Research questions

H1

The administrative offence of unauthorised provision of legal services (§ 52d of the Advocacy Act) applies to the operator of a software tool only where the elements of a continuing activity and of payment tied to legal assessment are met, not to the supply of the tool as such.

Sources in hand

• Act No. 85/1996 Sb., on the Legal Profession (zákon o advokacii), § 52d, § 1 odst. 2 and § 2 odst. 3 (Zákony pro lidi). Wording checked against the consolidated text from the ČAK (Czech Bar Association) (PDF).
• Act No. 40/2009 Sb., the Criminal Code (trestní zákoník), § 251, unauthorised business (Zákony pro lidi).
• An analysis of the perfective verb aspect (the Czech "poskytne", "provides") and of the consideration dissolving into a flat fee, see the note Research Question 01.
• A US parallel: OpenAI's defence rests on the argument that a subscription is not payment for legal assessment and that a tool is not a person, see the memorandum of 15 May 2026 in Nippon v. OpenAI.
• The professional-body reading: AI in any form cannot be a provider of legal services within the meaning of § 1 of the Advocacy Act, and the lawyer's responsibility when using it remains untouched. Stanovisko ČAK k užívání umělé inteligence při poskytování právních služeb, September 2023, PDF at cak.cz.
• The professional-body layer is in motion: the ČAK section for artificial intelligence and new technologies is preparing an update to the opinion, a long-term strategy, and IT standards for lawyers and legal-technology developers alike. Interview with Barbora Vlachová, member of the ČAK board, epravo.cz, 9 June 2026.
• The line between service and information in the Czech literature: a contract template for anyone is legal information, a bespoke contract is a legal service, and a template adjusted by AI on the client's instructions hangs in between. Michal Kuk, Přístup k právu a AI, Bulletin advokacie 7/2025, s. 27.

The anti-circumvention layer too (§ 2 odst. 3, second sentence, on intermediation) is tied to the person of the provider. It therefore fails for the same reason as the basic definition of the offence.

Statusin progress in RQ-01 · missing: the Ministry of Justice's administrative-offence practice under § 52d (request under Act 106/1999) and commentary doctrine on § 1 odst. 2

H2

The BGH test from the smartlaw case, distinguishing a legal service from a legal-information product by the degree to which the assessment is individualised, carries over to the Czech setting and provides a workable dividing criterion.

Sources in hand

• BGH, judgment of 9 September 2021, sp. zn. I ZR 113/20 (the smartlaw contract-document generator): mechanically assembling a document from predefined answers is not a legal service, because it lacks any legal assessment of the specific case within the meaning of § 2 odst. 1 RDG (the German Legal Services Act). Primary text still to be added.
• The Czech foothold for carrying it over: "the giving of legal advice" and "other forms of legal assistance" in § 1 odst. 2 of the Advocacy Act (Zákony pro lidi).
• The same dividing line conceptually: an impassable boundary between legal information and a tailored legal conclusion. Eran Kahana, Stanford CodeX, and the analysis in the note Nippon v. OpenAI.
• The American mirror of the same test: the practice of law requires the application of legal knowledge and skill (Clark v. Gannett, King v. First Capital, cited per the OpenAI memorandum).
• The same conclusion about generators from the academic side: large language models can produce a generic, superficially convincing contract, with no guarantee of internal consistency or fitness for a specific transaction. Eliza Mik, Contractual Deepfakes: Can Large Language Models Generate Contracts?, arXiv, 2026.

If the test holds, H1 and H2 fit together: supplying the tool is a product, individualised assessment is a service, and the boundary runs along the degree of individualisation.

Statusawaiting the primary BGH text and Czech doctrine · candidate for note RQ-02

H3

The AI Act does not fill this gap: legal-advice systems from private providers are not high-risk systems under Annex III, which is aimed only at the administration of justice.

