Help us improve
Share bugs, ideas, or general feedback.
From skills-for-humanity
Audits algorithms, models, ranking systems, and automated decisions for discriminatory patterns and unfair outcomes. Use before deploying any system that makes decisions about people.
npx claudepluginhub human-avatar/skills-for-humanityHow this skill is triggered — by the user, by Claude, or both
Slash command
/skills-for-humanity:s4h-ethics-bias-checkThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
Algorithms that treat everyone the same can still discriminate. A ranking that optimises for engagement may systematically deprioritise certain groups. A model trained on historical data may encode historical injustice. A feature that works well on average may fail badly for users who aren't the implicit default.
Measures ML model performance across demographic groups to detect discriminatory outcomes. Required for regulatory compliance (EU AI Act, CFPB, EEOC) and ethical AI deployment.
Conducts a structured ethical review of AI/ML features, models, or products covering fairness, transparency, privacy, safety, accountability, and societal impact with risk scoring and mitigations.
Designs review workflows, checklists, and processes to detect and mitigate bias in AI outputs, including types of bias, detection methods, and mitigation strategies.
Share bugs, ideas, or general feedback.
Algorithms that treat everyone the same can still discriminate. A ranking that optimises for engagement may systematically deprioritise certain groups. A model trained on historical data may encode historical injustice. A feature that works well on average may fail badly for users who aren't the implicit default.
This check surfaces those patterns before they ship.
Step 1: Define the system What is the algorithm, model, or automated decision? What is its input? What is its output? Who does it make decisions about? What happens to people based on its output?
Framing check: Confirm the specific system before continuing. State what you've identified — the actual system being audited, its outputs, and the population it affects — in one sentence, then use AskUserQuestion:
Step 2: Identify the implicit default Every system has a default user in mind — often implicitly. Ask:
The implicit default is often the demographic that experiences least friction. Others bear the cost of that assumption.
Step 3: Check for direct bias Does the system use protected characteristics (age, gender, race, disability, location as proxy for race, etc.) as features, or correlates that map closely to them? Does it produce different outcomes for different demographic groups? Is that difference justified (e.g. a medical dosage model that accounts for body weight) or unjustified (e.g. a loan model that penalises postcodes that correlate with race)?
Step 4: Check for proxy bias Protected characteristics don't need to be explicit to create discriminatory outcomes. Audit the features used:
Step 5: Check the feedback loop Many deployed systems create the conditions that confirm their own predictions. A recommendation system that shows users content they're predicted to engage with can entrench filter bubbles. A fraud model that flags more accounts from a certain demographic leads to more investigations of that demographic, which produces more evidence that "justifies" the bias. Ask: does this system create or amplify the patterns it's trying to predict?
Step 6: Assess the harm distribution When the system makes errors, who bears the cost?
Before proceeding, use the AskUserQuestion tool. State your interpretation of the situation in 1–2 sentences — what is being analyzed and what the core question is — then ask:
Proceed based on their selection. If the user reframes, incorporate the correction before running any analysis.
System Being Audited: [What it does, inputs, outputs, who it affects]
Implicit Default [Who the system is optimised for; who may be disadvantaged by that assumption]
Bias Findings
| Dimension | Finding | Severity |
|---|---|---|
| Direct bias (protected characteristics) | [finding] | 🔴 / 🟡 / 🟢 |
| Proxy bias (correlated features) | [finding] | 🔴 / 🟡 / 🟢 |
| Training data issues | [finding] | 🔴 / 🟡 / 🟢 |
| Feedback loop risk | [finding] | 🔴 / 🟡 / 🟢 |
| Error cost distribution | [finding] | 🔴 / 🟡 / 🟢 |
Most Significant Concern [One specific, concrete finding that warrants most attention]
Recommended Actions
A clean bias check is not a guarantee of fairness — it is evidence of serious effort. Bias is often subtle and emerges at scale. Where this check surfaces concerns, treat them as decisions to be made consciously, not problems to be explained away.
For systems with high-stakes outputs (credit, hiring, healthcare, content moderation), this check is a minimum. Consider ongoing monitoring post-deployment, not just a pre-ship audit.
After delivering this output, use AskUserQuestion to offer the next move:
/s4h-ethics-check — Run a full ethical assessment of the bias-affected reasoning/s4h-logic-fixer — Correct bias-induced logic errors/s4h-decision-criteria-weighting — Re-weight decision criteria after removing bias