California's SB 574 as an AI self-check for German and European law firms

Key facts

• What: SB 574 is a California statute on how lawyers (and arbitrators) may use generative AI. Its core sentence: “An attorney shall not delegate the practice of law to generative artificial intelligence.”
• Signed: 30 September 2026 (Chapter 858, Statutes of 2026). In force: 1 January 2027.
• Where codified: new Business and Professions Code § 6068.1, an amendment to Code of Civil Procedure § 128.7 on citation checks, and Code of Civil Procedure § 1282.1 for arbitrators.
• Four duties for lawyers: no delegation of legal work to generative AI; verification of outputs, correction of hallucinated output and personally verified citations; confidential, personal identifying and other nonpublic information only in access-restricted systems; disclosure of generative AI use to the court.
• Germany: the law does not apply. Its four duties still make a good self-check: could a firm prove its AI use with documents, or only vouch for it in words?

On 30 September 2026 the Governor of California signed SB 574, one of the first statutes anywhere to set out how lawyers must work with generative AI. Bloomberg Law called it a first-of-its-kind bill. It takes effect on 1 January 2027.

SB 574 is not German law, but it makes a useful benchmark. Its four duties are the same questions that clients, courts and partners in Germany and across the DACH region ask as soon as AI comes up. Taken as a self-check, each duty raises one question: what could a firm document today, and where is the evidence missing? The piece is written for partners, legal operations, and knowledge and innovation leads at law firms, and for in-house legal departments that instruct outside counsel. It is neither advice on US law nor on German professional rules in an individual case.

What does SB 574 require from lawyers?

The bill text is published on the California Legislature’s official site, marked “Approved by Governor September 30, 2026.” Its core sentence is short: “An attorney shall not delegate the practice of law to generative artificial intelligence.” Lawyers may use AI, but they may not hand it the legal work. Responsibility for the result stays with the person who signs.

The duties of care follow from there. Lawyers must take reasonable steps to verify AI outputs, expressly including case and statutory citations, and correct erroneous or hallucinated output. Separately, under Code of Civil Procedure § 128.7 a court filing may not contain any citation that the responsible attorney has not personally verified, whether or not it came from AI. Confidential, personal identifying and other nonpublic information may only go into systems where access is restricted to the lawyer and authorised persons bound by confidentiality. Lawyers who use generative AI must tell the court; the bill requires them to “disclose the use of generative artificial intelligence to the court for all documents submitted to the court …” Holland & Knight points out that the wording leaves open whether disclosure is needed only for filings where AI helped, or more broadly.

The bill title also names arbitrators, judicial officers and alternative dispute resolution providers. For arbitrators, the new Code of Civil Procedure § 1282.1 applies: they may not delegate any part of their decision-making to a generative AI tool, and may not rely on AI-generated information from outside the record without first disclosing it to the parties. The bill’s new complaint process for ADR providers, by contrast, concerns breaches of ethics standards and has nothing to do with AI. For firms with arbitration practices, it is therefore worth taking a closer look at this part.

Since the first briefs with invented citations, nobody has needed a statute to know that AI-generated authorities must be checked. What is new is that a legislature now regulates the way of working itself. From 2027, a lawyer filing in California has to produce a sound result and also be able to say how it was made. In many firms today, that second question cannot be answered with the evidence to back it up.

Why the question is reaching Germany anyway

Markus Hartung, managing partner of the law firm Chevalier and director of the Bucerius Center on the Legal Profession from 2010 to 2019, wrote about the bill in his NJW column “B-Note”, published by beck-aktuell on 18 September 2026, before the Governor signed it. Nobody knows exactly what the delegation ban covers, he notes, because California has never defined “practice of law”; he calls that a sore point. He still finds the approach remarkable. Translated from the German: “This regulates how lawyers have to work, not just what they must refrain from. […] We have nothing like it.” Responsibility stays where it was, though. LTO sums up an article by Christiane Bierekoven, who chairs the IT law working group of the German Bar Association (DAV), in the Anwaltsblatt like this: “Die anwaltliche Haftung besteht auch für KI-Ergebnisse.” In English: a lawyer’s liability extends to AI results.

What already applies to German firms

Germany has no SB 574, but AI use in a law firm is not unregulated either.

