Lesson 1: What AI Can and Cannot Do in Legal Work

Before you start

This course gives general information about using AI in legal work. It is not legal advice and does not replace your regulator's rules, your firm's or chambers' policy, or your own professional judgement. Where a lesson quotes a figure, it gives the date, because this field changes month by month.

In this first level you will:

1. Learn what a generative AI assistant is actually doing when it answers you.
2. See dated evidence of how the profession uses these tools.
3. Sort common legal tasks by how risky they are.

What a large language model actually does

General assistants such as ChatGPT, Gemini, Claude and Microsoft 365 Copilot run on large language models (LLMs). An LLM is trained on huge amounts of text and learns which words tend to follow which. When you ask it something, it builds an answer one piece at a time, each time choosing a likely next word.

Think of a very well-read pupil who has never been inside a law library and answers from memory, in confident, polished prose. Most of the time the answer is close. Sometimes it is invented, and it sounds exactly as confident as when it is right.

Remember it in two sentences: An LLM produces the text that is statistically plausible, not text it has checked against a source. A case name, citation or quotation can therefore look right and read fluently while being completely made up (a "hallucination").

Three consequences follow:

• Fluency is not accuracy. Good grammar and a neat citation format prove nothing.
• The model does not know when it is wrong. It will not reliably flag its own inventions.
• Some tools help but do not remove the risk. Tools that are "grounded" in a set of documents or a legal database are less likely to invent material, but you still check the output.

Evidence: how the profession is using AI (dated)

Use dated sources, not impressions. Four worth knowing:

• February 2025, LexisNexis UK survey (published 3 April 2025): 46% of legal professionals said they were using generative AI for work, up from 11% in July 2023. In-house teams reported the highest use (72%) and the Bar the lowest (28%).
• September 2026, LexisNexis "The AI-Dependent Lawyer" (UK, 500+ lawyers): 34% used AI tools for legal work every day and a further 32% several times a week. The top uses were drafting, legal research and document review. 15% of lawyers at large firms said they were "now dependent on AI" to do their job.
• May 2025, Solicitors Regulation Authority: the SRA authorised Garfield.Law, described as the first AI-driven law firm in England and Wales, for small debt claims. Named regulated solicitors stay accountable for its work.
• October 2025, Courts and Tribunals Judiciary: updated AI guidance for judicial office holders replaced the April 2025 version. It covers hallucinations, bias and confidentiality, and stresses personal responsibility for anything produced in a judge's name.

Evidence: what goes wrong

On 6 June 2025 the Divisional Court gave judgment in R (Ayinde) v London Borough of Haringey [2025] EWHC 1383 (Admin), heard together with Al-Haroun v Qatar National Bank. In Ayinde, five fictitious authorities appeared in judicial review grounds. In Al-Haroun, a solicitor relied on research by his client that included many non-existent cases. The court said:

• Generative AI tools such as ChatGPT are not capable of conducting reliable legal research.
• A lawyer cannot rely on a lay client for the accuracy of citations put before the court.
• Citing false authorities is likely to lead to referral to a regulator such as the SRA or BSB, and the most serious cases could amount to contempt.

The lawyers involved were referred to their regulators.

Task: sort ten legal tasks by risk

DO: Before reading the answers, put each task into one of three groups yourself.

• Suitable with checking: low stakes, and you can verify the output easily against material you hold.
• High risk: possible to do with AI, but errors are costly or hard to spot, so you need close expert checking.
• Not suitable: the task needs human judgement or accountability, or it relies on facts the tool cannot verify.

The categories depend on the stakes and on whether you can check the output, not on how hard the task is. Research is high risk because a made-up case reads exactly like a real one.

Task: a basic check routine

DO: Whenever you use AI output in your work, follow these steps:

1. Decide the category before you start, using the table above.
2. Read every word of the output. Do not skim.
3. Verify each authority, quotation and fact against the primary source, such as the official judgment or the legislation.
4. Correct or delete anything you cannot verify.
5. Take ownership. The work goes out under your name, and you remain responsible for it.

Which tools you may use, and what you must never type into them, are covered in the next levels. Until then, do not put client information into any tool your organisation has not approved.

Try it on your own work

1. List five tasks from your past week.
2. Put each one into one of the three categories.
3. For one "suitable with checking" task, write down how you would check the output.

Key takeaways

• An LLM predicts plausible text, so it can produce fluent legal content that is false.
• Dated surveys show AI use in UK legal practice rising quickly: 46% in February 2025, and 34% using it daily by September 2026 (LexisNexis).
• Ayinde (June 2025) shows the consequence of unchecked AI citations: likely referral to a regulator.
• Sort tasks by stakes and how easily you can check the output.
• Verify every authority against the primary source. Responsibility stays with you.