Guides
AI Literacy for Law Firms: What to Teach and What Article 4 Expects
AI literacy for law firms under Article 4 of the EU AI Act: an obligation of effort, no certificates, keep a record. A curriculum by role and a training record.
On this page
Every law firm in the EU whose staff use ChatGPT, Copilot or Claude has had an AI-literacy duty under Article 4 of the AI Act since 2 February 2025. Since 2 August 2026 national authorities supervise it, and the Digital Omnibus has reworded it into an obligation of effort: take measures, keep a record, no certificates. Outside the EU the same duty arrives through competence rules, warning notices and CLE mandates. This guide sets out what to teach, to whom, and how to prove it.
What Article 4 requires after the Omnibus
Article 4 was the first operative provision of the AI Act to bite, and the one the Digital Omnibus on AI touched without deferring. As of September 2026:
| Element | Before the Omnibus | After the Omnibus (reported as Regulation (EU) 2026/1744, in force 27 July 2026) |
|---|---|---|
| Core wording | Providers and deployers "shall take measures to ensure, to their best extent, a sufficient level of AI literacy" | "take measures to support the development of AI literacy" — an obligation of effort |
| Who is covered | Any provider or deployer; the Commission FAQ names organisations whose staff use tools like ChatGPT | Unchanged |
| Applicable since | 2 February 2025 | Unchanged; not deferred with Annex III high-risk duties (now 2 December 2027) |
| Supervision | No designated supervisor until August 2026 | National market-surveillance authorities since 2 August 2026 |
| Fines | No dedicated EU tier | Still none; national, proportionate penalties |
| Evidence expected | Internal training record | Unchanged; no certificates or tests |
| Support | — | New paragraphs on Commission and Member State support, Board recommendations, and a living repository of 40+ practices |
The Commission's AI-literacy FAQ, updated 27 July 2026, is the practical text. Three points matter for a firm. First, the duty scales with context: what a firm using a general chatbot for drafting needs is not what a firm deploying an AI-assisted judicial tool would need. Second, no certificate or test is required, but an internal record of who was trained, when and on what is expected. Third, the absence of an EU fine tier does not make the duty toothless. Latham & Watkins' analysis, cited in the techniques research for this site, notes that a documented training gap is an aggravating factor in any other AI Act or professional-conduct proceeding and a director-liability exposure. The Article 4 explainer has the legislative history and the Omnibus page covers what else moved. In Austria, the KI-Servicestelle at RTR, established under § 20c KOG, runs a "KI-Kompetenz" section that is the natural first stop for an Austrian firm; the deployer entry explains why a law firm is one.
The same duty by other names outside the EU
Nothing in the US or UK is called an AI-literacy obligation, and all of the following amount to one:
| Jurisdiction | Instrument | What it expects | Date |
|---|---|---|---|
| USA (40 states) | ABA Model Rule 1.1, Comment 8 | Keep abreast of "the benefits and risks associated with relevant technology" | Adopted progressively; 40 states as of 2026 |
| USA | ABA Formal Opinion 512 | Competence, confidentiality, supervision of AI use; understanding the tool's capabilities and limits | 29 July 2024 |
| Texas | Opinion 705 | "A reasonable and current understanding of the technology"; no blind reliance | February 2025 |
| California | Practical Guidance covering generative and agentic AI | "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified" | 14 May 2026 |
| New Jersey | Supreme Court CLE rule | One technology credit every two years | From 1 January 2027 |
| Florida / New York / North Carolina | CLE rules | 3 technology hours per cycle / 1 cybersecurity hour a year / 1 hour a year | In force |
| England and Wales | SRA Warning Notice on misuse of AI | Understand a tool "well enough to review its output critically"; 42 misuse reports July 2025–July 2026 | 17 August 2026 |
| England and Wales | Ayinde v Haringey [2025] EWHC 1383 (Admin) | Heads of chambers and managing partners are responsible for their lawyers' AI use | 6 June 2025 |
| Austria | ÖRAK practice rules and § 40 Abs 3 RL-BA 2015 | EU servers, disclose to clients, lawyer keeps final responsibility, no client data for training | 2026 |
| Germany | BRAK Hinweise zum Einsatz von KI | Tools that store or forward client data for training breach § 43a Abs. 2 BRAO | December 2024 |
| Europe | CCBE Guide (Oct 2025) and Technical Guide (adopted 27 March 2026) | Inform clients; understand deployment models from on-premise to SaaS, open-weight models and agentic AI | 2025–2026 |
Courts have started prescribing education as a sanction, which is the clearest sign of what they think is missing. Beus Gilbert PLLC v. BYU (D. Utah, 9 September 2026) imposed $3,000 plus two AI-ethics CLE courses; Booker v. Kroger (N.D. Ga., 28 August 2026) imposed $8,000 plus CLE. The SRA notice page, the ABA 512 page and the state opinions page have the texts; the Austria and Germany page covers the bar guidance in German.
