Strictly Educational — Not legal advice.

The Federal Court of Australia has issued a landmark practice note on generative AI. Here's what it says, why it matters, and what you need to do about it.

On 16 April 2026, Chief Justice D S Mortimer signed into effect the Use of Generative Artificial Intelligence Practice Note (GPN-AI) — the Federal Court of Australia's first comprehensive framework governing how artificial intelligence tools may be used in court proceedings.

This is not a gentle suggestion. It is binding guidance with real consequences. And it applies to everyone — not just lawyers.

What the Practice Note Actually Says

The GPN-AI opens with a notably balanced posture. The Court explicitly welcomes generative AI, recognising its potential to increase efficiency, reduce legal costs, and enhance access to justice. The named tools are familiar to anyone in the legal tech space: ChatGPT, Claude, Harvey, Google Gemini, Microsoft Copilot.

But the Court is clear that enthusiasm must be disciplined by responsibility. The note is structured around three core obligations and four high-risk areas.

The Three Core Obligations

  1. Competence — Any person using generative AI must have a "basic understanding of its capabilities, and its limitations and risks." This is a low bar deliberately set — the Court is not demanding technical expertise, but it is demanding informed use. Ignorance is not a defence.
  2. Non-interference with justice — AI use must not adversely affect the administration of justice. Existing legal and professional responsibilities are not suspended because a tool was involved. If anything, they are amplified. The note is explicit: there will be circumstances where using AI at all is inappropriate.
  3. Disclosure on demand — and proactively — If the Court requires it, you must disclose whether and how AI was used. But the note goes further: there are categories of use where disclosure is mandatory without being asked. All persons must be able to tell the Court what AI was used, how it was used, and for what purpose.

The Three High-Risk Areas

Pleadings, Submissions and Filed Documents

This is where the hallucination problem becomes a litigation crisis. The Court catalogues the known failure modes with uncomfortable precision: fictitious cases, non-existent citations, false quotes, incorrect legal analysis, and — pointedly — AI confirming wrong information when asked to check itself.

The consequences flow directly from existing duties. Presenting false information to the Court is not just an AI problem — it is a contempt problem, a professional conduct problem, and a costs problem.

Where AI has been used to draft or assist with filed documents, the responsible person (lawyer or self-represented litigant) must personally verify:

  • That pleaded facts are based on what can actually be proved
  • That cited authorities exist and support the proposition stated
  • That cited evidence exists, is before the Court, and is likely admissible
  • That factual inferences drawn are reasonably open
  • That chronologies are accurate
  • That document lists conform with the Federal Court Rules 2011

The note makes clear this is not an exhaustive list. The human remains responsible. Full stop.

Affidavits, Expert Reports and Evidentiary Materials

This section carries the most serious implications. When a person swears or affirms an affidavit or witness statement, they are legally representing that it reflects their own recollection, knowledge and experience. AI-assisted drafting that blurs or displaces that authorship is not just a procedural problem — the note flags criminal laws prohibiting the falsification of evidence.

For expert witnesses, the stakes are equally high. An expert's overriding duty is to the Court, not the client. Their report must contain their own opinion and reasoning process. AI outputs substituted for expert analysis fundamentally compromise that duty.

Mandatory disclosure is triggered where AI was used to:

  • Summarise or analyse the information on which a witness relies
  • Create images, video, audio or other multimedia presented to the Court
  • Prepare materials in any other way that could affect admissibility or weight

Disclosure must appear at the start of the document, identifying where AI was used and how.

Confidential, Suppressed and Private Information

This section deserves particular attention from litigators working in commercial, IP, family, or any proceeding involving suppression orders.

The warning is stark: feeding information into a publicly accessible AI tool may make it available to others. You may not know where it is stored, how it is used, or who can access it.

The Court identifies four categories of information that must not be entered into AI tools without careful prior analysis:

  • Information subject to confidentiality, suppression or non-publication orders
  • Privileged information
  • Information subject to the implied undertaking (the Harman obligation)
  • Otherwise confidential or private information

Critically, the note addresses "closed" or enterprise AI tools as well. Even in a ringfenced environment, if outputs are later used for purposes beyond the original scope, the implied undertaking may still be breached. The Court is candid: achieving certainty about effective confidentiality controls "may be inconvenient or even quite difficult." That difficulty is a factor to weigh before using AI tools in litigation at all.

The Consequences Section Is Brief — and Deliberate

Section 5 is just one paragraph. It does not need to be longer. Where AI is used inconsistently with the practice note or any Court order, "all persons should expect that there could be consequences including adverse costs orders and issues as to compliance with legal and professional obligations."

Adverse costs orders in Federal Court proceedings can be substantial. For law firms, "issues as to compliance with legal and professional obligations" is a reference to regulatory and disciplinary exposure.

Why This Matters Beyond Australia

The GPN-AI reflects a global inflection point. Courts around the world are grappling with the same questions — and most are still watching and waiting. Australia's Federal Court has chosen to act.