• EU AI Act, Article 4: the AI literacy duty has applied since 2 February 2025 and was recast by Regulation (EU) 2026/1744 (in force since 27 July 2026). Organisations that deploy AI systems, law firms included, must now take measures to support the development of their staff’s AI literacy. They do not have to guarantee any particular level of competence.
• German Federal Bar (BRAK): on 8 January 2025 the BRAK published guidance on the use of AI in law firms (“Hinweise zum Einsatz von künstlicher Intelligenz”, as at December 2024). It is non-binding and covers, among other things, duties of care when using AI, confidentiality and transparency duties.
• Professional rules: the duty of confidentiality (§ 43a(2) BRAO) and criminal liability for disclosing client secrets (§ 203 StGB) apply to every AI tool that processes client data. The rules for engaging service providers (§ 43e BRAO) come in as soon as an external provider gains access to client secrets; according to the BRAK guidance, the mere possibility of access is enough. The GDPR applies alongside them.

None of these rules prescribes how a firm should build AI into a specific process. In California, SB 574 at least sets ground rules for the way of working, but leaves open what exactly the delegation ban covers. § 6068.1 contains no liability or sanction provision of its own; breaches of the citation rule fall under the existing sanctions in Code of Civil Procedure § 128.7.

Austria: courts are already receiving AI-written filings

Austrian courts are already dealing with the other end of the problem: AI-written submissions that have actually been filed. In a decision of 9 June 2026 (RV/6100033/2026), reported in KPMG’s Tax News of 23 September 2026, the Federal Fiscal Court (BFG) held that submissions produced by so-called artificial intelligence are not suitable grounds in an appeal and need no further discussion. In a guest article in Die Presse on 15 September 2026, tax law professor Friedrich Fraberger cites a decision of the Supreme Court of Justice (OGH, 14 Os 95/25i) and a further BFG decision (RV/7101070/2026). He spells out the practical cost when clients bring their own AI drafts: checking the AI’s output is often far more work than writing the brief or the advice from scratch. We are hearing similar reports from the German market.

Clients want proof

Clients are the most direct source of pressure. The Legal Tech Monitor 2026 by the German Legal Tech Association (Legal Tech Verband Deutschland), based on around 360 responses, found that almost 40 percent of in-house respondents already ask outside counsel for evidence of their legal tech and AI use. 73 percent said they would consolidate work with firms that can show cost savings from AI. Both figures are about proof of what a firm actually does.

The evidence has to show process flows and the use of AI clearly. An internal AI policy is far from an adequate answer.

The four duties as a self-check for your firm

For each duty, the test question is: could the firm produce a document as evidence, or only a description from memory? (This is not about complying with California law.) The test works best on a single recurring process, such as client intake, a standard contract review, or preparing filings in high-volume litigation.

1. No delegation of legal work

The underlying question is where in a process AI does the work and where a person decides. The evidence is a process description that sorts every step into one of three types: a fixed rule, AI-assisted work, or a legal decision by a lawyer. It helps to mark the reverse too, meaning steps where no AI is involved, such as deadlines, jurisdiction or fees. The typical gap is that AI use happens in individual conversations with a language model that no one ever looks at again. In a process mapped in e! by Lexemo, each step shows whether AI is involved, and every run leaves an audit trail.

2. Verifying citations and outputs

Most firms check, but few record that they checked. The evidence is a review note showing who checked which output, when, against which source, and what was corrected. The most reliable form is a fixed sign-off step in the process, where a named person has to approve before a document leaves the firm. Fraberger’s point belongs here: review is work, and it belongs in the pricing, especially when the draft came from the client.

3. Confidentiality

This is where a firm can get the most on paper, and quickly. Start with a list of approved AI tools, stating where data is processed, whether inputs may be used for training, and which contracts apply, including data processing agreements and, as soon as the provider gains access to client secrets, the requirements of § 43e BRAO for engaging service providers. Add a rule on which categories of data may go into which tool. The usual gap is private or free accounts that individual colleagues use on the side and that appear on no list.

4. Disclosure

Germany has no general duty to disclose AI use to the courts. The question can still come up from a client, from opposing counsel, or in matters with a US connection. A simple test is whether the firm could say, within one working day, whether and where AI was involved in a given filing. A short note on the file is often enough, and the same note answers the client question from the Legal Tech Monitor.

Run honestly, the test rarely ends with “we need another policy”. What is usually missing is a description of the firm’s own processes, and one process is enough to start.

Frequently Asked Questions

What does California’s SB 574 regulate?