Why an hour on "AI ethics" does not meet it
The gap between what firms provide and what the duty expects is measurable. As of 2026:
- 54% of firms in the 8am Legal Industry Report (1,300+ respondents) offer no AI training and have no plans to; 43% have no policy; only 9% have written, enforced policies.
- 34% of professionals in Thomson Reuters' 2026 survey use unsanctioned "shadow AI"; organisations with a named AI strategy report AI meeting expectations 66% of the time against 22% without one.
- Only 9% of the 1,553 lawyers in the State Bar of Texas 2026 AI survey are "extremely confident" evaluating AI-generated documents; 64% name confidentiality as their top concern.
- 45.5% of the 112 federal judges surveyed for the Sedona Conference Journal (published March 2026) had received no court-provided training, while 61.6% use AI tools.
- A single-source analysis by The Legal Stack (July 2026, treat cautiously) counted 400+ approved AI-related CLE hours, of which 78% are primers, ethics or regulation; 67% of practitioners said CLE did not cover the tools they use, and fewer than three courses addressed agentic workflows.
Policies without training do not protect anyone. Butler Snow had an AI policy from June 2023; in Johnson v. Dunn (N.D. Ala., 23 July 2025) three of its lawyers were publicly reprimanded and removed from the case over ChatGPT citations. Morgan & Morgan's lawyers in Wadsworth v. Walmart (D. Wyo., 24 February 2025) were using the firm's own internal platform. Scientific American's May 2026 review of why warnings fail points to automation bias and "cognitive surrender": people reverse correct judgments when an AI disagrees, and awareness campaigns alone show only modest effect. Literacy, in other words, has to be practised, not announced. The firm AI policy playbook is the companion to this page; the policy names the rules, the training makes them reflexes.
A curriculum by role
The FAQ's "context-dependent" standard means one course for everyone is the wrong design. The table below is a defensible baseline for a firm running a frontier assistant on a business tier plus at least one legal tool.
| Role | Must understand | Must be able to do | Hours (year 1) | Assessment evidence |
|---|---|---|---|---|
| Partners and supervising lawyers | Which tool tiers train on inputs; ABA 512 supervision and fee rules; disclosure duties to clients and courts; what an agent can do without confirmation | Review AI-assisted work critically; brief a junior on verification; explain AI use to a client; decide fixed versus hourly pricing for AI-suited work | 4–6 | Signed review of one AI-assisted deliverable; short client-explanation exercise |
| Associates and trainees | The three-check verification protocol; prompt structure (role, task, sources, format, constraints); Projects and playbooks; jurisdiction and cutoff limits | Run a research task by hand and then with the tool and diff the two; verify every authority; produce a first draft to a playbook; log tool and version | 8–10 | Verified research memo with log; drafting exercise marked by a senior; oral questions on authorities |
| Paralegals and business services | Confidentiality rules; which accounts are enterprise and which are consumer; data-training toggles; what never goes into a public tool | Use the approved tools for intake, summaries and admin; recognise a request that needs escalation | 2–3 | Settings walkthrough completed; scenario quiz |
| IT, risk and knowledge | Vendor data policies (retention, ZDR eligibility, sub-processors); agent permissions, allowlists and audit logs; Article 50 transparency for client-facing bots | Configure enterprise tiers; run vendor due diligence; switch on and read audit trails; maintain the prompt library | 6–8 | Vendor review completed; audit-log test; annual settings attestation |
| Everyone | Why models hallucinate; the sanctions record; the firm's traffic-light rules | Spot a fabricated citation; find the firm's policy and the escalation route | 1 | Attendance and a five-question check |
Two design notes. Verification and confidentiality appear in every row because those are the two failure modes that produce sanctions and data incidents; the verification protocol and the privacy settings walkthrough are the two artefacts to hand out on day one. And associates get the most hours because they are the ones whose training is otherwise disappearing; the junior lawyers guide explains why "checking the AI" is not a curriculum.
