DR PETER FIELDS • AI & IDEAS

THE AI NOTEBOOK / WHAT MATTERS AND WHY

AI and the Justice System, Part 1 · The Legal AI Stack

Dr Peter Fields •


Download this article (PDF)

AI Explained

How AI is changing legal work and the way disputes are decided

Part 1 · The Legal AI Stack

By Dr Peter Fields

THE 2-MINUTE SUMMARY

A legal AI tool combines a model with the sources, permissions and software that sit around it. A general chat app can connect to a legal publisher. A specialist platform can use a model you already know from a general app.

Ask which jurisdiction the tool covers, what the search can find and whether it can show you the passage behind an answer. A real citation can still support the wrong proposition.

Check what happens to client material at each stage. A promise that a model supplier keeps no copy can sit alongside a platform that stores your documents or chat history.

Buy against a job you can test. Compare the time it takes to reach a checked answer, the errors reviewers have to catch and who remains responsible for the work.

Two tools and one question

Suppose a commercial litigation team needs to check how the courts of England and Wales have applied the limitation rules for concealed fraud, then relate those authorities to a disclosure bundle. You can ask the same question in a chat window or a specialist legal platform. Both may produce an orderly memorandum. By the time someone signs it off, much depends on what the tool could search and what the lawyer can check.

What is behind the screen?

Start with three overlapping categories. A general chat app can help organise an argument or explain an unfamiliar term. It might also have web search. Its access to legal databases depends on what has been connected, which subscription the user has and what the connector covers. In May 2026 Anthropic and Thomson Reuters announced a connector bringing CoCounsel Legal into Claude. Anthropic added a separate set of legal tools in October. [1]

A model developer can also offer a version of its model configured for legal work. In September 2026 OpenAI launched Astra for Law with GPT-6 Astra and a dedicated US legal search index. On 17 November, OpenAI said it had moved the service to GPT-6.2 Astra and added sources from the UK, Australia and Canada. Enterprise customers can request access. A firm researching English limitation law still needs to check which English sources its particular setup can reach. [2]

Specialist platforms and publishers offer another route. One may be valuable because it can extract the same clause from a set of contracts and put the answers in a table. Another may give you access to a publisher’s case law and tools for checking how a decision has been treated. The supplier may use several models. Harvey’s August 2026 Tenet research preview, for example, used Moonshot’s downloadable Kimi K3 as a starting point and trained it further with Fireworks for legal work. That is a description of the release. The performance question has to be tested separately. [3]

What happens to the documents?

The distinction between a platform and the model it calls matters when you upload a client file. Harvey, for example, describes how it stores files and uses a search index. Its terms distinguish its usual commitments from features with different processing arrangements. Ask who stores the original and the chat, what each supplier can keep, and what happens if someone uses web search or connects another service. A zero-retention promise for one supplier answers one part of that enquiry. [4]

In UK and Munir, a case about false citations and professional supervision, the Upper Tribunal said that uploading confidential client material to a public AI tool amounted to disclosure, breached confidentiality and waived privilege. Its observations are about what follows from the upload. The tribunal did not find that someone at the provider had read the material or that other users could retrieve it. Legal advice privilege ordinarily protects confidential communications between client and lawyer for the purpose of giving or receiving legal advice. In April 2026 Sir Colin Birss expressed the view that a solicitor’s use of a secure AI system in doing that work need not disturb privilege. A client asking a chatbot directly for advice raises a separate question. The Confidentiality and Privacy articles discuss the distinction in more detail. [5, 6]

The SRA warning notice and the BSB guidance explain how existing duties apply. The practitioner remains responsible for confidentiality, checking and supervision. Before using client material, check the firm’s approval, the terms agreed with the client and what the supplier’s contract actually provides. Before buying a tool, establish who checks the work and how the firm would handle an error. [7]

Geek Box – A real case can still be the wrong case

Retrieval-augmented generation is a way of finding passages and giving them to the model with the question. A good search can surface the right report. The model may still treat a judge’s aside as the rule the court actually decided, or miss the effect of a later appeal. A citator helps trace how the case has been treated. You still need to read what the later court did with it. [8]

What the tests tell us

Magesh and colleagues tested three US legal research tools in spring 2024 and reported hallucinations in 17–33% of queries under their own definitions of correctness and grounding. In a separate Vals evaluation, four products scored 74–78% on a weighted measure of performance, compared with 69% for lawyers from one US firm. Vals started with 200 US research questions and excluded one from scoring. For English law, Linklaters has published a 50-question test of general models. The tests answer different questions. Try the particular version you would use on the sort of English case you handle. [8, 9, 10]

How to use it better

Before buying, test the tool on the work you have in mind. Give it a question where a decision has been reversed, one where a statutory amendment changes the answer and one where there is no good authority for the proposition. Mark source selection and legal analysis separately. Count the time someone spends opening the authorities and putting things right. The subscription price is only one part of the cost.

Part 2 asks what happens to fees and instructions when the first pass takes less time.

Sources and notes

1. Anthropic and Thomson Reuters, legal tools and CoCounsel integration (12 May 2026). Anthropic, legal artifacts (28 October 2026). May release • Connector • October release

2. OpenAI (17 September 2026), Introducing Astra for Law, and the 17 November 2026 update. Announcement • November update

3. Harvey (20 August 2026), Harvey Tenet Research Preview. Research preview

4. Harvey, How Harvey Manages Customer Data and Service Terms. Check the terms applicable to the account and feature in question. Data handling • Service terms

5. UK and R (on the application of Munir) v Secretary of State for the Home Department [2026] UKUT 81 (IAC), promulgated 17 November 2025, at [21] and [60]. Decision

6. Sir Colin Birss (22 April 2026), Legal Professional Privilege in the Age of AI, §§19–28, especially §26. See also R (Prudential plc) v Special Commissioner of Income Tax [2013] UKSC 1, [17]–[24], [51]–[52]. Speech • Supreme Court judgment

7. Solicitors Regulation Authority, Misuse of AI warning notice (17 August 2026). Bar Standards Board, Guidance on the Use of Artificial Intelligence and Other Technologies (18 May 2026). SRA • BSB

8. Varun Magesh, Faiz Surani, Matthew Dahl, Mirac Suzgun, Christopher D. Manning and Daniel E. Ho (2025), Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools, Journal of Empirical Legal Studies 22(2), §4.3 and Table B1. Paper

9. Vals AI (14 October 2025), VLAIR – Legal Research. See the executive summary and scoring rubric. Report

10. Linklaters (10 February 2025), The LinksAI English law benchmark (Version 2). Evaluation

Download this article (PDF)