Sources in hand

• Regulation (EU) 2024/1689 (the AI Act), Article 6(2) and Annex III, point 8(a), EUR-Lex, CELEX 32024R1689. Point 8(a) covers systems used by a judicial authority or on its behalf and comparable use in out-of-court dispute resolution; private advisory systems do not appear in Annex III (verified as of 6 July 2026).
• Article 50 of the AI Act (transparency): the only layer that reaches such systems is the duty to inform about interaction with AI, effective from 2 August 2026. A duty to say that something is AI is not a requirement on the quality of the output.
• Context on the delays: Regulation (EU) 2026/1744 (published 24 July 2026) pushed the high-risk obligations back to between December 2027 and August 2028. Even if the classification did apply one day, the obligations are receding.
• A running commentary on the whole regulation (TL;DR, deadlines, chapters), the AI Act card on this site.
• A contrast from the same annex: assessing job applicants is high-risk, private legal advice is not. Šimon Svoboda, Riziko diskriminace uchazečů o zaměstnání při využívání AI systémů k jejich hodnocení, Revue pro právo a technologie 29/2024, s. 29.
• Agentic AI opens the gap further: systems that autonomously plan and carry out multi-step actions run up against the regulation's requirements (human oversight, transparency across chains of actors, behavioural drift at runtime) even where the regulation does apply. Nannini et al., AI Agents Under EU Law, arXiv, 2026 (working paper).

What matters for the strength of the hypothesis is that the gap is not an oversight: the EU legislature regulates systems by risk and did not mark private legal advice as risky.

Statuscore verified against the text of the regulation · missing: the recitals to point 8 and the Czech implementing framework (supervision)

H4

De lege ferenda, output-based regulation (requirements on the quality, verification, and transparency of the output) is more effective than input-based regulation through the professional body.

Sources in hand

• Garfield AI (England): a regulator-approved "AI law firm", where the system does the paperwork and a human argues in court, exactly an output-based model. The Economist, The rise of vibe lawyering (29 June 2026), summary in the TL;DR.
• Illinois Supreme Court Policy on Artificial Intelligence (effective 1 January 2025): the use of AI "is expected and should not be discouraged", with a human bearing responsibility for the final product, illinoiscourts.gov.
• Jones v. Kankakee County Sheriff's Department, 164 F.4th 967 (7th Cir. 2026): responsibility for a filing under Rule 11 as an output anchor (cited per the OpenAI memorandum).
• Kahana's safe-harbour conditions (deterministic brakes, auditability, knowledge of the jurisdiction) as a concrete form of requirements on output and architecture, Stanford CodeX, the analysis in the note.
• The technical limit of input control: AI-text detectors are unreliable and easily circumvented, Stanford HAI and NIST AI 100-4.
• Tools of output-based regulation in the literature: regulation through technical standards and certification (chapter 2) and product-defect liability with mandatory insurance (chapter 5), Mesarčík, Gyurász et al., Právo a umelá inteligencia (PF UK v Bratislavě 2024, open textbook, CC BY-NC-ND).
• An ex post layer already exists in Czech law: protection of personality (§ 81 and § 82 of the Civil Code) and the principle of data accuracy (Article 5(1)(d) GDPR) as a defence against hallucinations, including the NOYB complaint against OpenAI. Langhoffer, Nonnemann, Právo vs. halucinace: Co dělat, když umělá inteligence lže?, Právní rozhledy 13–14/2025, s. 436.
• The Czech courts already assess the output, not the tool: the use of machine translation of documentary evidence does not in itself make a judgment unreviewable, NSS 11 September 2025, sp. zn. 5 Azs 95/2025-42, annotation in Revue pro právo a technologie 32/2025.
• And the sanctions practice targets the output, not the use of the tool: the Ústavní soud (Constitutional Court) imposed a procedural fine of CZK 25,000 on a lawyer for a complaint resting on non-existent case law (ruling of 1 December 2025, sp. zn. I. ÚS 3004/25), and the Nejvyšší správní soud (Supreme Administrative Court) refused to award costs for a submission with invented citations (judgment of 15 October 2025, č. j. 3 As 34/2025-80, cited per Hlovjak, Šlechta, Možné důsledky nesprávného použití AI v procesních podáních, epravo.cz, 13 January 2026). Both courts expressly allow the use of AI, while the ČAK disciplinary line has not yet been decided.
• Responsibility for the output holds outside legal services too: a business is liable for its chatbot's mistaken information just as for any other content on its own website (negligent misrepresentation). Moffatt v. Air Canada, 2024 BCCRT 149, 14 February 2024, full text on CanLII.
• An empirical reason to regulate the output specifically: legal hallucinations in general-purpose models range between 58% and 88% on verifiable case-law queries, depending on the model. Dahl, Magesh, Suzgun, Ho, Large Legal Fictions, Journal of Legal Analysis 2024, arXiv.
• An experiment on output quality right in the professional press: a case-law analysis of adequate causation in the Austrian OGH was generated by ChatGPT 5 from a structured prompt, with the author merely checking the output and printing the prompt. Petr Bezouška, AI: Adekvátní příčinná souvislost v judikatuře rakouského OGH, Bulletin advokacie 12/2025, s. 24.