Several features of this note are worth watching internationally:

  • It is technology-neutral. The note governs generative AI tools generically, not specific products. It will not require amendment every time a new model is released.
  • It applies to self-represented litigants. This is not a lawyers-only framework. Unrepresented parties — a growing cohort in every jurisdiction — are explicitly within scope, directed to the Litigants in Person Practice Note (GPN-LIP) for further guidance.
  • It acknowledges its own limitations. The Court openly states that given the pace of AI development, the note cannot remain exhaustive. It has published supplementary "Generative AI Resources" for both lawyers and non-lawyers, signalling ongoing revision as the technology evolves.

Key Takeaways for Legal Practitioners and Litigants

If You Are a Lawyer

  • Audit your AI-assisted workflows now. Every output needs human verification before it goes to court.
  • Build disclosure statements into your document templates for AI-assisted drafting.
  • Never enter discovery documents, affidavit evidence, or confidential information into a public AI tool without explicit risk analysis.
  • Review your firm's enterprise AI arrangements against the implied undertaking requirements.

If You Are a Litigant in Person

  • You can use AI to help understand your case and draft documents — but you are personally responsible for everything you file.
  • Do not cite a case you have not independently verified exists. Not in a submission, not in an email to the registry.
  • If asked by the Court whether you used AI, you must answer truthfully and specifically.

If You Are Building Legal Tech

  • Disclosure infrastructure is now a product requirement, not a feature request.
  • "Closed" AI environments need demonstrable confidentiality controls, not just contractual representations.
  • The Court has signaled it will be reviewing this note. Build for change.

From Practice Note to Practice: A Skills-Based Compliance Framework

Reading a practice note is one thing. Knowing what to do differently on Monday morning is another. One approach that illustrates how the GPN-AI can be operationalised is a skills file — a structured translation of each legal obligation in the practice note into a discrete, actionable compliance step.

A skills file built against GPN-AI decomposes the practice note into twelve mandatory skills, each mapped directly to a source paragraph in the note. The framework illustrates what genuine compliance actually looks like in practice.

What Is a Skills File?

A skills file is a structured reference document that translates a set of rules, obligations or standards into a format an AI agent can read, understand and apply consistently. It is not a prompt. It is not a chatbot instruction. It is a persistent, reusable compliance layer that travels with the AI's working context wherever it is loaded.

The core idea is simple: AI language models are highly capable at following structured instructions when those instructions are clear, specific and present at the start of a task. A skills file exploits this capability deliberately. Rather than asking an AI to "comply with the law" — an instruction so broad it is practically useless — a skills file breaks that obligation down into discrete, testable behaviours, each triggered by a specific condition.

The anatomy of a skills entry typically contains:

  • Applicability — the precise condition that triggers this skill (e.g., "when generative AI has been used to assist in drafting an affidavit")
  • Actionable step — the concrete behaviour required (e.g., "place disclosure at the start of the document, identifying where and how AI was used")
  • Strict prohibition — what must never occur (e.g., "must not present AI-generated content as the deponent's own knowledge")
  • Classification — whether the skill is mandatory or advisory
  • Source reference — the specific paragraph in the governing instrument that the skill implements

This structure matters because it mirrors the way a competent professional internalises rules. A lawyer who has read the Federal Court Rules does not recite them verbatim before every filing — they have internalised the rule into a habit of conduct. A skills file does the same for an AI agent: it converts a legal instrument into a set of conditioned behaviours that fire automatically when the relevant context arises.

How a skills file is used with an AI agent depends on the tool, but the principle is consistent across platforms. The skills file is loaded into the AI's context at the start of a working session — either as part of the system instructions, as a project-level document, or as a file the agent is directed to read before proceeding. The agent then treats the skills as active constraints on everything it does in that session. It does not need to be reminded. It does not need to be asked. The skills are already in scope.

This is the critical difference between a skills file and an ordinary instruction. An ordinary instruction tells an AI what to do in this conversation. A skills file tells an AI how to behave across every task within a defined domain — persistently, consistently, and in a way that can be audited because the file itself is a permanent, readable record of the governing rules.

For GPN-AI compliance, this means the AI does not just know the practice note exists. It knows which obligations are triggered by which actions, what it must do, what it must not do, and where to find the source rule if a question arises. The skills file is the practice note, operationalised.

Why a Skills File Matters

The skills file approach reflects something important about the GPN-AI as a regulatory instrument. The practice note deliberately frames obligations in general terms — "be guided by your existing legal responsibilities," "exercise caution," "be transparent." That flexibility is intentional; the Court cannot anticipate every configuration of AI use in every proceeding.

But general obligations demand personal translation. A lawyer or self-represented litigant who has only read the practice note knows they must comply. A person who has built a skills file against it has decided how they will comply, when each obligation is triggered, and what evidence they will be able to produce if the Court asks.

The practice note explicitly contemplates ongoing review. The skills file is a living document — each revision of GPN-AI should prompt a review of the skills mapped to it. That is the architecture of durable compliance in a rapidly evolving regulatory environment. For self-represented litigants in particular, this approach has immediate practical value.

The Bottom Line

The Federal Court of Australia has not banned AI. It has done something more sophisticated: it has insisted that the humans remain accountable. The AI is a tool. The signature on the document — whether a lawyer's or a litigant's — still belongs to a person, and with it, all the legal obligations that signature carries.

GPN-AI is available on the Federal Court of Australia website. It should be read alongside the Central Practice Note (CPN-1) and the Litigants in Person Practice Note (GPN-LIP).