EU AI Act: High-Risk AI in Justice and Democratic Processes (Annex III Point 8)

Overview: Protecting Courts and Elections From Unchecked AI

The eighth and final area in Annex III of the EU AI Act (Regulation (EU) 2024/1689) addresses two institutions on which every other right depends: the administration of justice and democratic processes. The recitals are direct about the stakes. AI tools supporting judicial decisions can affect the right to an effective remedy and a fair trial, and must not interfere with judicial independence or the principle that judges decide. AI aimed at swaying votes touches the foundations of democracy itself, a concern sharpened by the rise of generative systems capable of producing persuasive content at scale. Point 8 places both under the high-risk regime rather than banning them, on the logic that carefully governed judicial support tools and clearly bounded political technology can exist — but only with documentation, oversight and accountability.

What Annex III Point 8 Covers

The provision contains two use cases:

• 8(a) Administration of justice: AI systems intended to be used by a judicial authority, or on its behalf, to assist a judicial authority in researching and interpreting facts and the law and in applying the law to a concrete set of facts, or to be used in a similar way in alternative dispute resolution
• 8(b) Democratic processes: AI systems intended to be used for influencing the outcome of an election or referendum or the voting behaviour of natural persons in the exercise of their vote in elections or referendums

Both use cases carry significant carve-outs. For 8(a), the recitals clarify that the classification does not extend to purely ancillary administrative activities that do not affect the actual administration of justice in individual cases: anonymisation of judgments, communication between staff, or resource allocation in court administration. For 8(b), the provision expressly excludes AI systems whose output natural persons are not directly exposed to — tools used to organise, optimise or structure political campaigns from an administrative or logistical point of view, such as canvassing route planning or volunteer scheduling.

The Justice Use Case in Practice

The covered territory in 8(a) is decision support at the core of adjudication: systems that retrieve and rank case law for a pending matter, draft analyses applying legal rules to case facts, propose sentencing ranges, or evaluate the strength of arguments. The same logic applies to alternative dispute resolution where the proceedings produce legal effects for the parties — online arbitration platforms and ombudsman services using AI to propose or shape outcomes are within scope.

What stays outside is meaningful for court administrators: case management dashboards, transcription, translation of documents for accessibility, scheduling systems and statistical reporting do not assist in researching, interpreting or applying the law to a concrete case. The boundary question to document for each tool is whether its output can plausibly shape the substantive outcome of an individual case. A citation search engine sits near the line; a tool that summarises the merits of the parties' arguments is over it.

Note that point 8(a) addresses use by or on behalf of judicial authorities. Legal research tools sold to private parties for their own case preparation are not within this point, although other parts of the regulation, including transparency duties for AI-generated content, may apply.

The Elections Use Case in Practice

Use case 8(b) targets systems whose intended purpose is influencing votes: micro-targeting engines that select and tailor political messages to individual voters, persuasion optimisation systems, and AI-driven political advertising delivery built to shift voting behaviour. The exclusion for logistics tools means that the classification turns on exposure: if natural persons are directly exposed to the output of the system — the message, the ad, the recommendation — and the intended purpose is electoral influence, the system is high-risk.

This provision operates inside a wider legal lattice. Regulation (EU) 2024/900 on the transparency and targeting of political advertising restricts targeting techniques and requires labelling. The Digital Services Act imposes systemic risk duties on very large platforms regarding electoral integrity. The AI Act adds the product-level layer: providers of election-influence systems must build risk management, data governance, logging, human oversight and accuracy evidence, register the system in the EU database, and undergo conformity assessment. Deepfake content rules under Article 50(4) apply cumulatively: AI-generated or manipulated image, audio or video content must be disclosed, with specific arrangements for evidently artistic or satirical content.

Who Must Comply

For 8(a), providers are typically legal technology vendors and, importantly, justice ministries or court services that develop tools in-house and put them into service under their own name — public bodies can be providers. Deployers are courts, tribunals and ADR bodies. Public authority deployers must complete a fundamental rights impact assessment under Article 27 before first use, and registration occurs under Article 49. Under Article 74(8), market surveillance for AI used in the administration of justice is assigned to data protection authorities or other authorities meeting equivalent independence conditions, reflecting the constitutional sensitivity of supervising tools used by courts.

For 8(b), providers are the developers of voter-influence systems; deployers may be parties, campaign organisations or consultancies. All face the standard high-risk obligations, and the practical effect is that opaque persuasion technology becomes difficult to operate lawfully in European elections.

Practical Compliance Steps

1. Courts and ministries: inventory every AI tool in judicial workflows and classify against the ancillary-versus-adjudicative boundary, documenting reasons
2. Verify judicial independence safeguards: outputs must be advisory, judges must see system limitations and error characteristics, and training should address automation bias on the bench
3. Legal tech vendors: prepare Annex IV technical documentation, define the intended purpose precisely, and complete conformity assessment before August 2, 2026
4. Campaign technology providers: re-examine product positioning — a tool marketed for electoral persuasion is high-risk by intended purpose, and contractual disclaimers do not change classification
5. All deployers: retain logs, assign trained oversight, and prepare to give affected persons meaningful explanations under Article 86 where decisions rest on system output

Concrete Examples

Example one: a member state pilots an AI assistant that drafts preliminary analyses of asylum appeal files for judges, summarising facts and suggesting applicable precedent. That assists in researching and interpreting facts and law — high-risk under 8(a), requiring impact assessment, oversight design and registration before rollout.

Example two: an online consumer ADR platform uses a model to propose settlement amounts that become binding if both parties accept. High-risk under 8(a) as AI used in a similar way in alternative dispute resolution.

Example three: a campaign uses AI to plan door-knocking routes and schedule volunteers. Directly excluded as logistical organisation. The same campaign using AI to generate individually tailored persuasion messages for undecided voters is operating a high-risk system, and any synthetic media it produces must be labelled under Article 50.

Action Before August 2, 2026

Justice institutions move deliberately, which is precisely why work must start now. Court services should complete inventories and impact assessments in 2026, build procurement language requiring conformity-assessed tools, and establish a review board pairing judges with technical experts. Vendors should finalise documentation and consider that judicial customers will demand evidence beyond the legal minimum, because every tool they adopt will eventually be challenged by a losing party. On the electoral side, the August 2026 date lands ahead of several national election cycles; providers of campaign technology should assume that the first enforcement actions in this area will be public, fast and politically charged. The systems that survive will be those whose providers can show exactly what the tool does, on what data, with what accuracy, and under whose supervision — which is, in the end, what citizens are entitled to expect of any machinery operating inside their courts and elections.

Judicial Independence as a Design Requirement

One feature distinguishes 8(a) from every other Annex III point: the protected interest is not only the individual litigant but the institution of independent adjudication. The recitals insist that the use of AI tools can support but must not replace the decision-making power of judges, and that final decision-making must remain a human-driven activity. Translated into system design, this means judicial AI should be built to inform rather than converge: presenting competing precedents rather than a single answer, exposing uncertainty rather than projecting confidence, and recording what the judge saw so that appellate courts can later assess whether the tool framed the decision. Court services should also consider disclosure policy — whether parties are told that an AI tool was used in preparing a case analysis — because procedural fairness arguments will arrive in appellate litigation regardless of what the AI Act itself requires. Several European judicial councils have begun issuing internal guidance on these questions, and aligning deployments with both the AI Act and that institutional guidance is the safest course for ministries and vendors alike.

This article is for informational purposes only and does not constitute legal advice. Regulatory requirements change frequently — verify current rules with official sources. Built by Sawai Gyoseishoshi Office, Hiroshima, Japan.