Sector Pulse
Legal AI Pulse: UK Courts, HSF Kramer, OpenAI, India, China, Gulf
UK judges, a global law firm, OpenAI, India's former president and Gulf arbitration centres weighed in on legal AI, and China's courts already have rules.
Courts, law firms and arbitration centres on four continents addressed legal AI in the same week. The message from judges is caution; the message from vendors and firms is speed; the evidence behind the speed is mostly self-reported.
Key takeaways
- On 18 November, UK Supreme Court justice Lord Sales argued that a "human in the loop" can become a rubber stamp, and that AI may de-skill the lawyers and judges needed to check it [1].
- Herbert Smith Freehills Kramer (HSF Kramer) hired six legal engineers from Norm Ai to help its 2,700 lawyers build agentic workflows [2][3].
- OpenAI says its Astra for Law configuration earned full credit on 20.8% of 120 assignments in a legal benchmark, against 13.1% for the best rival model it tested. These are OpenAI's figures [4][5].
- Arbitration institutions in Dubai, Cairo, Doha and Singapore described cautious, mostly internal uses of AI; Singapore's SIAC is still studying what to automate, and Dubai's DIAC says it is drafting use guidelines [6].
United Kingdom: the judge's warning
Lord Sales, speaking at a DAC Beachcroft conference, named three reasons for caution in courts: accuracy, confidentiality and legitimacy [1]. His sharpest point was about people, not software. If AI performs the junior work through which lawyers learn, he said, firms may stop producing the experienced lawyers who can spot a machine's mistakes. He suggested some training may need to happen without AI, and that "human in the loop" can decay into a "human rubber stamp" when caseloads press [1]. He also raised a possible access-to-justice paradox: cheap AI-written claims arrive faster than courts can decide them [1].
Firms and vendors move, a US court limits
HSF Kramer's announcement is concrete: six hires from Norm Ai, all trained as lawyers and then as legal engineers, joining the firm's global AI team [2][3]. Reuters framed it as part of a talent contest with AI-native firms [2].
OpenAI's 17 November post says Astra for Law, a legal configuration of GPT-6.2 Astra, scored 20.8% on LAB-AA, an Artificial Analysis version of Harvey's benchmark, against 15.0% for the base model and 13.1% for the best other developer's model. OpenAI also reports 0.03 material hallucinations per assignment, against 0.59 for Anthropic's Claude Opus 5.5, and says access is now open by request to every ChatGPT Enterprise customer [4]. LawSites notes that on Vals AI's research benchmark the model is a "top performer", not the leader (112 of 208 questions), and that most of the gain on Harvey's own test came from the new base model [5]. A separate technical report on arXiv, which lists no affiliations, claims first place on Harvey LAB against five baselines; the two claims use different comparison sets, are self-reported, and are not independent of each other [10].
Meanwhile a court has already drawn one line. In United States v. Heppner, Judge Rakoff ruled from the bench on 10 February, with reasons in a memorandum signed 19 February, that a defendant's own chats with Claude were not covered by attorney-client privilege or work product, noting the outcome might differ had counsel directed the use [7][8].
India and China: oversight and rulemaking
Former President Ram Nath Kovind told a Visakhapatnam audience on 14 November that lawyers "cannot transfer our professional credibility" to software, and that verification means opening the original judgment, not asking the tool to confirm itself [9]. He praised the Kerala High Court's AI policy for barring AI from generating judicial findings or orders [9]. China has taken a rule-first route: the Supreme People's Court's Opinions on AI-related disputes (Fa Fa [2026] No. 10), issued on 7 September, address the legal status of AI-generated content, training-data use and infringement cases [11][12]. They govern how courts hear AI disputes; we did not find them to cover judges' own use of AI.
