AI and the duty of competence: what Australian lawyers must do
A solicitor asks a general purpose chatbot for authority on a point of unfair dismissal. It returns four cases, properly formatted, with plausible medium neutral citations and a paragraph of reasoning for each. Three of them do not exist. The fourth exists but says close to the opposite. Nothing on the screen distinguishes them, because the tool is not retrieving anything. It is predicting text that looks like a citation, and it does that job perfectly.
Australian regulators have not written a rule that says do not do that. They have not needed to. The duty to deliver legal services competently and diligently, the duty not to mislead the court, and the duty of confidentiality were all drafted without reference to any particular technology. A fabricated authority in a filed submission is a breach whether it came from a language model, a stale precedent bank or a rushed paralegal on a Friday afternoon.
What has changed is the speed at which a plausible error can reach a court, and the fact that courts are now asking about it directly. This piece sets out which obligations actually bite, what the superior courts have put in writing, and the verification workflow that holds up when someone asks you to justify it.
The rule you are looking for already exists
Practitioners keep waiting for a dedicated AI rule to tell them where the line is. The line is already drawn, in instruments most of us read once during admission and have not opened since. Rule 4.1.3 of the Australian Solicitors' Conduct Rules requires you to deliver legal services competently, diligently and as promptly as reasonably possible. Rule 9 governs confidentiality. Rule 19.1 prohibits deceiving or knowingly or recklessly misleading a court. Rule 37 makes you responsible for supervising the legal services delivered under your name. Barristers sit under an equivalent set of obligations in the Uniform Conduct (Barristers) Rules.
None of those rules contain a technology exception, and that is deliberate. The standard is what a competent practitioner would do, and competence has always included understanding the reliability of the sources you rely on. If you would not sign a submission drafted from a photocopied textbook of uncertain vintage without checking it, the analysis for a language model is identical. The tool changed. The standard did not.
The practical consequence is uncomfortable. There is no safe harbour available by saying you did not realise the tool could invent things. Competence includes knowing what your tools do.
What the courts have put in writing
Australian courts moved from silence to formal guidance faster than most practitioners noticed. The pattern has been consistent: restrict the use of generative tools in material that must be a witness's own account, require disclosure in defined circumstances, and remind practitioners that existing duties already apply to everything else.
The critical detail is that the obligations are jurisdiction specific. A practice note in one Supreme Court does not tell you what the tribunal down the road expects, and appearing across state lines means checking more than one. Reading a summary of a practice note is not the same as reading it. The restrictions on affidavits, witness statements and expert reports are drafted narrowly, and the narrow drafting is the part that matters.
Not every AI task carries the same risk
Firm wide bans and firm wide enthusiasm both fail for the same reason: they treat every use as one thing. The risk is not evenly distributed. It concentrates in a specific place, where the model is asked to produce something you cannot immediately check against a source sitting in front of you.
Sorted honestly, the risk ladder is unsurprising. The most dangerous request is the one that feels most useful, which is asking a general model for authority you do not already have. The safest is the one that feels least impressive, which is restructuring your own material while a lawyer reads every line.
Competence is a workflow, not an opinion about AI
Whether you personally think these tools are useful has no bearing on your obligations. What discharges the duty is a repeatable process that produces evidence. If you cannot describe the steps you took, you have no answer to a court, a client or an insurer who asks.
The workflow below adds minutes, not hours, and it is the difference between an efficiency gain and a disciplinary referral.
Supervision, costs and the awkward questions
Most firms have a shadow AI problem rather than an AI policy problem. Fee earners are already using consumer tools on personal accounts, quietly, because nobody has told them what is permitted and they suspect the answer is nothing. That is the worst arrangement available: all of the risk, none of the oversight, and no record of what went where.
There is also a costs conversation nobody wants to have. If a task that was billed at three hours now genuinely takes twenty minutes, your costs disclosure and any fixed fee arrangement need to reflect what actually happened. Charging for time not spent is a separate problem from anything to do with technology, and it does not become acceptable because the efficiency came from software you paid for.
- Ask each fee earner what tools they are already using, including personal accounts on personal devices.
