AI Ethics and Bias Audit Template: Evaluate AI Systems Before Production Deployment
Deploying an AI system without a structured ethics and bias review is not just a reputational risk — it is increasingly a legal one. Regulations including the EU AI Act, proposed US federal AI governance frameworks, and sector-specific rules under EEOC, FCRA, and ECOA impose substantive obligations on organizations that deploy AI in consequential decisions. Beyond compliance, AI systems that produce biased or opaque outputs erode customer trust and employee confidence in ways that take years to rebuild. This template gives you a systematic process to evaluate any AI system before it touches production. Use it before go-live on new deployments, after significant model updates, and on an annual cadence for systems already in production.
How to Use This Template
This template is designed for the person responsible for AI deployment — a data science lead, product manager, compliance officer, or AI consultant — working with input from the business stakeholders whose workflows the AI affects. You will need access to the model's documentation or the vendor's technical disclosures, a representative sample of the model's inputs and outputs, and someone who can run evaluation queries against your data.
Complete one audit per AI system or per major AI-powered feature. For complex systems with multiple AI components (e.g., a recommendation engine that also contains a fraud detection layer), audit each component separately.
Scoring: each section produces a score from 1–5. A system ready for production should score 4 or 5 in every section. A score of 3 in any section requires a documented remediation plan with a deadline before launch. A score of 1 or 2 in any section is a launch blocker.
Document every finding and keep the audit record for at least three years. Regulatory inquiries frequently request evidence of pre-deployment review.
Section 1: Use Case Risk Classification
Before testing for bias, classify the risk level of the use case. Higher-risk applications require more rigorous review and stronger safeguards.
High-risk applications (require full audit + external review):
- Hiring, promotion, compensation, or workforce reduction decisions
- Credit, lending, or insurance underwriting
- Healthcare diagnosis, treatment recommendation, or clinical triage
- Law enforcement or criminal justice
- Access to housing, education, or government benefits
- Content moderation at scale
Medium-risk applications (require full audit):
- Customer segmentation that affects pricing or service levels
- Lead scoring that determines who receives sales outreach
- Employee performance evaluation or productivity monitoring
- Content recommendation that can significantly affect user behavior
Lower-risk applications (require abbreviated audit, Sections 1–4 minimum):
- Internal productivity tools (drafting, summarizing, searching)
- Operational automations with human review steps
- Analytics and reporting without automated decision output
Record the risk classification and the reasoning. If two reasonable people would disagree about the classification, escalate to legal/compliance before proceeding.
Section 2: Data Provenance and Training Audit
The outputs of any AI system reflect the data it was trained on. Evaluating that data is the first step in evaluating the system.
2a. Training data documentation: does the vendor or internal team provide documentation of what data the model was trained on? If not, this is a transparency failure — score 2 or below.
2b. Historical bias: was the training data drawn from historical records that reflect past discriminatory practices? (Example: a resume screening model trained on historical hires at a company that historically underhired women will encode that pattern.) Document any known or suspected historical biases in the training set.
2c. Representation: are all relevant demographic groups present in the training data in sufficient volume? Models perform worse on underrepresented groups — not because of malice, but because they have seen fewer examples. Ask the vendor: what is the demographic breakdown of the training data, and was it balanced or weighted?
2d. Label quality: for supervised models, how were training labels generated? Labels created by human annotators can encode annotator bias (e.g., annotators who associate certain names or dialects with negative sentiment). If the vendor used crowdsourced labeling, ask about inter-annotator agreement and bias audits of the labeling pool.
2e. Data recency: how old is the training data? Models trained on data that is more than 2–3 years old may not reflect current norms, laws, or population demographics.
Section 3: Bias Testing Protocol
Testing for bias requires running the AI system against a structured evaluation set and measuring performance across demographic groups.
3a. Identify the protected characteristics relevant to your use case and jurisdiction. In the US, these typically include race, color, national origin, sex, religion, age (40+), disability, and pregnancy under federal law. State laws add additional categories (sexual orientation, gender identity, source of income, and others).