The six sessions
A session plan that covers the curriculum in a term, for a firm of any size, each session 60–90 minutes with a homework task:
- How the models work and how to prompt them. Tokens, context windows, knowledge cutoffs (Opus 5's is May 2026; GPT-6 Astra's is 30 April 2026), why the model guesses rather than abstains. Homework: rewrite three of your own prompts using the structured prompt method.
- Research and verification. The hallucination record (2,041 decisions in Damien Charlotin's database as of 14 September 2026, 811 involving lawyers), the three checks, the research log, why "ask the AI to check itself" is an aggravating factor. Homework: verify a ten-citation memo and log it.
- Drafting and document analysis. Playbook-driven review, tracked changes in Word, first drafts from precedents. Homework: a clause redline against the firm playbook.
- Confidentiality, data and privilege. Consumer versus enterprise tiers, retention periods, the United States v. Heppner (S.D.N.Y., February 2026) ruling that consumer-chatbot memos were not privileged, the UK Upper Tribunal's warning that pasting a client letter into a public tool puts it "in the public domain". Homework: audit your own accounts and settings.
- Agents and automation. What an agent is, approval gates, allowlists, what never to give an agent. Homework: map one workflow that could run on a gate. The agent security guide is the reading.
- Rules, disclosure and pricing. Article 4 and Article 50, ABA 512, the SRA notice, ÖRAK's four rules, billing for AI work. Homework: draft the engagement-letter sentence.
For firms that would rather buy the sessions than build them, the AI Lab for Lawyers course follows the same arc in four live two-hour sessions plus about two hours of homework, with recordings and a certificate of completion that slots into the training record; it is taught by Dr. Niklas Schmidt, a Wolf Theiss partner, and has had 1,000+ participants from 50+ firms. Whatever you use, the sessions should be hands-on in the firm's own tools, on the firm's own precedents, with client data replaced by synthetic examples.
The training record
The record is the only artefact Article 4 asks for, and it is the one thing an authority, an insurer or a court will ask to see. A minimal template:
| Field | Example entry |
|---|---|
| Person and role | A. Müller, associate, corporate |
| Session and content | Session 2: research and verification (three-check protocol, research log) |
| Tools covered | Claude Team (Opus 5), CoCounsel Legal, Westlaw KeyCite |
| Date and duration | 14 September 2026, 90 minutes |
| Trainer and format | Internal, live, hands-on |
| Assessment | Ten-citation verification exercise, passed; log reviewed by supervising partner |
| Refresher due | On next major model change, and no later than 14 March 2027 |
| Policy version acknowledged | Firm AI policy v3 (August 2026) |
Refresher triggers should be written into the record, because the tools change faster than annual cycles. GPT-5.5 retires from ChatGPT on 14 October 2026. Claude Opus 5 "verifies its own work without being told to", which means prompts written for older models that add a verification step now cause over-verification. Astra for Law launched on 17 September 2026 and has almost no independent evaluation. Each is a reason to run a 30-minute update, log it, and move on. Malpractice carriers have not excluded AI as of May 2026, but application forms already ask whether the firm lets AI draft documents; a training record is the answer to the follow-up question.
