The Legal AI 'Governing Law Mismatch' Benchmarking Report 2026: How Often AI Contract Review Tools Apply the Wrong Jurisdictional Standard to Enforceability Questions — and Which Clause Categories Are Most Exposed

Research briefing — a review of public sources. The Legal Stack did not conduct a survey, benchmark or interviews for this piece.

Updated October 2026. An earlier version described how AI contract tools "perform" on governing-law questions without any test data. This version relies only on published benchmarks, statutes and case law, and labels our analysis as analysis.

What is and isn't known

No independent, published benchmark measures how often AI contract review tools apply the wrong jurisdiction's rules to enforceability questions. Vendors rarely say in public which legal regimes their risk flags have been validated against.

There are two kinds of evidence we can rely on:

• Independent benchmarks that show AI performance falls when questions span more than one jurisdiction.
• Clear doctrinal divergences in high-traffic clause categories, where applying the wrong standard produces a wrong answer that someone can check.

This briefing pairs the two and sets out procurement steps. Where we draw conclusions, they are our analysis.

The benchmark signal: multi-jurisdiction questions are harder

The most relevant data comes from Vals AI's October 2025 legal research benchmark, as reported by LawSites. AI tools scored 79–81% overall, against a lawyer baseline of 71%. But all the systems struggled with multi-jurisdictional questions that required synthesizing the law of several states. On those, performance fell by an average of 11 points compared with single-state questions.

Thomson Reuters, LexisNexis and vLex did not take part, so the results cover Alexi, Counsel Stack, Midpage and ChatGPT.

Earlier, in May 2024, Stanford RegLab and HAI tested legal research tools on more than 200 queries. They found hallucination rates above 17% for Lexis+ AI and Ask Practical Law AI, and above 34% for Westlaw AI-Assisted Research.

Neither study tested contract review against a contract's governing-law clause. Still, both point the same way: answers that depend on jurisdiction are a known weak spot.

Where the law diverges: four clause categories

Each category below is one where applying the wrong standard gives an answer that can be checked against the source. These are the places to test a tool.

Non-competes: California vs New York (and the FTC rule that isn't)

California. California Business and Professions Code §16600 voids "every contract by which anyone is restrained from engaging in a lawful profession," subject to exceptions such as the sale of a business. SB 699 was signed on 1 September 2023 and took effect on 1 January 2024. It added §16600.5, which makes void non-competes unenforceable "regardless of where and when the contract was signed." It also gives employees a private right of action. Fenwick reported these details, and noted a companion bill, AB 1076, that would codify Edwards v. Arthur Andersen and require written notice to employees by 14 February 2024.

New York. New York took a different path. Governor Hochul vetoed a broad non-compete ban on 22 December 2023, as A&O Shearman reported. Non-competes therefore remain enforceable in New York if they are reasonable and protect a legitimate business interest.

The FTC rule. There is no federal ban to fall back on. On 5 September 2025, the FTC announced that it would no longer pursue its appeals defending the nationwide non-compete rule in Ryan, LLC v. FTC and Properties of the Villages v. FTC. Steptoe reported that the FTC would instead take a targeted approach, including a 4 September 2025 complaint and proposed consent order against Gateway Services.

Why this matters for tools. A tool that flags a reasonable non-compete in a New York-governed agreement as void, by applying California law or a federal rule that never took effect, is wrong in a way that can be checked. The reverse error also matters. Because of §16600.5, choosing another state's law may not protect a non-compete covering a California-based employee.

Liquidated damages and penalties: England vs the U.S.

England. In Cavendish Square Holding BV v Makdessi; ParkingEye Ltd v Beavis [2015] UKSC 67, decided on 4 November 2015, the UK Supreme Court restated the penalty rule. Herbert Smith Freehills summarized the decision. A clause is penal only if it imposes a detriment "out of all proportion to the innocent party's legitimate interest" in performance. The traditional "genuine pre-estimate of loss" approach is no longer the test. The rule applies only to secondary obligations triggered by breach. A conditional primary obligation cannot be a penalty. For negotiated contracts between properly advised parties, there is a "strong initial presumption" that the parties are the best judges of what is legitimate.

California. California Civil Code §1671(b) provides that a liquidated damages provision "is valid unless the party seeking to invalidate the provision establishes that the provision was unreasonable under the circumstances existing at the time the contract was made." Consumer contracts follow a stricter rule under §1671(c)–(d).