Statusconceptual · missing: a comparison of the regulatory regimes (SRA vs. ČAK) and the proposal part

Library

The sources I drew on, sorted by topic. Where a source comes from a licensed database, there is naturally no link.

AI in the judiciary and evidence

• Petr Zima, Ke dvěma aspektům využití umělé inteligence v oblasti práva, Právní rozhledy 20/2023, s. 712. AI as evidence (a neural-network business valuation that diverged from reality by a factor of 2.7) and the prediction of decisions, including the French ban on analytics of judges.
• Anežka Karpjáková, Zcela automatizované AI systémy a tvorba odůvodnění soudního rozhodnutí v civilním procesu, Právní rozhledy 22/2024, s. 737. The requirements for the statement of reasons under § 157 OSŘ against the black box, explainable AI, and candidate proceedings (deposits, annulment of instruments).
• Viktor Gazda, Předpověď recidivy umělou inteligencí a právo na soukromí, Právní rozhledy 8/2023, s. 283. The proportionality of algorithmic prediction, experience with the COMPAS system.
• Ondrej Kubík, Možnosti využitia umelej inteligencie pri hodnotení vierohodnosti svedeckých výpovedí, Trestněprávní revue 3/2025, s. 133. A Slovak perspective on AI in the forensic psychology of evidence.
• Libor Pavlíček, Kafkův Proces pohledem AI, Bulletin advokacie 1–2/2026, s. 46. An essay on three methods of knowledge in law and a proposal to build the Radbruch formula into the design of legal AI.
• Jakub Dohnal, Jak může umělá inteligence pomoci firemním právníkům a advokátním kancelářím?, epravo.cz, 5 August 2024. On the judgment of the Městský soud v Praze (Municipal Court in Prague) 10 A 99/2023: the unreliability of ChatGPT on questions of fact is a matter of common knowledge under § 121 o. s. ř., so its output is not a means of evidence.
• Jiří Hadaš, Umělá inteligence a právo. Patrně první rozsudek ve věci umělé inteligence, epravo.cz, 3 November 2023. On the judgment of the Spolkový ústavní soud (German Federal Constitutional Court) 1 BvR 1547/19 and 1 BvR 2634/20 of 16 February 2023: automated analysis of police data in Hesse and Hamburg as a disproportionate interference with informational self-determination; according to the court, adaptive systems intrude on fundamental rights especially deeply.
• Pavla Krejčí, Vexatorní podání jako neodvratitelný důsledek rozvoje AI, epravo.cz, 29 May 2025. AI has cut the cost of producing formally flawless filings. On the series from a single complainant with almost 860 proceedings (NSS 7 As 116/2023-11), on the procedural fine for a filing pursuing a clear abuse of rights (§ 44 odst. 1 s. ř. s.), and on the English model of civil restraint orders under the Senior Courts Act 1981, s. 42.
• Zbyněk Loebl, Jak bude vypadat online soud za 10 let?, epravo.cz, 28 February 2025. Online justice and the digital disadvantage of parties, the concept of a personal AI as a digital representative of a party in proceedings.
• State v. Loomis, 2016 WI 68, Nejvyšší soud Wisconsinu, 13 July 2016, full text. Using the COMPAS risk score at sentencing did not violate the right to a fair trial, but the score must not be the determining factor and the court must know the tool's limits (proprietary nature, group data, the risk of overestimating minorities). A primary source on algorithmic scoring in the judiciary.

