AI output inherits skew — gendered coding, stereotyped defaults, exclusionary phrasing — and its fluency makes it read as neutral, so users ship it downstream unchecked.
← Back to patterns
Guidelines for Human-AI Interaction (HAX Toolkit, G1-G18)
Trust & Verification
EmergingBias Flag
The output carries its own skew warnings
AI output inherits skew — gendered coding in a job ad, stereotyped defaults in an image, exclusionary phrasing in customer copy — and the fluency of the prose makes it read as neutral. Bias Flag annotates the output itself: the specific span or depiction, a named skew category, a short evidence note, and a neutral alternative one click away. Review the skew like you review a spelling error — at the moment it can still be fixed.
Metaphor
Spirit Level
Interaction
Span Flag, Skew Category Chip, Neutral Alternative
Use case
Job Ads & HR Copy, Public-Facing Text, Image Generation Defaults
Failure mode
Fluent Skew
Control mechanisms
Flagged Spans · Category Labels · Evidence Note · Replace Suggestion · Per-Flag Accept/Dismiss
Reference examples
Textio · Grammarly inclusive language suggestions · Microsoft Editor inclusiveness checks
Interactive Example
The live side-by-side is the comparison.
This prototype already contains the contrast between chat and the pattern response. It stands as the primary comparison surface for this page.
Skew does not read as skew. It reads as well-written.
A job ad comes back written to brief — ambitious team, fast-paced — as a rockstar engineer who thrives under pressure, a digital native, a work-hard-play-hard culture. Every phrase is fluent, so the only instrument the reviewer has is how it reads, and it reads fine. The careers page ships it, and the coding surfaces where nobody can edit it: in who never applies.
The annotated version renders the same ad with four spans underlined and three named categories in the legend — gendered coding, age coding, ableist idiom. Click one and it opens a category chip, a one-line evidence note, and a single neutral alternative: rockstar becomes an experienced engineer who delivers reliably against deadlines. Replace or dismiss, per flag — the dismissal stays on the record.
The restraint is the design: four flags, not forty. A flag without evidence teaches people to dismiss the ones that have it, and on creative work where the requested voice is the point, the correct count is zero. Textio and Grammarly already do this to human writing; pointing the same instrument at the model's own output is the part still missing.
Skew caught in the draft is an edit. Skew caught in the applicant pool is a statistic.
DOPE Lens
How this pattern improves AI interaction quality.
Observability
Skew is visible as flagged spans in the output, not an abstract disclaimer above it
Directability
Each flag carries a neutral alternative; accept or dismiss per flag, with dismissals logged, not silent
Predictability
Flag categories are a stable, named set — the same kind of coding triggers the same kind of flag
Explainability
Every flag names its category and shows the evidence — which words, which pattern, why it reads as skew
Fit
Where it earns its place.
Use it
when
when
Generated text or imagery ships to audiences where unnoticed skew has real cost — hiring, customer copy, public communication.
Avoid
when
when
Creative work where the requested voice or style is the point — a flag on deliberate register is a false positive.
In the wild
Where this pattern already ships.
The pattern is shipping but still maturing across products. Entries below show how different teams have approached it; expect variation as the space settles.
Screenshot pending
Textio
Flags gendered and exclusionary coding in job posts and performance feedback, span by span, with rewrite suggestions and evidence — the canonical skew-flagging product, applied to human text.
Visit Textio ↗
Screenshot pending
Grammarly · Inclusive language suggestions
Grammarly
Underlines potentially exclusionary phrasing inline and offers neutral alternatives with a short why — accept or dismiss per flag. Flags human writing; flagging the AI's own output is the emerging edge.
Visit Grammarly ↗
Screenshot pending
Microsoft Editor · Inclusiveness checks
Microsoft
Optional inclusiveness category in Word's refinement checks — age, cultural, gender bias — as reviewable spans with alternatives, off by default.
Visit Microsoft ↗
Research Frame
Why this pattern exists, and where it breaks.
Flag the specific span or depiction in the output, name the skew category, show the evidence, and offer a neutral alternative one click away.
A prose answer has no channel for "this part may be skewed" — the confident paragraph asserts its own neutrality.
Over-flagging becomes moral noise that gets dismissed wholesale — flags must be specific, evidenced, and rare enough to mean something.
Common questions
What people ask about this pattern.
When should I use the Bias Flag pattern?
+
Generated text or imagery ships to audiences where unnoticed skew has real cost — hiring, customer copy, public communication.
When should I avoid the Bias Flag pattern?
+
Creative work where the requested voice or style is the point — a flag on deliberate register is a false positive.
What problem does Bias Flag solve?
+
AI output inherits skew — gendered coding, stereotyped defaults, exclusionary phrasing — and its fluency makes it read as neutral, so users ship it downstream unchecked.
Why is chat the wrong fit for this?
+
A prose answer has no channel for "this part may be skewed" — the confident paragraph asserts its own neutrality.
Grounded in Research
Not just opinion.
The papers, guidelines, and usability studies this pattern rests on.
Saleema Amershi et al.
Microsoft Research · 2019 · refreshed 2026-07
The canonical 18-guideline taxonomy for human-AI interaction. Strongest anchors: G9 efficient correction, G11 why-did-it-do-that, G17 global controls. Its three zero-coverage guidelines — G6 bias mitigation, G14 cautious updates, G18 change notification — became this library's Bias Flag and Model Changelog patterns in July 2026. Nearly the whole Agent Workflow category sits outside its 2019 single-assistant frame.
Related Patterns
Curated connections, not just neighbors.
Hand-picked relationships — what this pattern pairs with in practice, what it replaces, what it builds on, and what builds on it.
Often paired with
AI Provenance
AI Provenance marks AI-drafted spans no human has reviewed; Bias Flag marks spans that may be skewed and offers a neutral fix.
Often paired with
Confidence Signals
Confidence Signals mark how sure the AI is of each claim; Bias Flag marks skew in framing, a different failure on the same output.
Often paired with
Refusal Receipt
Refusal Receipt explains the no; Bias Flag annotates the output that does ship.