The Law Society AI advisory: what Singapore firms must do

Earlier this year, a Singapore High Court judgment ordered two lawyers to personally pay S$5,000 each. Their closing submissions, in a civil suit over money the late Tan Thuan Teck was said to have lent his brothers, cited two cases that did not exist. One combined the name of a real case with a different case’s citation. The other was invented from nothing. Opposing counsel caught it, not the lawyers who had filed it ([2026] SGHC 49).

Three weeks after that judgment surfaced, the Law Society sent members the Law Society AI advisory, a short notice formally titled the Advisory on the Use of Publicly Available AI Tools, dated 2 April 2026. It tells lawyers to check a tool’s terms of use before putting client work through it, because a chatbot does not suspend the duty of confidentiality. It is short, interim, and does not by itself tell a firm what to do about a workflow. This piece covers what the advisory says, what it does not cover, and what a firm should actually change this week.

What the Law Society’s AI advisory says

The advisory runs to a page, addressed “Dear Members,” and reads as a holding notice rather than a finished policy. The Law Society’s Generative AI Committee is still drafting fuller guidance, and in the meantime it gives members a small number of non-exhaustive rules of thumb, verified on the Law Society’s website, 2 April 2026:

• Before using a publicly available AI tool, read and understand its terms of use, and check that those terms are consistent with the lawyer’s professional duties, confidentiality above all.
• Understand the legal effect of what the terms actually say. If the terms are in a language other than English, an automated translation is not treated as good enough to rely on.
• Do not upload or submit confidential client information to a publicly available AI tool. The advisory draws a direct line to tools lawyers already treat with caution: a LinkedIn post, a tweet, a Google search box. A public AI chat window carries the same exposure.

The advisory sits alongside two longer documents it references rather than replaces: the Law Society’s own Guide on the Adoption of LegalTech for Law Practices, and the Ministry of Law’s Guide for Using Generative AI in the Legal Sector, published 6 March 2026. Read together, the sequence is: MinLaw sets out the full obligation in March, the Law Society narrows in on the sharpest edge of it in April, and the professional conduct rules underneath both were never optional to begin with.

Where “publicly available” draws the line

The advisory’s scope is narrower than its headline suggests. It is not a statement on generative AI in legal practice generally. It is a statement about one category: tools anyone can sign up for on a personal account, ChatGPT, Claude, Gemini, and similar, used with no contract between the firm and the vendor. That is exactly the category a firm has the least control over, and exactly the category most associates and paralegals already reach for on their own initiative.

The test the advisory implies is not complicated, and a firm can run it without hiring anyone. For each task someone in the firm already does with a public AI tool, ask two questions. First, does the input contain anything a client would call confidential, a name, a figure, a fact pattern, a draft clause tied to a live matter? Second, has anyone at the firm actually read the tool’s terms and confirmed the provider does not retain or train on that input? If the answer to the first question is yes and the second is no, the task does not belong on that tool, whatever the associate’s login screen says.

Work that survives that test on a free tool tends to be genuinely low-risk: drafting a generic internal memo template, summarising a public statute, checking English grammar in a non-confidential email. Anything built from an actual client’s facts does not survive it. That second category needs a tool the firm has vetted and, ideally, a system that runs on the firm’s own tenant rather than a vendor’s shared consumer product.

What the advisory leaves for a firm to build

A one-page advisory was never going to replace a governance program, and it does not try to. It gives a firm a tripwire for the riskiest single behaviour, staff pasting client facts into a free chatbot, but it says nothing about who owns the AI-use policy, how a firm scores which workflows are worth automating, or how it proves to a client that a reviewed system is safer than an associate’s personal ChatGPT tab. That is the ground the MinLaw guide’s five-step framework covers, and our explainer on what the guide asks of law firms walks through the three principles, the confidentiality tier, and the human-in-the-loop test in full. Our MinLaw compliance checklist turns those five steps into a working list a partner and an IT lead can run through in-house.

The gap between “don’t paste confidential data into ChatGPT” and “here is the reviewed system your staff should use instead” is where most firms stall. Reading the advisory and stopping there leaves a firm able to say what its people should not do, with nothing built to replace the free tool they were already reaching for.

What a firm should do this week

Three things, none of which requires a vendor. First, find out what staff are actually using. Not the tool the firm sanctioned, the tool associates and paralegals have quietly adopted, usually a personal account on whichever chatbot is free. A short, anonymous survey gets an honest answer faster than an email asking people to confess. Second, write the line down: which tasks may touch a free public tool, and which must stay on a system the firm controls, and circulate it as one paragraph, not a policy nobody reads. Third, add one sentence to the standard engagement letter noting that the firm may use AI tools in preparing work product, with an option for the client to ask questions. The Law Society’s own guide ships a sample clause to start from.

None of that is a build. It is the paperwork a firm can produce in a week, and it is also exactly what the guide’s three principles ask for before a single dollar goes toward a system: professional oversight proportionate to risk, confidentiality matched to the tool, and honesty with the client about what is being used. A firm that has done this much is in a materially different position than one relying on an advisory nobody circulated past the partner who opened the email.

The grant path once the firm is ready to build

Most Singapore law firms clear the SME test the Enterprise Development Grant applies: Singapore-registered, at least 30% local equity, and group turnover of S$100 million or less, or group headcount of 200 or fewer. EDG covers up to 50% of a qualifying project, subject to EnterpriseSG approval and never guaranteed, verified July 2026 on enterprisesg.gov.sg. A diagnostic that produces the policy, use-case register, and vendor due diligence the guide asks for, and a build that puts a reviewed workflow on the firm’s own tenant, both sit inside a qualifying project scope. The process and a worked example are in our EDG grant guide for AI projects, and the full cost of a diagnostic-plus-build engagement is in what AI consulting costs in Singapore.

Whatever a firm builds to close that gap has to outlast the vendor who built it, or the firm is back to an unwritten rule and a personal ChatGPT account within a year. That argument, and why training the firm’s own staff to run the system matters more than the system itself, is in why AI projects fail.