3b. Assemble or request a bias evaluation dataset that includes known demographic attributes for each record. If your production data does not contain demographic labels (as is common), use proxy variables where legally permissible or request the vendor's own bias evaluation results.
3c. Measure the following metrics across each demographic group:
- Selection rate: what percentage of each group receives a positive outcome (hired, approved, recommended)?
- False positive rate: when the model predicts a positive outcome that is incorrect, does it fail more often on certain groups?
- False negative rate: when the model misses a positive outcome, does it fail more often on certain groups?
3d. Apply the 4/5 Rule (also called adverse impact ratio): if the selection rate for any group is less than 80% of the highest-scoring group, adverse impact is indicated. This is the standard used by US regulatory agencies in employment contexts. Document the result explicitly.
3e. For language models: test outputs for differential quality and tone by group. Provide equivalent prompts substituting names, locations, and terms associated with different demographic groups and compare outputs for consistency and quality.
3f. Document all test results, including null results (no disparity found). Null results are part of the audit record.
Section 4: Transparency and Explainability
Users and affected individuals have a right to understand how AI systems reach decisions that affect them. This section evaluates whether your system can provide that transparency.
4a. Decision explainability: can the system produce a human-readable explanation for individual outputs? This is particularly important for consequential decisions (rejection, denial, adverse action). 'The model predicted X' is not an explanation. 'The primary factors were Y and Z, weighted by A' is an explanation.
4b. Disclosure to affected parties: if the AI system makes or informs a decision that affects a customer or employee, are they told an AI was involved? Draft the disclosure language and confirm it meets applicable legal requirements (FCRA adverse action notices, for example, require specific language about automated decision-making in consumer credit contexts).
4c. Human override: is there a documented process for an affected person to request human review of an AI-assisted decision? For high-risk use cases, this is increasingly a legal requirement rather than a best practice.
4d. Internal documentation: is the model's purpose, architecture, training data, performance metrics, and known limitations documented internally? Could a new employee who inherits the system understand what it does and why it behaves the way it does?
4e. Change log: is there a record of when the model was updated, what changed, and what bias re-testing was conducted after the update?
Section 5: Accountability and Governance
Accountability requires that a named person or team is responsible for the AI system's behavior and has the authority to act on problems.
5a. Named owner: who is the accountable owner of this AI system? This person must have both the technical authority to make changes and the organizational authority to halt or modify the system if a problem is discovered. 'The vendor' is not an acceptable answer — a vendor cannot be held accountable under your organization's compliance structure.
5b. Escalation path: if the AI system produces a harmful or biased output in production, who is notified? What is the timeline for response? What is the threshold that triggers a system pause?
5c. Review cadence: how often is the system's performance and bias profile reviewed? Quarterly is the minimum for high-risk applications; annual is acceptable for lower-risk internal tools.
5d. Vendor accountability: if you are using a third-party AI system, what contractual obligations does the vendor have regarding bias testing, model updates, and incident response? Who is liable if the system produces discriminatory outputs? Review the vendor agreement with legal before deployment.
5e. Audit trail: does the system log inputs, outputs, and the data used to produce each output? This log is essential for investigating complaints, responding to regulatory inquiries, and diagnosing problems after the fact. Define the retention period and access controls for AI logs.
Section 6: Ongoing Monitoring Plan
A one-time pre-launch audit is not sufficient for systems in production. This section documents the ongoing monitoring commitments.
6a. Performance drift monitoring: AI models degrade over time as the world changes and the production data distribution diverges from the training distribution. Define the performance metrics you will track in production, the baseline at launch, and the threshold that triggers a re-evaluation.
6b. Demographic outcome monitoring: for consequential systems, track selection rates by demographic group in production on a monthly basis. A significant shift from baseline (more than 5 percentage points) triggers a bias re-audit.
6c. Feedback collection: how will affected users report problems with the AI's outputs? Is there a visible feedback mechanism? Is feedback reviewed by a human on a regular cadence?
6d. Trigger events for re-audit: document the events that require a full re-audit outside the regular cadence: significant model update, expansion to a new use case, new regulatory requirement, a complaint that alleges discrimination or adverse impact, or an internal finding during routine monitoring.