Gulf, Egypt and Singapore: institutions go slowly
At Dubai Arbitration Week, a panel of institutional leaders described mostly administrative uses [6]. DIAC's registrar said it is developing guidelines using a traffic-light model that sorts AI uses into acceptable, disclosure-required and prohibited. Cairo's CRCICA uses AI internally for research and statistics, Qatar's QICCA limits it to summarising data and correspondence, and SIAC chose not to add AI provisions to its rules, keeping award scrutiny human-led [6].
Strengths, weaknesses and open questions
| Region | What happened | Status | Evidence |
|---|---|---|---|
| UK | Supreme Court speech | Opinion, no rule | Speech text |
| Global | HSF Kramer hires; Astra for Law | Live, vendor claims | Third-party benchmarks, results reported by OpenAI |
| US | Heppner privilege ruling | Court decision | Court opinion |
| India | Kovind lecture | Commentary | Press report |
| China | SPC Opinions | Issued 7 Sept | Law-firm summary, translation |
| Gulf, Egypt, Singapore | Arbitration panel | Plans, drafting | Conference report |
Strengths: faster research and drafting, and institutions writing rules early. Weaknesses: benchmark numbers come from the vendor, fabricated citations remain a documented risk, and confidentiality limits what courts and institutions can adopt. Open questions: who audits benchmark claims, whether a 20.8% full-credit score is good enough for client work, and how firms train juniors if AI does their tasks. Pricing for Astra for Law was not stated in the sources we read, and availability is by request.
Our take
In our analysis, the week's real story is the gap between deployment and verification. Firms and vendors are buying capability; judges and arbitral bodies are asking who checks it. Lord Sales's de-skilling point is the one nobody has an answer to yet.
What to watch next
- Independent evaluations of Astra for Law beyond OpenAI's own [4].
- DIAC's guidelines and whether other institutions copy them [6].
- Other US courts testing the Heppner reasoning [7].
- Any UK judicial guidance following Lord Sales's speech [1].
FAQ
Can judges use AI to decide cases? Not yet in the sources we read: Lord Sales says the line depends on case type and review, and Kerala's policy bars AI from generating findings or orders [1][9].
Is an AI chat privileged? In Heppner, no, because the defendant acted alone; the court left open a different result if counsel directs the use [7].
Is Astra for Law the best legal AI? OpenAI says it leads one benchmark and is a top performer on another; that is a vendor claim [4][5].
This article is analysis, not legal or investment advice. AI Worldwire uses AI assistance in research and drafting; an editor reviews every article. The AI system that assisted this article, Claude, is made by Anthropic.
Sources
- Artificial Intelligence, the Courts, the Legal Profession and the Legal Process (speech), Lord Sales, UK Supreme Court, 18 November 2026. Primary.
- Law firm HSF Kramer taps legal engineering team to fuel AI capabilities, Reuters, 16 November 2026.
- Herbert Smith Freehills Kramer hires a team of legal engineers, HSF Kramer, 17 November 2026. Primary.
- Astra for Law leads in legal intelligence, OpenAI, 17 November 2026. Primary; vendor claims.
- OpenAI opens Astra for Law to all ChatGPT Enterprise customers and releases new benchmark results, LawSites, November 2026.
- Dubai Arbitration Week 2026: AI in arbitration, SCC Times, 13 November 2026.
- United States v. Heppner, memorandum opinion, US District Court, Southern District of New York, signed 19 February 2026 (bench ruling 10 February). Primary.
- Attorney-client privilege and work product in the age of generative AI, White & Case (law firm alert), 2026.
- Lawyers Cannot Hand Over Their Credibility to AI, Says Former President Ram Nath Kovind, LawBeat, 17 November 2026 (lecture of 14 November).
- Sage S2.1 Flash Technical Report, arXiv, 16 November 2026. Primary; authors' own results.
- Cross-Border AI Deployment Regulatory Framework, China-EU Comparison and Compliance Checklist, Taylor Wessing (law firm), 17 November 2026.
- SPC Opinions on Hearing Cases of AI-Related Disputes, China Law Translate (English translation of the court's text), 16 September 2026.