Legal AI: Evaluating Efficiency and the Rule of Law

Speed, cost, accuracy, accessibility and procedural fairness are different measures of a legal service. An AI tool can improve one and worsen another. The practical question is which changes occur in the actual process and what safeguards are needed before people rely on it.

Separate an ethical argument from measured effects

The 2025 chapter by Metin and Kerikmäe on legal technology and the rule of law argues for attention to the people affected by legal services, beyond provider efficiency. Greenstein's analysis examines how opaque automated decisions challenge rule-of-law values. These are conceptual legal arguments, not measurements showing that every AI deployment improves or undermines justice.

Empirical research can test more specific questions. A 2025 study of judicial-AI perceptions used bail and sentencing vignettes and found that perceptions varied with AI use and participant groups. It measures responses to scenarios, not actual judicial fairness or Swiss public opinion.

The diagram is an evaluation aid: record speed and fairness as separate measures, because they can move in different directions and do not reduce to a single score.

Examine the process from the user's position

Ask whether the person understands the proposed action, can correct relevant information, knows who is responsible and has a usable route to challenge an error. Test these questions with people facing realistic constraints, including language, disability, limited time and unequal access to technology.

An explanation should communicate the grounds that matter to the decision. A plausible account generated after the fact is not necessarily a faithful explanation. Where a source supports a legal proposition, inspect that source rather than treating fluency as evidence.

Where the GDPR applies, Article 22 gives people the right not to be subject to a decision based solely on automated processing that produces legal effects concerning them or similarly significantly affects them, subject to its exceptions. The Court of Justice held in SCHUFA, C-634/21, on 7 December 2023, that an automatically generated credit score can itself be such a decision where a third party draws strongly on it to establish, implement or terminate a contractual relationship with that person, and in C-203/22, on 27 February 2025, that, under the right of access in Article 15(1)(h), the person may require an explanation of the procedure and principles actually applied. In Switzerland, FADP Article 21 requires the controller to inform the person of an automated individual decision and, on request, to let them state their view; the person may ask for the decision to be reviewed by a natural person, subject to the exceptions in that article.

Magesh and colleagues' evaluation, published in the Journal of Empirical Legal Studies in 2025, found remaining hallucinations in tested commercial legal research systems despite retrieval support. It concerns the versions tested at the time and US research tasks, not every current product. It supports checking actual propositions and omissions in the intended workflow.

Identify the applicable duties and dates

The current EU AI Act, Article 6 and Annex III, identifies specified justice-related uses, including certain assistance to judicial authorities or on their behalf and analogous alternative dispute resolution. It does not classify every private legal tool as high risk. The purpose, conditions, role and relevant exceptions matter.

Under Articles 111 and 113, as amended by Regulation (EU) 2026/1744, the relevant Annex III high-risk duties apply from 2 December 2027 and the duties for product-related high-risk systems under Article 6(1) and Annex I from 2 August 2028, with existing-system transitions. Article 50 transparency has a separate timetable and scope. Record the obligation that actually applies instead of requesting a conformity declaration from every supplier regardless of classification.

The same regulation added two prohibitions to Article 5(1), points (ba) and (bb), which apply from 2 December 2026 under Article 113, third paragraph, point (a). They cover AI systems that generate or manipulate realistic images, videos, audio or similar material of an identifiable person's intimate parts or of an identifiable person engaged in sexually explicit activities without that person's freely given, specific, informed, unambiguous and explicit consent, and systems that generate or manipulate child sexual abuse material or performances within the meaning of Article 2(c) and (e) of Directive 2011/93/EU, except where a “without right” defence applies under national law. Under Article 5(1a), placing such a system on the market or putting it into service is prohibited only where that output is its intended purpose, or a reasonably foreseeable and reproducible outcome without adequate safeguards, and use is prohibited only where the deployer uses the system for that purpose; for point (ba), Article 5(1b) excludes manipulation that neither increases the exposure of depicted intimate parts nor alters the nature of any depicted sexually explicit activity. Include these prohibitions when you check Article 5 before any classification. No label or notice makes a prohibited use lawful.

For Swiss legal practice, the statutory anchors are Article 13 of the Lawyers Act (BGFA) and Article 321 of the Swiss Criminal Code. Under Article 13 BGFA, lawyers within the scope of that Act are bound by professional secrecy, without time limit and towards everyone, about everything clients entrust to them in their professional capacity, and must ensure that their auxiliaries preserve it; Article 321 makes the disclosure of such secrets by lawyers and their auxiliaries punishable on complaint. The SAV's February 2025 guidance informs assessment of professional secrecy, supplier arrangements and independent review. The same page hosts the CCBE guide on generative AI of 2 October 2025, which asks lawyers not to enter client data without appropriate safeguards and, where the use case requires, to verify output before using it. Hosting location alone does not decide compliance. Professional guidance, an ethical charter and a statutory obligation also have different legal status.

Swiss regulatory status (checked 26 September 2026). Switzerland has no AI-specific legislation yet; the FDPIC states that the current FADP applies directly to AI-supported data processing. On 12 February 2025 the Federal Council stated that Switzerland intends to ratify the Council of Europe AI Convention, which applies primarily to state actors. The Federal Department of Justice and Police is to prepare a consultation draft by the end of 2026, in particular on transparency, data protection, non-discrimination and supervision. Switzerland signed the Convention on 27 March 2025. When this was checked, the Federal Office of Justice (German text) did not yet list a published draft.

Make the assessment reviewable

Ada Studio proposes recording benefits and harms in the same evaluation, with acceptance criteria chosen before reviewing the results. Assign operational responsibility without pretending that a project chart settles statutory liability.