Governance

AI and the Uniform Law: the conduct rules that already apply

By Michael Nadalin, Founder, Market Lead · 16 September 2026 · 8 min read

Picture the Friday afternoon version. Written submissions are due at 4pm. A junior solicitor has used a general chatbot to pull the authorities together, and one of the cases has a plausible name, a plausible medium neutral citation and no existence at all. Nobody checks it. The document is filed.

The Uniform Law has no rule headed 'artificial intelligence', and people read that absence as a regulatory gap. It is not one. Every obligation that decides how that Friday plays out, from the duty not to mislead the court to supervision to what you can bill for the work, was in force before anyone at the firm opened a chatbot.

This piece works through those obligations under the Legal Profession Uniform Law, which applies in New South Wales, Victoria and Western Australia, and the Australian Solicitors' Conduct Rules made under it. Most other states and territories have adopted the same conduct rules with local variations, so the analysis applies there too.

No AI rule exists, and you do not need one

The conduct rules are technology neutral. They regulate what you do, not the tool you do it with. A false citation is a false citation whether it came from a tired associate, a careless search or a language model. The regulators have taken the same line: the joint statement on AI use from the Law Society of NSW, the Legal Practice Board of Western Australia and the Victorian Legal Services Board and Commissioner frames AI use through existing duties, not new ones.

That cuts both ways. You do not need special permission before using AI, and you get no special defence when it goes wrong. 'The software produced it' appears nowhere in the disciplinary provisions. Under section 296, unsatisfactory professional conduct includes conduct that falls short of the competence and diligence a member of the public is entitled to expect of a reasonably competent Australian lawyer. It makes no difference what produced the error.

The belief versus the rule that already applies
There are no rules on AI yet
Rules 3, 4, 9, 19 and 37 apply to every AI output you rely on
The vendor is responsible for accuracy
The practitioner who signs, sends or files the document is responsible
Confidentiality only matters for privileged material
Rule 9 covers any information confidential to a client, privileged or not
Faster work means the same bill and a better margin
Section 172 requires costs to be fair and reasonable for the work actually done
Only litigators need to worry
Competence, confidentiality and costs duties apply equally to advisory and transactional work
Each common assumption about AI use has an existing answer in the Uniform Law or the Solicitors' Conduct Rules.

Rule 19 and the citation you did not check

Rule 19.1 prohibits a solicitor from deceiving, or knowingly or recklessly misleading, the court. Filing an authority you have not read, taken from a tool known to invent authorities, sits very close to recklessness. The paramount duty to the court and the administration of justice in rule 3 sits over all of it, and the Barristers' Rules impose equivalent duties on counsel.

The courts have gone further than restating principle and set out procedure. Supreme Court of New South Wales Practice Note SC Gen 23 requires that, where generative AI has been used to prepare written submissions, the author verifies that every citation, authority and legislative reference exists, is accurate and is relevant. It also bars generative AI from generating the content of affidavits, witness statements and character references, and requires leave before it is used for expert reports. Australian courts have already referred practitioners to regulators over false citations in filed documents.

Verifying a citation means reading the source. Asking the same tool whether it is sure does not count, because language models will confirm their own inventions with complete confidence.

A verification routine that satisfies rule 19
1
List every authority
Extract each case, section and secondary source the document relies on, including those buried in footnotes.
2
Open the primary source
Retrieve the judgment or legislation from an authorised report, AustLII or the official legislation register. A summary is not the source.
3
Check the proposition
Confirm the passage at the pinpoint actually supports the point made, and that the authority has not been overturned or confined.
4
Look for the case against you
Rule 19.6 still requires disclosure of directly relevant adverse authority known to you. A tool asked to support an argument will not volunteer the case that defeats it.
5
Record who checked
Note on the file who verified each authority and when. If the question is ever asked, you want a documented answer.
Five steps before any AI assisted document goes to a court or tribunal.

Confidentiality: rule 9 and where the prompt goes

Rule 9 prohibits disclosing information confidential to a client, subject to narrow exceptions. Pasting a client's facts into a public chatbot discloses them to a third party. Whether that amounts to a breach depends on what the provider does with the text: whether it is stored, for how long, where, who can access it, and whether it is used to train future models.

The answers are in the terms of service and data processing terms, not on the marketing page. Consumer tiers of well known tools have commonly allowed the provider to use inputs unless the user opts out, and those settings change. Read the current terms for the tier your staff actually use, not the enterprise tier you assume they use.

Court orders add a further layer. SC Gen 23 restricts entering material subject to suppression or non-publication orders, material covered by the implied undertaking, and material produced on subpoena into generative AI tools unless specific safeguards are met. That restriction applies even if the client would happily consent.

AI tool setups ranked by confidentiality exposure
Personal accounts on free consumer tools
No firm contract, inputs may be retained or used for training, no audit trail, and staff control the settings.
Highest
Individual paid subscriptions
Sometimes better controls, but the contract is with the individual and the firm can neither see nor enforce the settings.
High
Firm licensed general tools on enterprise terms
Contractual limits on retention and training plus admin controls, but data location and subprocessors still need checking.
Moderate
Legal specific tools with Australian data terms
Terms written with privileged material in mind and clearer retention and residency, but only as strong as the contract.
Lower
Questions about the law with no client facts
Asking about legislation or case law without identifying client information removes most of the rule 9 exposure.
Lowest
Highest exposure first. The terms you have actually signed decide where a tool sits, not the vendor's brochure.

Competence and supervision: rules 4 and 37

Rule 4.1.3 requires legal services to be delivered competently, diligently and as promptly as reasonably possible. Competence now includes a working knowledge of how the tools in your practice fail. A solicitor who does not know that language models produce fluent, wrong answers is not in a position to use one competently.