A 30-day rollout
- Week 1. Adopt or update the policy, name an owner, and list every AI tool in use including the unsanctioned ones (34% of your staff are using one). Decide the approved tiers and switch off training on any consumer accounts that remain.
- Week 2. Run sessions 1 and 4 for everyone; they are the two that reduce risk fastest. Start the training record the same day.
- Week 3. Run sessions 2 and 3 for fee-earners; hand out the verification protocol and the prompt library.
- Week 4. Run sessions 5 and 6 for supervising lawyers and IT; set the refresher triggers; brief clients whose outside-counsel guidelines ask about AI training, which increasingly they do.
At the end of the month the firm has a record that satisfies the Commission's FAQ, the SRA's "critically review" standard and Comment 8, and a workforce that has practised the two habits that keep lawyers out of Charlotin's database. That is the whole of the obligation, and it is a great deal cheaper than one sanction.
Frequently asked questions
Does Article 4 of the EU AI Act apply to law firms?
Yes. The European Commission's FAQ, updated 27 July 2026, says Article 4 covers any organisation whose staff use tools like ChatGPT, which includes a law firm as a deployer. The duty has applied since 2 February 2025. After the Digital Omnibus it reads as a duty to take measures to support the development of AI literacy, an obligation of effort rather than a guaranteed result, and it was not deferred with the high-risk rules.
Do lawyers need an AI certificate to comply with Article 4?
No. The Commission's FAQ states that no certificates or knowledge tests are required. What it does expect is an internal training record showing who was trained, when, on what and how it was assessed. National authorities have supervised the duty since 2 August 2026, and there is no dedicated EU fine tier for Article 4; penalties are national and must be proportionate.
What AI training is required for lawyers in the US and UK?
There is no stand-alone AI-ethics CLE credit in any US jurisdiction as of September 2026, but 40 states have adopted the Comment 8 duty of technology competence, Florida requires three technology hours per cycle, New York one cybersecurity hour a year, and New Jersey one technology credit every two years from 1 January 2027. In England and Wales the SRA's 17 August 2026 warning notice expects solicitors to understand an AI tool well enough to review its output critically.
How many hours of AI training does a law firm need?
There is no statutory number. A defensible baseline for fee-earners is six to eight hours of hands-on training in the first year, split across fundamentals, verification, confidentiality, drafting and agents, plus a short refresher whenever a major model or tool changes. Business-services staff need two to three hours focused on confidentiality and data settings. Record the hours and the assessment.
Put this into practice
AI Lab for Lawyers: a live, hands-on course for practising lawyers
4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.
Related reading
Regulation & Ethics
EU AI Act Article 4: The AI Literacy Duty for Law Firms, Explained
Article 4 of the AI Act requires every firm whose staff use AI to support their AI literacy. In force since 2 February 2025, supervised nationally since August 2026.
7 min read
Regulation & Ethics
The Digital Omnibus on AI: What Changed in the AI Act and What Did Not
The Digital Omnibus AI Act amendment, Regulation (EU) 2026/1744, deferred high-risk duties to December 2027 but left transparency, literacy and the prohibitions in force.
6 min read
Regulation & Ethics
The SRA's 2026 Warning Notice on AI Misuse: What UK Solicitors Must Know
The SRA AI warning notice of 17 August 2026 followed 42 misuse reports in a year. Solicitors must understand a tool well enough to review its output critically.
8 min read
Playbooks
Write a Law Firm AI Policy in a Day: A Working Template
A working law firm AI policy template with ten clauses, commentary on each, EU and US variants and an hour-by-hour rollout plan that gets it signed in one day.
12 min read
Guides
Junior Lawyers and AI: Training, Hiring and the Missing Rung
Junior lawyers and AI in 2026: Big Law hired 538 fewer graduates, 2% of UK juniors say AI helps them learn, and firms are rebuilding training with AI seats.
11 min read
Regulation & Ethics
ABA Formal Opinion 512 Explained: What US Lawyers Must Do When Using AI
ABA Formal Opinion 512 applies six Model Rules to generative AI: competence, confidentiality, communication, candour, supervision and fees. What each duty requires in 2026.
8 min read