Why this matters for tools. A tool that tests an English-law clause only against whether it was a pre-estimate of loss is applying the wrong test. So is a tool that treats a California commercial liquidated damages clause as presumptively suspect.

Indemnities, caps and arbitration

We found no comparably clear, single-source divergence for indemnity caps or arbitration carve-outs that we could verify for this briefing. These clauses still need jurisdiction-aware review. Buyers should include them in testing, using standards their own counsel confirms for each governing law in the portfolio.

Analysis: why tools may embed a default jurisdiction

The following explanation is our analysis, not a tested finding.

General-purpose models are trained on a corpus in which U.S. sources, and some U.S. jurisdictions in particular, are heavily represented. Unless a product explicitly conditions its review on the governing-law clause, a "standard" risk flag may reflect whatever rule appears most often in that corpus. The Vals multi-jurisdiction drop is consistent with this explanation, though it does not prove it.

Product documentation rarely says which jurisdictions a playbook or risk model was validated against. So buyers usually cannot tell from the outside whether jurisdiction is a real input to the tool or an unstated assumption.

Procurement steps for GCs and legal ops

These recommendations are analysis.

1. Ask for a jurisdiction coverage statement. Ask the vendor which governing-law regimes its risk flags and suggested language have been validated against, at the level of individual clause types. Make the answer a contractual representation, or at least a documented deliverable.
2. Run a "same contract, different law" test. Take one agreement and change only the governing-law clause, for example New York, California and England and Wales. Run each version through the tool. Check:
3. whether the non-compete flag follows §16600/§16600.5 for California and reasonableness for New York
4. whether the liquidated damages analysis applies Cavendish for English law and §1671(b) for California commercial contracts
5. Include an employee-location twist. Add a version where a New York-law non-compete covers a California-based employee. A good tool should flag §16600.5.
6. Check for stale law. Ask whether the tool still describes the FTC non-compete rule as pending or in force, or applies the pre-2015 English penalty test. Either one is a quick signal of how current the tool is.
7. Weight multi-jurisdiction performance in scoring. Given the 11-point average drop in the Vals study, score cross-border and multi-state test cases separately and give them more weight if your portfolio is international.
8. Require change notices. Ask to be told when the underlying model or clause library changes, and re-run your test set when it does.

Bottom line

There is no public number for how often AI contract tools apply the wrong governing law. What the public record shows:

• Multi-jurisdiction questions are measurably harder for AI.
• Several common clause types diverge sharply between jurisdictions, and the correct rule can be checked against statutes and leading cases.

That makes the risk testable. A two-hour "same contract, different law" exercise, scored against the authorities above, will tell a buyer more than any vendor slide.

Sources

• LawSites, "Vals AI's latest benchmark finds legal and general AI now outperform lawyers in legal research accuracy" (23 Oct 2025): https://www.lawnext.com/2025/10/vals-ais-latest-benchmark-finds-legal-and-general-ai-now-outperform-lawyers-in-legal-research-accuracy.html
• Stanford HAI, "AI on Trial: Legal Models Hallucinate in 1 out of 6 (or More) Benchmarking Queries" (23 May 2024, updated 30 May 2024): https://hai.stanford.edu/news/ai-trial-legal-models-hallucinate-1-out-6-or-more-benchmarking-queries
• Fenwick, "California Expands Prohibition Against Non-Competes" (2023): https://www.fenwick.com/insights/publications/california-expands-prohibition-against-non-competes
• A&O Shearman, "New York State non-competes ban vetoed by Governor" (Dec 2023): https://www.aoshearman.com/en/insights/new-york-state-non-competes-ban-vetoed-by-governor
• Steptoe, "FTC abandons national noncompete rule, but employer compliance risks remain" (2025): https://www.steptoe.com/en/news-publications/stepahead-antitrust-and-competition-insights/ftc-abandons-national-noncompete-rule-but-employer-compliance-risks-remain.html
• Herbert Smith Freehills, "Supreme Court rewrites English law rule on penalties" (Nov 2015): https://www.hsfkramer.com/notes/litigation/2015-11/supreme-court-rewrites-english-law-rule-on-penalties
• California Public Law, "California Civil Code § 1671" (accessed Oct 2026): https://california.public.law/codes/civil_code_section_1671