If your AI helps judicial authorities research or interpret facts and law, applies law to a concrete set of facts, or is intended to influence the outcome of elections, referenda, or voting behaviour, the EU AI Act puts you in scope under Annex III, point 8. A narrow carve-out applies to administrative tasks that do not affect substantive justice.

This is the smallest Annex III category and the one with the highest fundamental-rights weight. Decisions made with the help of these tools shape sentences, civil rights, and elections. The Act treats the audit trail as the public record that justifies the decision.

The Digital Omnibus on AI (Regulation (EU) 2026/1744) delayed the Annex III high-risk application date to 2 December 2027. The operational-obligation penalty band is €15 million or 3% of worldwide annual turnover. Article 50 transparency obligations layer on top for any AI-generated text on matters of public interest — and are already in force.

The articles you have to satisfy

Justice-and-democratic-process AI requires every Annex III obligation, plus the Article 50 transparency layer where AI-generated content is published.

  • •Article 9 — risk-management system over the lifecycle, with explicit attention to foreseeable misuse.
  • •Article 11 + Annex IV — technical documentation: intended purpose, accuracy and limitations, performance on legal-domain tasks (case-law retrieval, legal-reasoning aids), the rationale for any decision threshold, training-data lineage.
  • •Article 12 — automatic event logs across the lifetime of the system.
  • •Article 13 — instructions for use with the explicit capabilities and limitations.
  • •Article 14 — human oversight by a qualified judicial or operational user; ability to override, refuse, or reverse.
  • •Article 17 — documented quality management system.
  • •Article 19 / Article 26(6) — log retention of at least six months. Judicial-records rules and electoral-oversight rules will typically extend this materially.
  • •Article 26 — deployer obligations: assign human oversight to a competent person, monitor operation, suspend use when risk is identified, keep the logs.
  • •Article 27 — Fundamental Rights Impact Assessment by the public-authority deployer before first use.
  • •Article 49 — registration of the system in the EU database.
  • •Article 50 — for AI-generated text on matters of public interest, disclose unless human review with editorial responsibility applies. For deepfakes, disclose.
  • •Article 72 — post-market monitoring.
  • •Article 73 — serious-incident reporting; 15 days, 10 if a fundamental-rights infringement is widespread or a person dies.

The audit trail is the proof, for each output the system produced, that all of these obligations were satisfied.

Why hallucination is a compliance event here

A legal-research aid returns a citation. The judge's clerk pastes it into a draft. The citation does not exist. The decision goes out with a fictional case in it. The next week the opposing counsel notices.

It has already happened in jurisdictions that did not have the AI Act. Under the Act, the deployer (the court or the lawyer's office, depending on the use) is responsible for assigning human oversight to a competent person, for ensuring the system was used in accordance with the instructions for use, and for keeping the logs. The provider is responsible for documenting the system's known limitations — including the propensity to produce false citations — and the accuracy metrics on real legal-domain tasks.

If the technical documentation says the system has been measured on legal-domain tasks at a certain accuracy and the logs show the clerk used it without verification, the audit trail at least makes the failure attributable and reviewable. If neither exists, the failure is also a regulatory breach.

The villain is Ship and Pray, in a judicial robe. A demo, a pilot, a tool live in chambers. The system was evaluated on the dataset the vendor had. Acceptance criteria were "the answers look plausible." Hallucination rate on case-law retrieval was never measured. Instructions for use are a PDF nobody read. The audit trail is the order that got published with the fake citation.

A regulator will reject that position. A court will reject it. In the justice context, the press will print it.

The justice-and-democratic audit-trail checklist

For each AI feature in scope, you should be able to produce on demand:

  1. •Annex IV technical documentation — including accuracy on the actual legal-domain tasks (case-law retrieval, citation generation, summarisation of statutes, application of law to facts) and the documented hallucination rate at the chosen configuration.
  2. •Risk-management dossier under Article 9 — foreseeable misuse, mitigations, residual risk, basis for acceptance.
  3. •Versioned model, prompt, and configuration files — every change with date, author, evaluation, and approval.
  4. •Pre-deployment evaluation reports — signed off by a domain expert (a senior lawyer, a judicial-policy specialist, an electoral-integrity expert) with the authority to refuse to ship.
  5. •Runtime event logs under Article 12 — sufficient to reconstruct each output the system produced, the input that produced it, the model version, and any human review.
  6. •Human-oversight records under Article 14 — overrides, refusals, escalations.
  7. •Article 50 transparency artefacts — where AI-generated text is published, machine-readable provenance marks, deepfake disclosures, public-interest disclosure where applicable.
  8. •Post-market monitoring data under Article 72.
  9. •Fundamental Rights Impact Assessment under Article 27 (for public-authority deployers).
  10. •EU database registration under Article 49.
  11. •Serious-incident register under Article 73.

If you cannot produce all of these for an output that ends up in a published decision or an electoral campaign, the system is operating without the audit trail the Act requires.

Where domain experts have to lead

Article 14 requires oversight by personnel with the competence to provide it. In justice, that means senior judges, legal-research specialists, judicial-policy experts, and electoral-integrity specialists. In democratic-processes contexts, it means election officials, civil-society experts in campaign integrity, and platform-governance specialists.

Standard evaluation tooling shuts them out. The eval set lives in a notebook nobody outside engineering has opened. The acceptance criteria are decided by a vendor. The hallucination rate is described as "low" with no number behind it. The instructions for use are a generic legal-disclaimer paragraph. By the time a domain expert reviews the tool, it is already in chambers.

Lovelaice is built for this gap. The domain expert defines what acceptable looks like before the model touches data — factual-correctness floors, citation-verification requirements, refusal behaviour on out-of-distribution prompts, sub-group performance where relevant, the language of the rationale text. They evaluate outputs directly through a simple interface, with blind review that removes bias from the scoring. Every evaluation is captured. Every prompt and configuration version is immutable. Every result is exportable.

When a national court, a parliamentary committee, or an electoral-oversight body asks how a particular output was produced, the answer is a report, traceable from input to reviewer to decision.

Retention and reporting windows to design for

  • •Article 50 transparency obligations apply to AI-generated text and deepfakes in the public-interest context: disclose unless human review with editorial responsibility applies.
  • •Six months minimum log retention. Judicial-records rules and electoral-oversight rules typically extend this materially.
  • •Ten years retention of Annex IV technical documentation and the EU declaration of conformity.
  • •15 days to file a serious-incident report. 10 days if a person dies or a fundamental-rights infringement is widespread.

Justice and democratic processes is the category where every output the AI produces may end up on a public record. Speed without proof is a fictional citation in a real ruling. Build the audit trail before the first output is read.

Sources

More in this series

The EU AI Act audit-trail series — one article per Annex III high-risk category:

Enjoyed reading? Add us as a preferred source in Google to support us.

Add Lovelaice as a preferred source