AI compliance for Australian law firms
AI compliance for law firms: confidentiality, courts and evidence
Most Australian businesses working out their AI position begin with privacy law. A law firm has to begin earlier, with confidentiality, privilege, the Harman undertaking and a practice note that says in terms what generative AI may and may not be used for in proceedings.
Those obligations existed before AI and they are not satisfied by a policy that says staff should be careful. They are satisfied by knowing which tool holds what, and by a person verifying what comes out.
General information about Australian obligations, not legal advice. Check the requirements of the courts and regulators in your own jurisdiction.
What do Australian courts require about generative AI?
Requirements differ by jurisdiction, and several courts have now issued practice notes. The Supreme Court of New South Wales issued Practice Note SC Gen 23 on 28 January 2025, commencing 3 February 2025 and applying to all proceedings from that date. Check the current requirements of the courts you actually appear in, because they are being reviewed and revised.
SC Gen 23 is worth reading in full because of how specific it is, and because it is a useful model for what any firm should be deciding internally. A summary of its structure follows. It is a summary, not a substitute for the practice note.
Uses it expressly allows
Subject to the confidentiality condition below: generating chronologies, indexes and witness lists; preparing briefs or draft Crown Case Statements; summarising or reviewing documents and transcripts; and preparing written submissions, subject to verification.
Evidence is off limits
Generative AI must not be used to generate the content of affidavits, witness statements, character references or other material intended to reflect a witness's evidence or opinion, and must not be used to alter, embellish, strengthen, dilute or rephrase a witness's evidence.
Restricted material
Material under non-publication or suppression orders, material subject to the Harman undertaking, material produced on subpoena and material under a statutory publication prohibition must not be entered into a generative AI program unless the practitioner is satisfied it stays in a controlled environment and is not used to train any model.
Two obligations that reach every firm, not only litigators
Affidavits, witness statements and character references must contain a disclosure that generative AI was not used in generating their content. That is a positive statement in a document, which means somebody in the firm has to actually know whether it is true.
Where generative AI has been used in preparing written submissions, the author must verify in the body of the submissions that all citations, authority, case law and legislative references exist, are accurate and are relevant, and that verification must not be carried out solely by using a generative AI tool. Using AI to check AI is expressly not enough.
Source: Practice Note SC Gen 23 (Supreme Court of New South Wales, issued 28 January 2025, commenced 3 February 2025). The practice note states it will be periodically reviewed. Requirements in other Australian courts and tribunals differ and should be checked directly.
Why unverified output is a professional problem, not a quality one
Because the consequence follows the practitioner. The ASD's Australian Cyber Security Centre cites a 2025 case in which a lawyer used AI to prepare a court document, the tool generated false cases, the lawyer did not verify them before submitting, and after the court discovered it the lawyer was barred from operating and owning a law practice.
SC Gen 23 sets out why this happens in language any firm can use internally: generative AI can produce apparently plausible, authoritative and coherent responses that are inaccurate or fictitious, including false citations and fabricated legislative, case or secondary references. It also notes that underlying data sets may be incomplete, out of date, or not relevant to New South Wales or Australia.
The practical control is not a warning in a policy. It is a rule that says which categories of work require a named person to have checked the citations against the source, and an understanding across the firm that checking with another AI tool does not count.
Confidentiality, privilege and what a public tool does with an upload
SC Gen 23 warns that prompts entered into a generative AI program may, unless disabled, be added to the model's database and used to respond to queries from other users, and that public chatbots may lack adequate safeguards to preserve confidentiality, privacy or legal professional privilege.
Where firms actually get exposed
- A file uploaded for summarising, containing far more than the question required.
- Personal accounts used for matter work, outside the firm's control entirely.
- A transcription tool in a client meeting nobody assessed.
- Discovery material put through a tool to speed up review.
- A browser extension with read access to whatever is on screen.
- An AI feature switched on inside the practice management system by a vendor update.
What a governed position looks like
- A short list of approved services, on firm accounts, assessed and recorded.
- An explicit rule on subpoenaed and restricted material, matching the practice note.
- Confirmation, in writing from the provider, that inputs are not used for training.
- A verification rule for anything going to a court or a client.
- Every AI account inside the firm's identity system, with MFA and offboarding.
- A record of who was told, and when.
Privacy law reaches most firms as well
A law firm with turnover above $3 million is an APP entity in its own right. Below that, the small business exemption may apply, but it is worth checking against the exceptions rather than assuming, and it does nothing at all about the duty of confidentiality, which is what actually governs client material. Firms also handle tax file numbers in some matters, and TFN obligations apply to any recipient regardless of the exemption.