Training models and intellectual property

• Eva Fialová, Legalita trénování umělé inteligence, Právní rozhledy 12/2026, s. 402. The TDM exceptions in Articles 3 and 4 of the DSM Directive, the judgments Kneschke v. LAION and GEMA v. OpenAI, memorisation as the limit, and the preliminary-ruling reference C-250/25 Like Company v. Google Ireland.
• Regina Yusupova, Generativní AI a vybrané otázky autorského práva, Duševní vlastnictví 3/2023, s. 3. Zarya of the Dawn, the approach of the US Copyright Office, and computer-generated works.
• Regina Yusupova, Generativní AI a vybrané otázky patentového práva, Duševní vlastnictví 4/2023, s. 3. DABUS and the refusal to recognise AI as an inventor across jurisdictions.
• Klára Mladá, (Ne)díla umělé inteligence a počiny dalších (ne)autorů, Obchodněprávní revue 2/2023, s. 122. Authorless output as movable property (§ 1045 and § 1074 of the Civil Code) and the conflict of platform terms with Czech law.
• Softwarové smlouvy, 2nd edition, 2025, the chapter Ochrana dat a umělá inteligence. Practical contractual clauses on training data and the exceptions in § 39c and § 39d of the Copyright Act.
• Zuzana Ottová, Jiří Dyčka, Umělá inteligence jako původce vynálezu ve světle případu DABUS, epravo.cz, 2 October 2024. EPO J 0008/20 concluding that a machine is not an inventor within the meaning of the European Patent Convention; the German BPatG 11 W (pat) 5/21 and the British UKSC 49/2023 both hold the inventor to be a natural person. The Městský soud v Praze follows on, č. j. 10 C 13/2023 of 11 October 2023: only a natural person can be an author, and a prompt is at most the theme of a work.
• Tim W. Dornis, Sebastian Stober, Generative AI Training and Copyright Law, arXiv, 2025. An interdisciplinary argument that generative training is not ordinary text and data mining and that memorisation creates a copyright problem independently of the TDM exceptions.
• The technical strand on memorisation: Carlini et al., Quantifying Memorization Across Neural Language Models, arXiv, 2022, memorisation grows with model size, data duplication, and context length. Freeman et al., Exploring Memorization and Copyright Violation in Frontier LLMs, arXiv, 2024, output filters against verbatim reproduction can be circumvented with a prompt.
• Primary US practice on authorship: the U.S. Copyright Office letter in Zarya of the Dawn, 21 February 2023, copyright.gov, images from Midjourney are not human authorship, while the text and arrangement remained protected. The registration guidance 88 Fed. Reg. 16190, 16 March 2023, govinfo.gov, non-protectable AI material is excluded from the claim and the assessment always turns on the degree of human control over the result.

Corporate decision-making with AI

• Lucie Josková, Rozhodování člena statutárního orgánu kapitálové společnosti, C. H. Beck 2024. The member of the body as "master of the process": AI may support decision-making, not replace it (§ 159 of the Civil Code), the delegation test of proper selection, understanding, and control, automation bias, and the inadmissibility of a decisive vote by AI under § 44 odst. 3 ZOK. The Czech counterpart to the Slovak chapter on company bodies in the textbook by Mesarčík, Gyurász et al.
• Tadeáš Diviš, Umělá inteligence a CorpTech v kontextu právní regulace, epravo.cz, 27 February 2026. Smart contracts and the requirement of mechanisms to intervene in automated performance, the standard of professional care in AI due diligence, and the impact of the forthcoming adaptation act on corporate compliance.

