Bias and fairness audits for AI: Swiss equality law, EU AI Act Art. 10, BBQ and StereoSet

What are bias and fairness audits?

Bias and fairness audits check whether an AI system treats people or groups systematically differently – based on characteristics like gender, age, language, origin, social status. "Bias" is the statistical skew in output, "fairness" the normative requirement that this skew does not lead to unlawful discrimination.

Bias sources in LLM pipelines are diverse. First: training data – if the 2023 web was dominantly male-Western-English-speaking, the model internalises matching patterns. Second: prompt design – phrasings can amplify expectations ("a successful CEO" is more often described as male). Third: retrieval setup in RAG – if indexed documents are predominantly German-language, French-speaking clients are disadvantaged. Fourth: decision threshold – a classification threshold well calibrated for one group can produce misclassifications for another.

As of May 2026 several established bias benchmarks are available. BBQ (Bias Benchmark for QA) from NYU tests social bias in QA settings across nine demographic categories. StereoSet tests stereotype susceptibility on gender, profession, race and religion. CrowS-Pairs (Crowdsourced Stereotype Pairs) measures a model's preference for stereotype-conforming versus anti-stereotypical sentences.

For Swiss practice these tests are a starting point, not an end. Swiss equality law (Const. Art. 8, Equal Opportunities Act) and EU AI Act Art. 10 (data quality, "free of biases") set the legal frame. Concrete application tests must be language- and cultural-region-specific.

Why it matters

Swiss law sets clear limits. Const. Art. 8 prohibits discrimination by origin, race, gender, age, language, social position, lifestyle, religious, ideological or political conviction, or physical, mental or psychological disability. The Equal Opportunities Act (GlG) makes this concrete for employment. Under FADP Art. 21 on automated individual decisions, the affected person has a right to reasoning – algorithmic discrimination must be disclosable and correctable.

EU AI Act Art. 10 requires for high-risk systems that training, validation and test data are "free of biases" to the extent technically achievable. Bias-test methodology must be presentable to oversight.

The fiduciary use case: client routing. If the AI auto-routes incoming client requests to case workers, sorting by "language quality", that can be discriminatory – a client with a Ticino Italian accent in a German-language email is systematically classified as "complex" and forwarded to junior staff, while a Zurich-German client goes to senior advisors. That is both legally problematic (Const. Art. 8 – language) and commercially (mandates are treated unequally).

Commercially fairness is a trust signal. Clients, employees and supervisors increasingly expect documented fairness tests. Without that you lose mandates to competitors who have them.

How it works – methods and benchmarks

BBQ (Bias Benchmark for QA). Parrish et al. 2022 (NYU). Test set with 58,000 questions across nine demographic categories (age, disability, gender identity, nationality, physical appearance, race/ethnicity, religious preference, social status, sexual orientation). Each question has two variants: ambiguous (insufficient info, correct answer = "don't know") and disambiguated (sufficient). Bias is measured when the model gives stereotype-conforming answers in the ambiguous case. Current frontier models reach very low bias scores near 0 (the goal) today, while older and open models sit considerably higher.

StereoSet. Nadeem et al. 2021 (Microsoft / Carnegie Mellon). Multiple-choice: given a context sentence, three possible completions: stereotype-conforming, anti-stereotype, unrelated. Measures whether the model prefers stereotype-conforming or anti-stereotypical completions – either is a bias signal. Ideal value: 50/50 between stereotype and anti-stereotype (model has no systematic preference).

CrowS-Pairs. Nangia et al. 2020. Crowdsourced Stereotype Pairs. 1500 sentences, each as a pair (e.g. "He works as a doctor" / "She works as a doctor") – bias score = share of model preferences for the stereotype pair. Simpler than StereoSet, hence often used as a quick check.

HELM (Holistic Evaluation of Language Models). Stanford CRFM, continuously updated. Aggregates several bias benchmarks (including BBQ) plus performance tests in a standardised framework. As of May 2026 the reference for academically rigorous bias assessment.

