1. Ethical Guidelines¶
Purpose
Frameworks to ensure that AI systems respect human values and do not cause unintended harm.
1. Purpose¶
Ensure that AI systems are developed and used in a way that respects human values and causes no unintended harm.
2. Ethical Principles¶
Human Oversight and Control¶
AI must not undermine human autonomy. Users must be able to understand how the system works and, where necessary, intervene (Human Oversight).
Justice & Fairness¶
AI systems must not lead to unjust discrimination. We apply the Fairness Check to eliminate bias at three levels (Representativeness, Stereotyping, Equal Treatment).
Transparency & Explainability¶
It must be clear to a user when they are communicating with an AI. Decisions made by the system must be explainable in an understandable way.
Privacy & Data Protection¶
Strict compliance with GDPR. Data is only used for the intended purpose and in accordance with the established Hard Boundaries. Source: [so-49]
Societal & Environmental Wellbeing¶
We strive for a positive impact on society and minimise the ecological footprint of our AI systems (energy efficiency).
3. The Fairness Check (Bias Audit) — Extended¶
Audit Levels¶
We assess every High and Limited risk system at three levels:
| Level | Question | Example |
|---|---|---|
| Representativeness | Is the data a good reflection of reality? | Are all customer segments represented in training data? |
| Stereotyping | Does the AI reinforce harmful clichés? | Does the system associate certain professions with specific genders? |
| Equal Treatment | Does every user group receive the same quality of responses? | Is the error margin equal for different age groups? |
Measurable Fairness Criteria¶
We use the following measurable criteria for fairness:
| Criterion | Definition | Formula | When to Apply |
|---|---|---|---|
| Demographic Parity | Probability of positive outcome is equal for all groups | P(Y=1|A=0) ≈ P(Y=1|A=1) | Selection/assignment without legitimising difference |
| Equalized Odds | True Positive Rate and False Positive Rate are equal per group | TPR and FPR equal for A=0 and A=1 | Decisions where both positive and negative errors have impact |
| Predictive Parity | Precision (positive predictive value) is equal per group | Precision equal for A=0 and A=1 | When confidence in positive predictions is crucial |
| Individual Fairness | Similar individuals receive similar treatment | d(f(x), f(x')) ≤ d(x, x') | Personalised service delivery |
Thresholds per Risk Level¶
| Risk Level | Maximum Difference Between Groups | Additional Requirements |
|---|---|---|
| Minimal | Qualitative assessment by Guardian | No quantitative requirement |
| Limited | ≤ 10% difference in Major error rate | Documentation of group comparison |
| High | ≤ 5% difference in Major error rate | Quantitative analysis + documented mitigation plan |
Performing the Fairness Check¶
Step 1: Identify Relevant Groups
- Which protected characteristics are relevant? (gender, age, ethnicity, etc.)
- Note: some characteristics are proxies for protected characteristics (postcode, name)
- Document choices in Risk Pre-Scan
Step 2: Collect or Annotate Data
- Option A: Group labels available in test data
- Option B: Manual annotation of Golden Set subset
- Option C: Proxy variables with justification
- Note privacy: pseudonymise where possible
Step 3: Measure Performance per Group
| Metric | Group A | Group B | Difference | Status |
|---|---|---|---|---|
| Factuality | 98.5% | 97.2% | 1.3% | OK |
| Major errors | 2/75 (2.7%) | 4/75 (5.3%) | 2.6% | OK (\< 5%) |
| Relevance | 4.3 | 4.1 | 0.2 | OK |
Step 4: Analyse and Mitigate
When thresholds are exceeded:
| Cause | Possible Mitigation |
|---|---|
| Data imbalance | Rebalancing, oversampling, synthetic data |
| Bias in source data | Expand data sources, debiasing |
| Prompt bias | Neutral phrasing, explicit instructions |
| Model bias | Threshold calibration, post-processing |
Step 5: Document and Report
Record in Validation Report:
- Which groups were compared
- Which metrics were measured
- Results per group
- Conclusion relative to thresholds
- Mitigation measures (if applicable)
Tooling for Fairness Check¶
| Tool | Type | Strength | Link |
|---|---|---|---|
| Fairlearn (Microsoft) | Python library | Integration with sklearn, multiple metrics | fairlearn.org |
| AI Fairness 360 (IBM) | Python toolkit | Extensive algorithms, good documentation | aif360.mybluemix.net |
| Aequitas | Python library | Focus on auditing, visual reports | github.com/dssg/aequitas |
| What-If Tool (Google) | Visualisation | Interactive exploration | pair-code.github.io/what-if-tool |
Limitations and Considerations¶
Fairness-accuracy trade-off: Optimising for fairness can lead to lower overall accuracy. Document the trade-off.
Incompatibility of criteria: Some fairness criteria are mathematically incompatible. Choose criteria that fit the use case.
Proxy discrimination: Even without direct protected characteristics a model can discriminate via proxies. Test for this.
Intersectionality: Fairness for individual groups does not guarantee fairness for combinations (e.g. young women). Consider subgroup analysis for High Risk.
4. The Role of the Guardian¶
The Guardian acts as the moral compass of the project:
- Guards the Hard Boundaries
- Performs independent ethical reviews
- Has veto mandate for ethical violations
- Approves Fairness Check results
- Escalates for unresolvable fairness issues
Guardian Tasks per Phase¶
| Phase | Guardian Activity |
|---|---|
| Discovery | Assess ethical desirability, define Hard Boundaries |
| Validation | Perform/review Fairness Check |
| Development | Validate mitigation measures |
| Delivery | Final ethical approval |
| Management | Periodic ethics reviews, bias monitoring |
5. Ethical Guidelines Checklist¶
5. Ethical Guidelines Checklist
- Ethical principles have been discussed with the team
- Hard Boundaries are defined in the Objective Card
- Relevant groups for Fairness Check have been identified
- Fairness Check has been performed according to risk level
- Results meet thresholds or mitigation is documented
- Guardian has given ethical approval
- Transparency obligation is implemented (Limited/High Risk)