prEN 18283 — AI bias management (draft harmonised standard)

prEN 18283 is the draft harmonised standard developed by CEN/CENELEC JTC 21 for AI bias management. It operationalises the requirements of EU AI Act Art. 10(2)(f)/(g) and Art. 10(5): the examination of possible biases in training data, measures to detect, prevent and mitigate them, and safeguards for the exceptional processing of special-category data. prEN 18284 (Data Quality) makes a normative reference to prEN 18283 as a means of conformity with Art. 10(2)(f)-(g).

Draft status — no presumption until OJEU citation

prEN 18283 is a confidential harmonised draft from CEN-CENELEC JTC 21. Its full text is not public. Once adopted as a definitive EN and cited in the Official Journal of the EU (OJEU), it will confer presumption of conformity with Art. 10. Until then, compliance with its guidance is good technical practice — not a presumption. AI Act mandatory deadlines for Annex III have been postponed to 2 December 2027.

IP — why clause ids are semantic

The text of prEN 18283 is confidential (paid standard). The clause identifiers in the engine are semantic (bias.*) and anchored in the public, binding text of Art. 10 of Reg. EU 2024/1689, not in the draft’s clause numbers. This page does not reproduce any draft text: it paraphrases the obligations of the public article.

---

What prEN 18283 covers

prEN 18283 defines a systematic process for providers of high-risk AI systems to:

1. Identify the affected groups and bias scenarios relevant to the intended purpose of the system, including groups beyond legally protected classes.
2. Examine training data for possible biases: data gaps, insufficient representativeness, and contextual limits.
3. Measure bias by subgroup with governed thresholds and adequate statistical power.
4. Mitigate identified biases via treatments (debiasing) and evaluate the residual bias risk after mitigation.
5. Monitor bias post-market with review triggers.
6. Document the process in a versioned bias management file.
7. Apply safeguards for the exceptional processing of special-category data (Art. 10(5)(a)-(f)) when necessary to detect and correct bias.

---

Clause catalogue

The file crates/froga-core/resources/standards/pren-18283-clauses.yaml declares the modelled clauses.

Bias examination and identification (Art. 10(2)(f))

Clause	Short title	satisfied_by
bias.affected_groups	Identification of relevant/affected groups from the intended purpose	evidence: affected_persons
bias.examination	Examination of possible biases: bias scenarios, at-risk group, hazards and potential harms	measures_citing: true

A clause with evidence: affected_persons is covered when the signed bundle contains the list of affected persons declared in froga.yaml. A clause with measures_citing: true is covered when at least one signed measure cites it and its associated control passes.

Relevant and representative data (Art. 10(2)(a)-(h), 10(3))

Clause	Short title	satisfied_by
bias.relevant_data	Relevant and sufficiently representative data; examination of data gaps and contextual limits	measures_citing: true
bias.data_sufficiency_gap	Data sufficiency to certify fairness: gap not closable by retraining	measures_citing: true

Subgroup bias measurement (Art. 10(2)(f))

Clause	Short title	satisfied_by
bias.subgroup_measurement	Subgroup bias measurement with governed threshold and statistical power	measures_citing: true

This is the clause that captures the quantitative fairness metrics (e.g. demographic_parity_diff and equal_opportunity_diff in classification, or group_score_gap of per-subgroup Dice in segmentation) declared as controls in the AssuranceProgram. The threshold is what the team declares in froga.yaml, not a number imposed by the standard.

Detection, prevention and mitigation (Art. 10(2)(g))

Clause	Short title	satisfied_by
bias.mitigation	Measures to detect, prevent and mitigate identified biases	measures_citing: true
bias.residual_evaluation	Evaluation of the residual bias risk after mitigation	criterion: residual_present

Special categories (Art. 10(5))

Clause	Short title	satisfied_by
bias.special_category_safeguards	Exceptional processing of special categories: safeguards Art. 10(5)(a)-(f)	measures_citing: true

Monitoring and file

Clause	Short title	satisfied_by
bias.monitoring	Post-market monitoring of bias and review triggers	measures_citing: true, evidence: monitoring_plan
bias.management_file	Signed bias management file: versioned and documented bias scenarios	evidence: signed_bundle

---

How it projects onto the engine

The froga engine implements the ISO 23894 paradigm as the primary engine. prEN 18283 is projected on top via this clause catalogue: froga conformance --standard eu/pren-18283@2026 evaluates clause coverage without rewriting the underlying process.

The same signed bundle that generates the ISO 23894 verdict can be projected onto prEN 18283 without re-annotation. Bias coverage emerges from the AssuranceProgram measures that cite this standard in their standard_clauses field.

Example chain for bias.subgroup_measurement:

• The team declares in froga.yaml a measure with control demographic_parity_diff < 0.15 and cites eu/pren-18283@2026#bias.subgroup_measurement in standard_clauses.
• froga run executes the pipeline, measures the value, and signs it in the bundle.
• froga conformance --standard eu/pren-18283@2026 finds the citing measure with a passing control → emits bias.subgroup_measurement: COVERED.

---

Identified gaps

🟡 Partial — The engine implements bias measurement and projection to prEN 18283. The full mitigation chain and special-category safeguards require explicit citing measures. See Status & gaps.

bias.data_sufficiency_gap: the data sufficiency gap for certifying fairness (Art. 10(3)) requires a measure documenting that the gap exists and cannot be closed by retraining. Without that measure, the clause is a GAP.

bias.special_category_safeguards: Art. 10(5) safeguards apply only when the system processes special categories to detect/correct bias. If the project does not declare them, the clause emits GAP (not necessarily a failure — they may not apply — but the engine does not infer non-applicability automatically).

No enforcement:blocking control on bias: the bias.subgroup_measurement clause may be PARTIAL if the fairness measure has enforcement: audit instead of enforcement: blocking. The engine honestly distinguishes between a monitored metric (advisory) and a gate-blocking control.

---

How to use froga conformance with prEN 18283

# Evaluate conformance against prEN 18283 from the signed bundle
froga conformance --standard eu/pren-18283@2026

# With history per RDD iteration
froga conformance --standard eu/pren-18283@2026 --history

# Export to .froga/conformance/
froga conformance --standard eu/pren-18283@2026 --out

The report emits COVERED / PARTIAL / GAP per clause, grouped by bias management cycle phase.

---

Relation to other standards

• EU AI Act Art. 10 — the legal obligation prEN 18283 operationalises; see EU AI Act.
• prEN 18228 — the risk management standard (Art. 9); fairness/bias is a dimension of risk in the product-safety paradigm. See prEN 18228.
• MDR (SaMD) — in medical systems, subgroup/scanner bias also covers GSPR 17.1. See MDR.
• ISO 23894 — the risk management process underlying the engine. See ISO 23894.