Regulation and public administration

• Adam Jareš, Umělá inteligence, automatizace a kódování v právu, Bulletin advokacie 1–2/2025, s. 33. The automation of administrative decision-making (parliamentary print No. 845), interactive templates of decrees, and the AI-legislator project of the Digitální a informační agentura (Digital and Information Agency).
• Regulace umělé inteligence, Právní rozhledy 22/2025, s. II. A legislative editorial on the draft Czech AI adaptation act from the MPO (Ministry of Industry and Trade).
• The action X.AI Holdings v. Commission, T-120/26 (brought 16 February 2026), Official Journal C/2026/2391. Eight pleas against fines under the Digital Services Act, among them the interpretation of the term "provider" and the single economic unit doctrine.
• Cobbe, Lee, Singh, Reviewable Automated Decision-Making, arXiv, FAccT 2021. Reviewability as a framework for the accountability of automated decision-making: explaining the model alone is not enough for regulatory oversight; what matters is record-keeping across the whole socio-technical process. Useful for the FRIA and for administrative AI.

AI in the workplace

• Anna Králíková, Odpovědnost zaměstnavatele a zaměstnance v souvislosti s využitím umělé inteligence, epravo.cz, 30 June 2026. A map of liability relationships: for an employee's use of AI the employer is liable to third parties, the employee is liable only to the employer, together with the deployer's obligations under Articles 4 and 26 of the AI Act.
• Michaela Vimpelová, Vliv umělé inteligence na zaměstnávání a pracovní právo, epravo.cz, 9 April 2024. High-risk systems for recruitment and employee assessment, the ban on emotion recognition in the workplace, and psychosocial risks as part of occupational health and safety.
• Sánchez-Monedero, Dencik, Edwards, What does it mean to solve the problem of discrimination in hiring?, arXiv, 2020. An analysis of hiring systems (HireVue, Pymetrics, Applied): vendors' claims of removing bias are little verified, and the EU and UK framework differs from the American one.
• The doctrine of less discriminatory alternatives: Black et al., The Legal Duty to Search for Less Discriminatory Algorithms, arXiv, 2024, and Laufer, Raghavan, Barocas, What Constitutes a Less Discriminatory Algorithm?, arXiv, 2025. For the same prediction problem there is often an equally accurate and less discriminatory model, but defining the search needs a standard of reasonableness.

Books in review

• Richard Susskind, How To Think About AI: A Guide For The Perplexed, Oxford University Press 2025. Reviewed by Kristýna Mlčáková, Revue pro právo a technologie 31/2025. The distinction between outcome-thinking and process-thinking, seven categories of risk including the risk of inaction.
• Bart van der Sloot, Regulating the Synthetic Society, Hart Publishing 2024. Reviewed by Damián Palašta, Revue pro právo a technologie 30/2024. Deepfakes, humanoid robots, AR and VR against the limits of the GDPR and the AI Act.

The profession in practice

• Pavel Kroupa, AI: Přichází doba umělé inteligence, Bulletin advokacie 10/2023, s. 3. An editorial by the ČAK board on the impact of AI on the legal profession and the first mention of the professional-body opinion.
• Olga Černá, Rozhovor s AI: Dejte mi práci, kterou dělat můžu, Komorní listy 1/2025, s. 23. An interview with a chatbot about automating the work of a judicial enforcement officer, instructive even in its limitations.
• Martin Maisner, Umělá inteligence a my ostatní, EPRAVO.CZ Magazine 2/2023. Because only an advocate may provide legal services, the advocate is fully liable even for errors taken over from AI that they relied on in good faith. At the time of writing the author was vice-chair of the ČAK.
• Jiří Matzner, Co když před námi začne umělá inteligence skrývat právo?, epravo.cz, 26 March 2025. The epistemic risk of generative models: the answers are probabilistic, so the model can effectively render lesser-known statutes invisible.
• Právní poradenství poháněné umělou inteligencí. Frank Bold spustil veřejné testování, epravo.cz, 9 May 2023. The first public test of AI legal advice by a Czech law firm (GPT-4 over the database of its own advice service), an early data point on the line between legal information and service from H1 and H2.
• Eliza Mik, Caveat Lector: Large Language Models in Legal Practice, arXiv, 2024. The risk of over-reliance: the fluency and persuasiveness of the output says nothing about its truth; the models work with the distribution of words, not with verified facts.