The plan, 3 of 3

AI in practice: who carries the risk

When an AI tool gets it wrong, the solicitor who signed the work answers for it, in court and before the regulator. Every solicitor is told to check the tool before using it. Nobody has done that checking for them.

What every solicitor is told to do

The Society's guide, A Solicitor's Guide to Responsible Use of Artificial Intelligence (January 2026), carries a pre-adoption checklist. Among the things it asks each solicitor to do before buying a generative AI tool, in its own words:

"Consider undertaking robust research on vendors and tools available for the generative AI use case and evaluate integration with existing IT systems, tools and apps."

"Prior to purchasing a generative AI tool, endeavour to undertake a full contract review to understand the key issues relating to copyright, intellectual property, data ownership, data sharing, and access to data provided to the tool. This includes details on how prompts, queries and data are stored to ensure that use of the tool would not breach confidentiality obligations."

"Check the source(s) of data upon which the tool has been trained, noting that databases may contain misinformation or incomplete information."

"Consider undertaking thorough testing of the generative AI tool prior to its adoption."

That is an enterprise procurement programme, addressed to every solicitor equally. A sole practitioner is asked to conduct a full commercial review of a vendor's standard terms, audit its training data and run benchmark testing before taking out a subscription. The vendor's terms are the same document for everyone who signs them. The work is identical for every practice in the state, and every practice is asked to do it alone.

What the guide does not do

These are statements about what a document contains. They are not criticisms of the people who wrote it.

• It commits the Society to nothing. The phrases "we will", "the Law Society will", "will publish" and "will provide" do not appear. Every substantive sentence describes a risk or directs the solicitor to act.
• It sets no method for checking. "Solicitors must therefore always thoroughly check its outputs", and "solicitors should not rely on generative AI to verify sources produced by AI. This has been known to fail." No minimum step, no record, no standard a solicitor can point to when asked how they verified the work.
• It names no court. "Supreme Court" and "SC Gen" do not appear. A solicitor is told that court practice notes apply without being told which one, in which court, or what it says.
• It is silent on insurance. The words "insurance", "indemnity" and "Lawcover" do not appear. Nothing published by the Society tells a solicitor where an AI-assisted error sits against their professional indemnity policy. On the policy wordings I have read, a client's claim arising from an erroneous output the solicitor adopted is an ordinary third-party claim, and nothing excludes it because a machine was involved. The point is the absence of guidance, not the absence of cover.
• Privilege falls between two documents. The AI guide says that placing client confidential information into a publicly available generative AI system "is akin to putting it in the public domain" and that "clients may lose privilege". The Society's separate guide to client legal privilege (November 2025) contains no reference to artificial intelligence, generative tools, cloud or technology at all. One Society document raises the risk; the Society document on the subject does not deal with it.

Meanwhile the judges have a list. The Guidelines for NSW Judges on the use of generative AI (21 November 2024) set out red flags for AI-generated material: non-existent citations, out-of-date or inapplicable authorities, obvious substantive errors, repetitive non-specific language, spelling and expressions from other jurisdictions. The Society has reproduced that list in its own correspondence. Its guide to solicitors contains nothing equivalent.

What the Society already does

The gap is specific, and it should not be overstated. The Society has published the guide. It runs an AI for Legal Professionals Hub. The Practice Management Course has a mandatory "Technology and Innovation" subject, although it does not name AI. An AI Taskforce was established in 2024. And in its submissions to the Law Council the Society has a developed position on AI in litigation, including a request that the Council of Chief Justices refer the harmonisation of court rules on AI to the Harmonisation of Rules Committee. That last point matters: the Society cannot bind a court and has never claimed to. It has asked the body that can, through the mechanism that exists.

Three fixes

1. Do the checking once.

   The Society publishes a plain, vendor-neutral assessment of the standard terms and data handling of the tools solicitors commonly use: what happens to prompts and client data, where it is stored, who can see it, what the vendor may do with it. That is exactly the review the guide asks every individual solicitor to perform. Done once, centrally, it converts eight tasks per practice into one task performed once. It requires no rule change and no court's agreement.

2. Say what checking means.

   A one-page verification checklist a solicitor can point to when asked how they checked the output, built from the red flags the judges already use. Alongside it, press for harmonised court rules on AI through the Council of Chief Justices, which is already Society policy, and publish the member-facing standard that matches them, so that a solicitor in the Local Court, the District Court, the tribunal or the family courts knows what applies to them, not only a solicitor in the Supreme Court.

3. Protect privilege.

   Add technology to the privilege guide. The Society has already told a tribunal that prompts and responses created in generative AI "may not attract legal professional privilege and therefore may be discoverable by the opposing side". Its members should be told the same thing, in the document they turn to on privilege.

And three more, because the first three will not survive on their own

• Education by practice size. When the anti-money laundering obligations arrived, the Society published an implementation guide segmented by practice size. The AI guidance is not segmented at all. Scenario-based training on how these products fail, how to check citations and how to preserve privilege, designed for the practice that has no IT department.
• A word on insurance from the people who can give it. In 2017 Lawcover saw its first cyber-enabled fraud claims and responded with free risk education, then a group cyber policy from 2018, before any substantial claims history existed. The Society and its insurer are the only bodies that can tell solicitors where an AI-assisted error sits against the policy. Saying so costs nothing.
• A standing home, not a taskforce. There is no standing Council committee for technology. The AI Taskforce is a presidential initiative with no published terms of reference and no place in the committee structure, and presidents change every year. An earlier innovation programme, begun in 2016, disappears from the committee reports after 2021. Multi-year work needs a Council committee or a Professional Support Unit function on the anti-money laundering model.

A disclosure, and a change of emphasis

I build and run the systems my own firm uses, and I am the founder of Gleg, which builds AI tools for Australian legal work. A candidate who proposes that the Society assess AI products has an interest in the market being assessed, and you should know that. It is one reason the plan above asks the Society to assess vendor terms and data handling, which are the same for every buyer, rather than to rate products. My lodged statement speaks of plain assessments of what a product does and does not do. Having read the Society's own guide closely, I think the terms-and-data review is the sharper and cleaner form of the same idea, and it is the one I would move.

Sources

• Law Society of NSW, A Solicitor's Guide to Responsible Use of Artificial Intelligence (updated January 2026), including the pre-adoption checklist and the commentary on Rule 9.
• Law Society of NSW, Client Legal Privilege guide (November 2025).
• Law Society of NSW, letters to the Law Council of Australia of 5 June 2025 (with AI Taskforce discussion paper), 16 September 2025 and 7 April 2026.
• Supreme Court of NSW, Guidelines for NSW Judges in respect of the Use of Generative AI (21 November 2024), as reproduced in the Society's correspondence.
• Law Society of NSW, Annual Reports 2024 and 2025; Committees Reports 2020 to 2025; Practice Management Course Guidelines 2025; practising certificate application 2026-27.
• Lawcover professional indemnity policy wordings 2024-25 to 2026-27 and the Law Society and Lawcover submission of 2023, as produced to the Legislative Council under its order of 25 March 2026 without claim of privilege.