- Find out whether client material has been pasted into a consumer service, and deal with it as a confidentiality issue if it has.
- Read the current practice note for every court and tribunal your practice appears in.
- Name the approved tools and the specific tasks each may be used for. A list of prohibitions with no permitted path drives use underground.
- State plainly that verification sits with the author, not the reviewer.
- Set a clear rule on confidential material: what may go into a prompt, what may not, and who decides the borderline cases.
- Add the file note requirement to your matter management checklist so it happens by default.
- Spend CPD units on this. Ten units a year across four fields, with at least one in ethics and professional responsibility and one in practice management, and this material fits squarely in both.
- Brief your professional indemnity insurer at renewal rather than at claim.
- Settle your billing position and update costs disclosures where the time estimates no longer reflect the work.
- Review the policy against any new practice notes, because the ones issued so far will not be the last.
What competence looks like from here
Abstention is a legitimate choice today and it will get harder to sustain. The duty of competence is measured against what a reasonable practitioner would do, and that benchmark moves with practice. It has moved before, from paper reporters to electronic databases, and nobody now argues that refusing to use a search function is a defensible way to run a research task. The same drift will happen here, slowly, and then all at once when a costs assessor first asks why a matter took as long as it did.
The useful question is not whether to use these tools but what to demand of them. Ask whether the tool searches actual Australian primary sources or composes an answer from training data. Ask whether every proposition links to a judgment or a section you can open in one click. Ask which jurisdictions and which date range it covers, and what the provider does with the text you enter. Purpose built Australian tools, Legal Brain among them, are designed around retrieval from primary sources for exactly this reason: not because it removes the verification duty, but because it shortens the distance between an answer and its source enough that verifying is realistic at volume.
The practitioners who will have the least trouble over the next few years are not the ones with the strongest views about artificial intelligence. They are the ones who can open a file, point to a note, and say precisely what they checked and when.
Frequently asked questions
Do I have to tell the court if I used AI to prepare a document?
It depends on the court. Practice Note SC Gen 23 in the Supreme Court of New South Wales requires disclosure in defined circumstances and restricts the use of generative AI in affidavits, witness statements and expert reports. Other courts and tribunals have their own notices with different thresholds, and some rely entirely on your general duties. Check the practice note for the specific court you are appearing in before you file.
Can Australian solicitors use ChatGPT for legal research?
No conduct rule prohibits it outright, but a general purpose model does not search Australian case law. It generates text that resembles a correct answer, which is why citations that do not exist appear formatted perfectly. If you use one, every authority must be opened and read in full, and no confidential client information should go into the prompt.
Who is liable if an AI tool produces an error that reaches the court?
The practitioner whose name is on the document. The duty not to mislead the court under rule 19.1 of the Australian Solicitors' Conduct Rules attaches to you, and rule 37 makes you responsible for supervising work delivered under your name. Vendor terms of service do not transfer a professional obligation, and no Australian court has accepted the tool as an explanation.
Does learning about AI count towards my CPD requirements?
Yes, where the activity meets the usual CPD criteria. Practitioners in Uniform Law jurisdictions need ten CPD units each year with at least one unit in each of ethics and professional responsibility, practice management and business skills, professional skills, and substantive law. Sessions on AI obligations and verification duties generally fit within ethics or practice management.
Two quick questions
No score is stored. Pick an answer to see why it is right.
-
1A solicitor files submissions containing a case that does not exist, generated by an AI tool. Which duty is most directly breached?
Rule 19.1 of the Australian Solicitors' Conduct Rules prohibits deceiving or knowingly or recklessly misleading a court. Filing an authority you never verified engages the recklessness limb, and the duty of competence under rule 4.1.3 is breached alongside it. Lack of intention is not an answer where the check was never performed.
-
2Which AI use carries the highest risk to an Australian practitioner?
A general chatbot holds no index of Australian judgments, so it composes text that looks like authority rather than retrieving it. The other three tasks all involve source material the practitioner already has in front of them, which means the output can be checked in minutes. Risk tracks how easily you can verify against a real source.
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.