Age, gender, ethnicity, removal of accessories

Key documented metrics could include: Demographic distribution statistics (age, gender, ethnicity) Subgroup performance metrics (precision, recall, FNR/FPR per group) Fairness metrics (e.g., demographic parity, equal opportunity) Data coverage matrices (lighting, occlusion, accessories like glasses/hats) These metrics help demonstrate balanced representation and consistent performance.

Document demographic coverage ratios vs. target population, per-subgroup performance metrics (TPR/FPR/FNR gaps), confidence intervals by subgroup, distributional similarity scores (e.g., KL divergence) between training and real-world data, and fairness deltas showing no statistically significant performance degradation across age, gender, ethnicity, or physical attributes.

I believe the name, age, gender and phone number should be documented

To prove fairness, developers should clearly show who is represented in the training data and how the system performs for each group. This includes age, gender, ethnicity, and physical features like glasses or hats. Comparing accuracy across groups helps confirm the system works equally well for everyone and does not unintentionally favor certain users.

Documented metrics should include demographic coverage ratios, balanced class distributions, subgroup-specific accuracy/false-positive/false-negative rates, and fairness metrics (e.g., equalized odds) across age, gender, ethnicity, and physical attributes.

Age, gender and race

To prove the data is unbiased, you must document Proportional Representation (balancing age, gender, and ethnicity) and Attribute Parity, ensuring physical traits like glasses or hats are represented across all skin tones. The definitive metric is the Disparate Impact Ratio, which confirms that error rates remain equally low for every demographic group.

Equal proportions of people of all ethnicities, as well as equal proportions of clothing styles typical of the area of deployment.

Demographic distribution tables showing proportions of age groups, gender identities, ethnic backgrounds, and physical characteristics (e.g. glasses, facial hair, head coverings). Performance parity metrics, such as: False positive rate (FPR) and false negative rate (FNR) per demographic group. Accuracy, precision, and recall disaggregated by subgroup. Statistical fairness measures, such as: Difference in error rates between majority and minority groups. Confidence intervals to show robustness of results. Data provenance documentation, explaining where data originated, how it was collected, and known limitations. Synthetic data validation, demonstrating that synthetic samples meaningfully improve representation without introducing artefacts or amplifying bias. While “perfect neutrality” is unrealistic, regulators will expect evidence of active bias mitigation and continuous monitoring rather than mere assertions of fairness.

Physical characteristics and age

Gender

A dataset can only be considered “unbiased” if it shows demographic coverage parity, balanced representation, and equivalent safety performance (especially FNR) across all driver groups, supported by transparent documentation and independent audits.

Demographic Coverage Ratios Minimum Samples per Subgroup Intersectional Coverage Condition & Accessory Coverage Label Consistency Outcome Parity Metrics

How diversified is the dataset being used to train the model. Number of races Age groups Gender Physical characteristics

A diverse, yet big enough sample size

Metrics that are willingly given by participants and that don't put their personal lives at risk. Their physical appearance, age, ethnicity are some examples of metrics that can be documented. Perhaps a log of how long they've been on the road can be included as well.