Rule 37 requires a solicitor with designated responsibility for a matter to exercise reasonable supervision over the solicitors and other employees working on it. If a paralegal or graduate uses AI on your matter, supervise the output as if they wrote it themselves, because for regulatory purposes the firm did. Principals also carry an obligation under the Uniform Law to take all reasonable steps to ensure the practice and its associates comply. A firm with no policy and no visibility of which tools staff use will struggle to show those reasonable steps.

Independence matters here too. Rule 4.1.4 requires you to avoid any compromise to your professional independence, and rule 17 requires forensic judgments to be your own. Adopting a tool's framing of the issues without testing it is a quiet version of that problem.

Supervising an AI assisted task
Before
Set the boundaries
  • Which approved tool and which tier
  • What client information may be entered
  • What the output will be used for
During
Keep the work checkable
  • Save prompts and outputs to the file
  • Flag every passage that came from AI
  • Verify authorities as they are added, not at the end
Before it leaves the firm
Sign off as the author
  • Supervisor reads the sources for key propositions
  • Check the court's own AI requirements and disclosures
  • Confirm the advice reflects independent judgment
What the supervising solicitor should be able to see at each stage.

Costs: sections 172 and 174 when the draft took minutes

Section 172 requires a law practice to charge costs that are no more than fair and reasonable in all the circumstances, proportionately and reasonably incurred and proportionate and reasonable in amount. The time spent and the skill involved are part of that assessment. If a tool produced a first draft in ten minutes and review took forty, billing three hours for it is hard to defend when the bill is assessed.

The honest options are simple. Bill the time actually spent, including genuine review time, or move the task to a fixed or value based fee disclosed up front. What you cannot do is bill notional time for work nobody performed.

Section 174 disclosure is where this is best settled. If AI changes how a matter will be charged, or you intend to pass tool costs through as a disbursement, say so at engagement and update the disclosure if the position changes significantly. An explanation given at the start carries far more weight than one offered during a costs assessment.

Where the costs obligations meet AI across a matter
Engagement
Section 174 disclosure: explain the basis of charging, including any fixed fee for AI assisted tasks and any tool costs passed through.
During the matter
Record actual time, including review. Keep AI outputs on the file so the work behind each time entry can be shown.
Significant change
If AI materially changes the estimate, update the disclosure instead of letting the final bill drift from it.
Billing
Apply the section 172 test: fair, reasonable and proportionate for the work actually done, not the time the task used to take.
Dispute
On a costs assessment, time entries for AI drafted work can be tested against what the file shows was done.
The same Uniform Law costs provisions apply at every stage. AI simply changes the facts they are applied to.

A one page AI policy that maps to the rules

Most firms do not need a forty page framework. They need a document short enough that people actually read it, with each line tied to a rule: approved tools and tiers (rule 9), client information that must never be entered (rule 9 and any court orders), mandatory verification of every authority before filing (rules 3 and 19), who supervises AI assisted work and how it is recorded (rule 37), and how that work is billed and disclosed (sections 172 and 174).

Then check the requirements court by court. The NSW Supreme Court practice note is the most prescriptive so far, other courts, including the Supreme Court of Victoria, have issued their own guidance, and all of it gets revised. Make one named person responsible for keeping the list current.

Choosing the right tool reduces the work but does not remove it. A research tool that checks citations against authoritative Australian sources and anonymises client-identifying detail before anything reaches a model, which is how Legal Brain is built, shortens the rule 19 and rule 9 steps. It does not replace them. No tool does, and a vendor who tells you otherwise is describing a product you should keep well away from a court document.

Frequently asked questions

Is there a specific rule about AI in the Legal Profession Uniform Law?

No. Neither the Uniform Law nor the Australian Solicitors' Conduct Rules contain an AI specific rule. Existing duties govern AI use instead: the paramount duty to the court, the duty not to mislead, confidentiality, competence, supervision and the costs provisions. Individual courts, such as the Supreme Court of NSW, have added their own practice notes on top.

Can a solicitor put client information into ChatGPT?

Rule 9 prohibits disclosing information confidential to a client, and entering it into a third party tool is a disclosure. Whether it breaches the rule depends on the provider's retention, training and access terms for the specific tier being used. Material subject to suppression orders, the implied undertaking or subpoena carries extra restrictions under court practice notes, so the safe default is no identifiable client information in consumer tools.

What happens if a lawyer cites a fake case generated by AI in Australia?

The solicitor risks breaching rule 19.1, which prohibits knowingly or recklessly misleading the court, as well as the paramount duty in rule 3. Depending on the circumstances, that can be treated as unsatisfactory professional conduct or professional misconduct. Australian courts have already referred practitioners to their regulators over false citations in filed documents.

Can lawyers bill for time saved by using AI?

Not as if the time had been spent. Section 172 of the Uniform Law requires costs to be fair, reasonable and proportionate, so time billing should reflect the time actually worked, including review. If you want to capture the value of faster work, use a fixed or value based fee disclosed under section 174 at engagement.

Check your understanding

Two quick questions

No score is stored. Pick an answer to see why it is right.

  1. 1Under Supreme Court of NSW Practice Note SC Gen 23, what must happen when generative AI is used to prepare written submissions?

  2. 2A graduate uses an AI tool to draft advice on a matter you are responsible for. Which conduct rule most directly makes the quality of that output your problem?

Research Australian law without handing over client data

Legal Brain searches Australian legislation and case law, shows you the source behind every answer, and anonymises client-identifying detail before anything reaches a model.

Request early access