Custom bias tests for the Swiss context. Generic benchmarks are often English-only. For DE-CH applications we build custom tests covering Swiss realities: does the AI treat a "Frau Müller" differently from a "Herrn Müller"? Are queries with French vocabulary in German emails answered with equal quality? Does the AI classify dialect-tinted requests as "less serious"? Such tests are built as an extension of the golden dataset – with 20-50 cases per axis.

Disparate impact analysis. For classification pipelines: check threshold per subgroup. If the AI hits 90% accuracy for one group and 70% for another, that is disparate impact. With legitimate performance differences (e.g. less training data for a language), this must be openly reported and compensated – either with additional training or with human-review support for the less-covered group.

Mitigations. When bias is detected: (a) prompt tuning ("treat all queries equally regardless of ..."), (b) output filter with bias detector, (c) model switch to less biased model, (d) additional training data for the underrepresented group, (e) human-in-the-loop escalation for suspicious cases.

Bias audit in 6 steps

1. 01Identify risk axes: for which attributes (gender, language, age, region) is bias likely problematic?
2. 02Run standard benchmarks: BBQ, StereoSet, CrowS-Pairs on the deployed model – measure baseline.
3. 03Build custom Swiss tests: 20-50 cases per axis, language- and cultural-region-specific.
4. 04Disparate impact analysis on real application data: accuracy, latency, quality per subgroup.
5. 05Prioritise and mitigate findings: prompt tuning, output filter, model switch, human review backstop.
6. 06Quarterly repetition, oversight report annually or on substantial pipeline change.

When bias audits are mandatory

For automated individual decisions under FADP Art. 21 – always. When the AI decides on mandate acceptance, lead scoring, credit score or HR pre-selection, fairness testing is mandatory.

For applications on the EU AI Act high-risk list (Annex III) – always. Justice, HR, education, credit, insurance, critical infrastructure. EU AI Act Art. 10 is mandatory here.

For client routing and client-response pipelines in fiduciary and law firms – recommended even without hard regulatory pressure. Reputation damage on detected discrimination is high.

For multilingual setups (DE/FR/IT/EN) – always. Language disparity is the most common bias class in Swiss applications. What hits 92% accuracy in German often only reaches 78% in Suisse-Romand French.

On model switches – bias profile changes. Different models and generations have different bias patterns. Test before migration.

For purely internal tools with no external decision effect (brainstorming, knowledge lookup), an annual check suffices.

When the effort is lower

For purely generative use cases without decision impact (slogan generator, image concept sketches, internal note suggestions), a full bias audit suite is overkill. An annual quick check with CrowS-Pairs on the deployed model suffices.

For pipelines whose output goes exclusively to humans with full review competence (senior lawyer, senior fiduciary) without further automation: a one-off initial audit, then annually. The human is the fairness backstop.

Beware: bias mitigations can introduce new bias. A prompt instruction "treat all queries equally" can cause the model to drop legitimate differentiation (e.g. for legal complexity) and lose quality. Bias mitigation must always be evaluated measurably, not just "patched on".

Benchmarks like BBQ are not fully transferable. They are English-language and US-demographic. For Swiss use cases they are a hint, not proof. Building custom tests is mandatory.

Trade-offs

FAQ

Related topics

Sources

1. Parrish et al. – BBQ: A Hand-Built Bias Benchmark for Question Answering · 2022-05
2. Nadeem et al. – StereoSet: Measuring stereotypical bias in pretrained language models · 2021-08
3. Nangia et al. – CrowS-Pairs: A Challenge Dataset for Measuring Social Biases · 2020-11
4. Stanford CRFM – HELM (Holistic Evaluation of Language Models) · 2026-04
5. EU AI Act, Article 10 – Data and Data Governance · 2024-07
6. EDÖB – Leitfaden zu automatisierten Einzelentscheidungen · 2025-11