Where AI turns up in practice, and what each raises
| Use | What it raises |
|---|---|
| Legal research | Fabricated authority. Note that SC Gen 23 treats dedicated legal research software searching legislation, judgments and legal literature differently from a general chatbot. |
| Drafting correspondence and advice | Client information leaving the firm, and accuracy in something going out under a practitioner's name. |
| Document review and discovery | Volume. Also the Harman undertaking and any material produced under compulsion. |
| Chronologies and indexes | Expressly permitted by SC Gen 23, subject to the confidentiality conditions on restricted material. |
| Written submissions | Permitted, with the verification obligation attached and the requirement that verification is not done solely by AI. |
| Affidavits and witness statements | Prohibited for generating content, with a positive disclosure required in the document. |
| Expert reports | Not without prior leave of the court in NSW, with specific disclosure and record-keeping conditions if leave is granted. Practitioners must draw the requirements to experts' attention when instructing them. |
| Client meeting transcription | Privileged conversation captured and processed by a third party, usually adopted with no assessment at all. |
| Billing, marketing and administration | Lower risk, and still your data leaving the firm. Worth being on the register even where it is routine. |
A firm with the right rule and none of the record
A fourteen-person suburban firm
The firm's AI policy prohibits entering confidential client information into AI tools. It was adopted after a partner read about the practice note. Staff use three generative AI accounts between them, two personal, mostly for drafting and summarising.
What exists
- A policy with a clear and correct prohibition
- Partners who understand the citation risk
- A sensible instinct about privileged material
- Something to say at the indemnity renewal
What does not
- No approved service, so the rule has no permitted alternative
- No acknowledgement record from any staff member
- No training, so nobody was told a matter summary is client information
- No assessment of any of the three accounts
- No basis on which to sign the affidavit disclosure with confidence
- No review since adoption and no owner for one
The last gap in that column is the one that would matter first. A disclosure in an affidavit that generative AI was not used in generating its content is a statement to the court. A firm that has not decided what tools exist, told its staff, or recorded that it did, is making that statement on the basis of hope.
What a firm should be able to produce
- The approved AI services, with who approved them and on what basis.
- Provider assessments covering training use, retention, processing location and subprocessors.
- Written confirmation on training use for any tool that touches client material.
- The rule on restricted material, matching the practice notes that apply to your matters.
- The verification rule, saying which work a named person must check and how.
- Staff acknowledgements, with dates, from everyone including contractors.
- Training completions by name, covering citations, confidentiality and uploads.
- Instructions to experts drawing the practice note requirements to their attention.
- AI accounts inside the firm's identity system, with MFA and offboarding evidence.
- Incident records, including breach assessments where personal information was involved.
- A review date and owner for the policy and the approved list.
The firm has the judgement. What it usually lacks is the record.
Nobody needs to explain confidentiality to a law firm. What tends to be missing is the dated record that the decision was made, that everyone was told, and that someone has looked at it since.
Policies with acknowledgement
An AI Governance Policy generated from the firm's own answers, with an approver, an audience, named acknowledgements and an assigned review owner.
Providers on the register
Each AI service recorded with data types received, storage country, offshore disclosure, subprocessor visibility, contract status, an owner and a review date.
An exportable position
A dated Evidence Pack covering control posture, policies, training by role, registers, open gaps and the chronology, for the indemnity renewal or a client's supplier questionnaire.
Cleverer does not review matter files, monitor AI use or connect to your practice management system, and it does not advise on your obligations to a court. It holds the compliance record around the decisions the firm makes.
Could your firm sign the affidavit disclosure with confidence?
It is a statement to the court about what tools were used, which means somebody has to know. If the honest answer is that nobody currently does, the Readiness Check is a short way to find out what else sits in the same position.
Questions from Australian law firms
Can Australian lawyers use generative AI?
Yes, within limits that differ by jurisdiction. The NSW Supreme Court's Practice Note SC Gen 23 expressly permits generative AI for chronologies, indexes, witness lists, briefs, summarising documents and transcripts, and written submissions subject to verification. It prohibits its use in generating the content of affidavits, witness statements and character references, and requires prior leave for expert reports. Check the current requirements of the courts you appear in.
Can client documents be uploaded to an AI tool?
The duty of confidentiality applies whatever the tool. SC Gen 23 sets a specific condition for material under suppression orders, the Harman undertaking, subpoena or a statutory publication prohibition: it must not be entered into a generative AI program unless the practitioner is satisfied the information stays within the controlled environment of the platform, is used only for that proceeding, and is not used to train any model. That is a workable standard for client material generally.
Does an affidavit need a statement about AI?
Under SC Gen 23, affidavits, witness statements and character references must contain a disclosure that generative AI was not used in generating their content, including by altering, embellishing, strengthening, diluting or rephrasing a witness's evidence. A firm needs to know what tools its people use before anyone can sign that with confidence.
Can we verify AI-generated citations with another AI tool?
Not under SC Gen 23. Where generative AI has been used in preparing written submissions, the author must verify that citations and references exist, are accurate and are relevant, and that verification must not be carried out solely by using a generative AI tool. Practitioners are also reminded that using AI does not qualify or absolve any professional or ethical obligation to the court.
Does the Privacy Act apply to our law firm?
If annual turnover exceeds $3 million, yes. Below that the small business exemption may apply, though it is worth checking against the exceptions rather than assuming, and tax file number obligations apply to any recipient regardless. In practice the exemption changes little for a law firm, because the duty of confidentiality already covers the material that matters.
What should we tell experts we instruct?
SC Gen 23 requires legal practitioners and unrepresented parties to draw the practice note's requirements to the attention of experts when instructing them. Generative AI must not be used to draft or prepare the content of an expert report without prior leave, and where leave is granted the expert must disclose what was prepared using it, which program and version, and keep a record of how it was used.