6e. Annual full re-audit: schedule the next full audit before launching. Governance gaps are almost always a result of not scheduling the next review before the urgency of the current one fades.
Section 7: High-Risk Use Case Addendum
For systems classified as high-risk in Section 1, complete this additional section before proceeding.
7a. External audit: for hiring, lending, or healthcare AI, obtain an independent bias audit from a third-party firm with domain expertise. Self-audits are insufficient for high-stakes regulatory exposure.
7b. Legal review: have outside counsel with AI and employment/lending/healthcare law expertise review the system and its deployment before launch. Specifically ask about: applicable federal and state law, disclosure obligations, adverse action notice requirements, and the regulatory landscape for AI governance in your industry.
7c. Employee or customer communication: how will you communicate to affected parties that AI is being used? What do they need to know to give informed consent where required? What options do they have to opt out?
7d. Union or works council notification: if you employ unionized workers and the AI system affects working conditions, check whether your collective bargaining agreement or applicable labor law requires notification or negotiation before deployment.
7e. Executive sign-off: high-risk AI deployments should require a named executive to review and sign off on the audit results before launch. This creates accountability and ensures the organization's leadership has acknowledged the risk profile.
Section 8: Audit Scorecard and Launch Decision
Rate each section 1–5 based on completeness and quality of responses:
| Section | Score (1–5) | Notes / Open Items |
|---------|-------------|--------------------|
| 1. Risk Classification | | |
| 2. Data Provenance | | |
| 3. Bias Testing | | |
| 4. Transparency | | |
| 5. Accountability | | |
| 6. Monitoring Plan | | |
| 7. High-Risk Addendum (if applicable) | | |
Launch decision:
- All sections 4–5: approved for launch. File the audit record.
- Any section at 3: launch conditional on remediation plan with named owner and date. Do not launch until plan is in place.
- Any section at 1–2: launch blocked. Resolve the gap and repeat the affected sections.
Sign-off required from: [AI System Owner], [Legal/Compliance Representative], [Executive Sponsor]
Audit date: ___________ Next scheduled audit: ___________
This audit does not constitute legal advice. Consult qualified legal counsel before deploying AI in high-risk contexts.
Frequently Asked Questions
Frequently Asked Questions
Both. When you deploy any AI system — purchased, open-source, or custom-built — your organization is responsible for its outputs in your context. A vendor's own internal bias audit does not transfer liability to them. You still need to evaluate the system against your use case, your population, and the regulatory requirements that apply to your business. Off-the-shelf tools often come with less documentation about training data and evaluation methodology, which makes independent testing more important, not less.
The 4/5 Rule (adverse impact ratio) compares the selection rate of each demographic group to the rate of the highest-scoring group. If Group A has a 50% selection rate and Group B has a 60% selection rate, the ratio is 50/60 = 83% — above the 80% threshold, so no adverse impact is indicated. If Group A had a 40% rate, the ratio would be 67% — below 80%, indicating potential adverse impact. US regulatory agencies including the EEOC use this standard in employment contexts. It is a useful screen but not the only bias test — consult a data scientist or HR compliance expert for high-risk applications.
Several approaches are available. For language-based inputs (names, addresses, text), use established proxy testing: substitute names and identifiers associated with different demographic groups and compare outputs. For tabular data, Bayesian Improved Surname Geocoding (BISG) can infer race and ethnicity probability from surname and geography — it is imprecise at the individual level but useful for population-level analysis. Your AI vendor may also have conducted demographic bias evaluations they can share. In some regulated industries (consumer credit, employment), regulators provide specific guidance on acceptable proxies for demographic testing.
For high-risk applications (hiring, lending, healthcare), conduct a full audit annually and after any significant model update. For medium-risk applications, annually is sufficient if ongoing monitoring shows stable performance. For lower-risk internal tools, a full re-audit every 2 years is reasonable. The most important thing is not the cadence but the commitment: schedule the next audit before you file the current one. Governance programs that rely on 'we'll do it when we need to' consistently do not do it.