SB 574 is a California statute, signed on 30 September 2026, that regulates how lawyers must work with generative AI. Its central sentence reads: “An attorney shall not delegate the practice of law to generative artificial intelligence.” Lawyers must also take reasonable steps to verify AI outputs and correct erroneous or hallucinated output, and every citation in a court filing must have been personally verified by the responsible attorney. Confidential, personal identifying and other nonpublic information may only go into access-restricted systems, and the use of generative AI must be disclosed to the court for documents submitted to it. Arbitrators may not delegate any part of their decision-making to generative AI, and may rely on AI-generated information from outside the record only after disclosing it to the parties.

When does SB 574 take effect?

SB 574 was enacted as Chapter 858, Statutes of 2026, and applies from 1 January 2027. It adds § 6068.1 to the California Business and Professions Code, amends Code of Civil Procedure § 128.7 (citations) and governs arbitrators’ use of AI in Code of Civil Procedure § 1282.1. How far the law reaches in a particular case, for instance in matters involving foreign firms, is a question of US law and should be clarified with US counsel.

Does SB 574 apply to German law firms?

No, not for work before German courts. SB 574 is addressed to the California bar. Whether a German firm with US matters or proceedings in California is affected in a specific case can only be answered with US legal advice. The four duties are still useful in Germany as a self-check, because they match the questions clients and courts ask here too.

Which rules on AI use apply to lawyers in Germany?

There is no AI-specific statute for lawyers comparable to SB 574. Under Article 4 of the EU AI Act, as amended by Regulation (EU) 2026/1744 (in force since 27 July 2026), deployers of AI, including law firms, must take measures to support the development of AI literacy among their staff; they do not have to guarantee any particular level of competence. The German Federal Bar (BRAK) published non-binding guidance on AI use in law firms on 8 January 2025 (as at December 2024). The general professional rules also apply: confidentiality (§ 43a(2) BRAO), the protection of client secrets under § 203 StGB and the GDPR, plus § 43e BRAO as soon as an external service provider gains access to client secrets.

Do lawyers in Germany have to disclose their use of AI?

There is no general statutory duty in Germany to disclose AI use to the courts. Clients, opposing counsel or courts in cross-border matters may still ask, and in-house legal departments increasingly do: almost 40 percent ask outside counsel for evidence of their AI use, according to the Legal Tech Monitor 2026. Towards clients, the BRAK guidance finds that the German professional rules (BRAO and BORA) in principle impose no duty to tell clients about AI use. Such duties can arise from the client contract or from unfair competition law (UWG), and the BRAK recommends being transparent about AI and, where in doubt, agreeing it contractually. A short note on the file recording whether and where AI was involved answers most of these questions.

Are lawyers liable for mistakes made by AI?

Yes. Using an AI tool does not shift responsibility for the work product. LTO sums up an Anwaltsblatt article by Christiane Bierekoven, chair of the German Bar Association’s IT law working group, this way: “Die anwaltliche Haftung besteht auch für KI-Ergebnisse” (a lawyer’s liability extends to AI results). A documented review step with a named person is the most reliable way to show that outputs were checked.

How can a law firm show clients how it uses AI?

Evidence convinces when it refers to concrete processes: a description of which steps are rule-based, which are AI-assisted and which are decided by a lawyer; a list of approved tools with details on data processing and training; a documented review or sign-off step with a named person; and a note on the file showing whether and where AI was involved. A general AI policy without that link to actual work rarely answers the question.

Sources

• California Legislature: SB 574 bill text
• Bloomberg Law, 30 Sept 2026: Newsom signs first-of-its-kind bill on lawyer, arbitrator AI use
• Holland & Knight, Oct 2026: California enacts rules governing lawyers’ use of generative AI
• beck-aktuell, 18 Sept 2026: Hartung, NJW column “B-Note”
• beck-aktuell, 17 Sept 2026: Legal Tech Monitor 2026
• KPMG Austria Tax News, 23 Sept 2026: BFG on AI-generated appeals
• Die Presse, 15 Sept 2026: Fraberger on AI briefs in tax matters
• LTO: Anwaltsblatt on legal risks of AI (DAV)
• BRAK, 8 Jan 2025 (as at Dec 2024): Guidance on the use of AI in law firms
• EU AI Act (Regulation (EU) 2024/1689), Article 4
• Regulation (EU) 2026/1744 (AI omnibus), OJ L of 24 July 2026: recast of Article 4 of the EU AI Act

This article gives a general overview as at 8 October 2026 and does not replace advice on an individual case.