Profile (2024-07; doi.org/10.6028/NIST.AI.600-1)
• EU AI Act — Regulation (EU) 2024/1689 (artificialintelligenceact.eu/implementation-timeline/)
• IEEE 7000-2021 — Model Process for Addressing Ethical Concerns During System Design
• C2PA Content Credentials Technical Specification 2.2 (spec.c2pa.org/specifications/specifications/2.2/)
• Japan APPI Amendment Bill 2026 — ppc.go.jp (Personal Information Protection Commission)
• Geert Hofstede — Cultures and Organizations (locale dimensions)
• Sorting Things Out — Bowker + Star (categorization politics)
• Ruha Benjamin, Race After Technology
• Joy Buolamwini, "Gender Shades" (algorithmic bias parallel)

Core Contract Research Basis (SKILL.md excerpt)

• Prioritize behavioral data over demographics. Personas should be built around user journeys and behavioral patterns, not demographic profiles. Match persona fidelity to team size and research capacity: large organizations benefit from statistical personas (quantitative + qualitative); most teams should use qualitative personas; small teams with limited research capacity can use lightweight personas. Source: nngroup.com/articles/persona-types/.

• Recognize that GenAI does not merely reproduce traditional persona biases — it makes them more convincing and harder to detect (evolutionary amplification). Apply bias audits more rigorously for AI-assisted personas than for manually created ones. A CHI 2026 scoping review of 81 articles (2022–2025) found that 45% of GenAI persona studies lack evaluation and 86% use only GPT models, creating circularity risk when the same model both generates and evaluates personas. Source: dl.acm.org/doi/10.1145/3772318.3790608.

• Include persona refresh anchors in multi-turn delivery packets. CHI 2026 research (N=3,473 conversations) shows LLM self-reported persona intensity remains stable across 18-turn interactions, but observer ratings reveal a gradual decline for moderate and high-intensity personas during extended conversations. DISTRIBUTE packets for multi-turn consuming agents (e.g., Echo walkthroughs) must specify recommended refresh intervals. Source: dl.acm.org/doi/10.1145/3772363.3799334.

• Flag racial and demographic identity representation risk in AI-generated personas. A 2025 ethical audit (arXiv:2505.07850) of personas generated by multiple LLMs found LLMs disproportionately foreground racial markers, overproduce culturally coded language, and construct personas that are syntactically elaborate yet narratively reductive — producing stereotyping, exoticism, erasure, and benevolent bias. Source: arxiv.org/abs/2505.07850.