An AI issues binding court rulings without a human judge
Will the judiciary of any G20 or EU country authorise an AI system to issue binding rulings in any class of cases without a human judge approving each one?
Chance by Q4 2031
13%
Cumulative probability
- Latest forecast
| Year | Q1 | Q2 | Q3 | Q4 |
|---|---|---|---|---|
| 2026 | <1% | |||
| 2027 | <1% | <1% | 1% | 2% |
| 2028 | 2% | 3% | 3% | 4% |
| 2029 | 5% | 5% | 6% | 7% |
| 2030 | 8% | 8% | 9% | 10% |
| 2031 | 11% | 12% | 12% | 13% |
Model reasoning
No court system in a major economy or European Union country lets artificial intelligence issue binding rulings without a judge signing each one. In 2025–2026 China, Russia, Brazil, Spain and others said such tools may assist but not decide. I expect only a slow rise, to about 13% by the end of 2031. The development that would most change this is a formal move letting machines decide small claims with both sides' consent, an idea Britain's senior civil judge has floated.
Current status (as of 12 Oct 2026)
I found no qualifying publication since the 16 Sep 2026 opening date, and none before it either. Over 2025–2026, nearly every relevant judiciary has written down the opposite rule: AI assists, and a judge decides and signs.
- Russia: The Supreme Court's Order No. 59-P (3 Sep 2026, before the opening date) adopts an AI concept that runs to 2030. Its stated principle is that AI is "a tool to help the judge, not a replacement; the final decision always remains with a human". Its targets have AI drafting 30–50% of judicial acts, which judges then approve (cpk42.com, 22 Sep 2026; Fontanka, 20 Sep 2026).
- China: The Supreme People's Court's March 2026 work report keeps AI in an "auxiliary" role: "the subject of judicial responsibility can only be the judge" (court.gov.cn/zixun/xiangqing/492921.html). A Jiangsu court article of 9 Oct 2026 restates that AI "shall not replace judges".
- Brazil: National Council of Justice Resolution 615/2025, art. 10, bans AI systems that do not allow human review (atos.cnj.jus.br/atos/detalhar/6001).
- Spain: The judicial council's Instruction 2/2026 says AI "may not in any case replace" judges (BOE-A-2026-2205, 28 Jan 2026).
- Italy: Law 132/2025 reserves every decision to the magistrate.
- Turkey: A senior official of the judges' council said in Sep 2026 that "AI will not issue verdicts" (ekranhaber.com).
- Argentina: The Buenos Aires province draft regulation (June 2026) bans "full automation of jurisdictional decisions" (Infobae, 25 Jun 2026).
- United States: The federal judiciary's guidance cautions against delegating core judicial functions (Reuters, 17 Sep 2026).
- Mexico: Querétaro's SONIA system requires a judge to review each ruling (Plaza de Armas, Sep 2026).
- United Kingdom: The Master of the Rolls, Sir Geoffrey Vos, said on 18 Aug 2026 that there should be no machine-made judicial decisions without the parties' informed consent, and that replacing judges needs legislation (judiciary.uk Lehane Lecture PDF; Legal Cheek, Aug 2026).
- Outside scope: Even the UAE, which is not a G20 member, adopted a code in Sep 2026 saying no decision may be made by AI alone.
Reference class and base rate
- Fully automated binding orders built on fixed rules have existed for decades: Germany's automated payment-order procedure, and Saudi Arabia's Virtual Enforcement Court, which has issued enforcement decisions "without human intervention" since 2022 (SPA, 10 Feb 2025; Arab News, 2024).
- No G20 or EU judiciary has authorised a learned-model or knowledge-based system to decide binding rulings without per-case approval. That is 0 cases in roughly 10 years of "smart court" programmes and about 3.5 years of large language models.
- So the base rate is near zero, and the recent trend is toward explicit bans.
How I read the edge cases
I read the criteria literally: a publication dated after the opening date that establishes an earlier authorisation could count. In practice this only matters for Saudi Arabia. I judge its system nonqualifying for three reasons:
- it was set up by the Ministry of Justice, an executive body (the minister is only acting president of the judicial council);
- it mechanically enforces debt instruments, which falls under the "fixed procedural formula" exclusion;
- no new release has appeared since 16 Sep.
A lenient resolution of a future Saudi press release is still a small part of my near-term hazard.
Paths to YES (2027–2031)
- Existing high-volume automated pipelines, such as payment orders or enforcement, add AI screening, with orders issuing automatically and a court saying so.
- Opt-in, consent-based AI decisions for small claims. The UK (after legislation) or an innovative US state supreme court are the likeliest places.
- China, Russia, Saudi Arabia or Turkey reverse course as AI matures and caseloads grow.
- Transformative AI capabilities raise pressure everywhere. With about 43 countries and thousands of subnational courts in scope, and narrow opt-in pilots or interlocutory orders counting, one actor could move.
Paths to NO
- The current bans are recent and explicit, and the Russian and Chinese plans run to 2030.
- Constitutional and human-rights barriers apply, including the right to a fair hearing in Article 6 of the European Convention on Human Rights and Recital 61 of the EU AI Act.
- Most jurisdictions would need legislation.
- Crucially, routine human sign-off is cheap and protects legitimacy. Batch "rubber-stamp" approval, which does not count here, is the path of least resistance even with very capable AI.
- The criteria are strict. An official judicial publication must establish both binding effect and the absence of per-case approval, and executive agencies, arbitration and rule-based automation are all excluded.
Model
- The quarterly hazard starts at about 0.3%, covering mainly unknowns and edge cases.
- It rises roughly linearly to about 1% per quarter by 2031, as AI capability and court backlogs grow.
- Cumulative probability is computed as 1 − ∏(1 − h) in Python: about 1.8% by end-2027, 4% by end-2028, 6.8% by end-2029, 9.9% by end-2030 and 13.2% by end-2031.
- This is a judgement-based hazard model, not a fitted statistic.
Strongest case against my estimate
AI progress is fast; a highly capable model released in 2026 has already alarmed governments. If AI becomes clearly better than judges at routine adjudication by 2029, a centralised digital state or a consent-based UK or US pilot could move faster than I assume, and narrow classes do count. This could push the five-year figure toward 20–25%. On the other side, the near-universal 2025–2026 bans and the low cost of keeping nominal sign-off argue for something nearer 5%. I settle at about 13%.
Resolution criteria
Question. By the last day of each quarter, has the judiciary of any G20 country or EU member state authorised an AI system to issue binding rulings in any class of cases without a human judge approving each one?
Definitions.
- G20 country or EU member state: a sovereign country that is an individual G20 member or an EU member state on the date of the qualifying authorisation. Membership in the African Union alone, guest participation in the G20, and EU institutions themselves do not qualify.
- Judiciary: a national or subnational court, judicial council, or other body legally empowered to govern the country's courts. A single court's authorisation within its own jurisdiction is sufficient; nationwide adoption is not required. Executive agencies and private arbitration providers are not the judiciary.
- AI system: a machine-based system using learned models or knowledge-based inference to determine the substance of a ruling from case-specific inputs. Software that merely files, formats, transmits, or mechanically implements a decision already determined by a person or a fixed procedural formula does not count. There is no restriction on developer.
- Authorise: formally adopt a rule, order, practice direction, or other official judicial decision permitting the system to issue qualifying rulings. A judiciary announcement expressly confirming that such permission has been granted also counts. A statute alone counts only if it records or is accompanied by the judiciary's adoption of the authority. Final authorisation with a specified future commencement date counts; proposals, consultations, procurement, and permission still contingent on further discretionary approval do not.
- Binding ruling: a judicial determination that has legal effect on the parties' rights, obligations, or conduct of proceedings without needing case-specific judicial approval. Final judgments, default or payment orders, sentences, and binding interlocutory orders qualify. A right to appeal, object, or obtain later human review does not disqualify a ruling that otherwise takes effect automatically.
- Class of cases: an identifiable category of actual judicial proceedings defined by subject matter, value, location, procedural status, or another stated eligibility rule. A limited or opt-in pilot qualifies if it authorises legally effective rulings in real cases.
- Without a human judge approving each one: the authorised process allows qualifying rulings to take legal effect without case-by-case approval by a human judge or equivalent judicial adjudicator. General supervision, advance approval of the system or its rules, clerical authentication, sampled audits, and review only when requested do not constitute approval of each ruling. Human adoption of each AI recommendation, including routine sign-off, does.
Resolves YES if a qualifying official publication establishes that a judiciary has granted this authority for at least one class of cases. Issuance of the first actual ruling is not required. The publication must establish both the binding legal effect of the AI's output and the absence of required individual judicial approval; silence about human review is insufficient.
Resolves NO if no qualifying publication exists by the horizon. Excluded are advisory recommendations, AI-drafted decisions signed off individually by judges, nonbinding mediation, private arbitration, administrative agency determinations outside the judiciary, simulated cases, and automated court administration without AI determination of a binding ruling. General permission to use AI as a judicial assistant does not count.
Resolving source. Official court or judicial-council rules, practice directions, orders, judgments, and judiciary announcements on their official websites or in official gazettes; enacted statutes and regulations published in official legal repositories may establish the legal basis but must also satisfy the judicial-authorisation requirement above. These are the narrower resolving sources in place of Public confirmation (A2); vendor statements and news reports alone do not count. Apply the Resolution window and Resolution date (A3) conventions. Relevant nonqualifying precedents include Estonia's Ministry of Justice statement, “Estonia does not develop AI Judge”, and the Chinese Supreme People's Court's December 2022 account of its AI guidelines, which says rulings must be made by judges. EU membership is documented in the EU's official country directory; G20 membership is established by official G20 presidency membership records.