Diese Seite ist noch nicht in Ihrer Sprache verfügbar.

← Alle Beiträge

GPN-AI Is Here: What the Federal Court's Generative AI Practice Note Means for Document Review

The Federal Court of Australia's generative AI practice note, GPN-AI, sets conditions on confidential material, verification and disclosure. Read as a specification, it gives litigation support teams the outline of an AI review protocol.

Claira-Team20.11.20266 Min. Lesezeit

The Federal Court of Australia issued its Use of Generative Artificial Intelligence Practice Note, known as GPN-AI, with a Notice to the Profession dated April 16, 2026. As the Hall & Wilcox litigation team reads it, the practice note applies to all persons who appear before, or file documents with, the Court. The Clifford Chance team spells out that this reaches witnesses, experts and third parties answering subpoenas as well as parties and their lawyers.

Its provisions on submissions and citations will matter most to counsel. For a litigation support team, the practice note reads just as well as a specification for document review: it sets conditions on the material that goes into a tool, on the checking that comes before anything is filed, and on the account of AI use a team may later have to give.

This story works through those conditions in order and shows where Claira, the AI document review built into Nuix Discover, fits. It is information drawn from the briefings linked below, and the text on the Court's own site is the authority for any matter.

Confidential, privileged and discovered material

The practice note starts from a practical risk. In the words Hall & Wilcox quote from it, "Information provided to a Gen AI tool may become available to others, as users may not know where that information is stored, how it is used, or who will have access to it." Privileged, confidential or private information must not be entered into a tool in a way that is inconsistent with the obligations applying to the use of that information.

The Piper Alderman team reads those obligations as covering court-ordered confidentiality, suppression and non-publication orders, privileged material, and the implied obligation not to use discovered information for a purpose other than the proceeding. They add that even a ringfenced tool can breach an obligation if its outputs are used for another purpose. Smokeball's guide to the court protocols notes that the risk of inadvertent disclosure may be lower for "closed" or "controlled" tools, while the Court still expects the confidentiality of information obtained through compulsory processes to be preserved.

For a review team, this means due diligence on data location, retention and model training before discovered material goes into any tool. For Claira, the help centre's privacy and security article documents the position. Each customer's data stays in the country of the deployment it uses, and Claira can keep a case's data in Australia. Claira sends each model request to a provider inside the deployment's country, documents, prompts and results are never used to train a provider's models, and providers do not log, cache or store content once they have returned a response. Claira keeps no copy of the document text or of the response, which lives in your Nuix Discover field. The article lists exceptions, such as a long bulk scan prompt kept with the provider until a run ends, and is worth reading in full.

Case Context, the structured briefing on a matter that lives in Claira, is visible to every user in the Discover case, so the help centre advises keeping privileged strategy commentary, work product and mental impressions out of it.

Verification and the list of documents

As Hicksons read clauses 4.5 to 4.7, the person responsible for a document must confirm that the legal authorities cited exist and support the propositions for which they are relied upon, and that evidence is before the Court and reasonably likely to be admissible. Piper Alderman's summary extends the confirmations to chronologies and to lists of documents conforming with the Federal Court Rules.

If AI coding contributes to the list a party serves, the responsible lawyer needs a basis for saying the list is right, and that basis comes from testing and sampling. The help centre recommends running a prompt on 10 to 25 documents with single scans before scaling it. A bulk scan then writes its answers to a Nuix Discover field, never overwriting a value already in it, so reviewers can sample the AI's calls like any other coding.

Privilege calls need the same discipline. The help centre's privilege review workflow describes Claira's analysis as an AI-assisted first pass and says every designation should be confirmed by qualified counsel before anyone relies on it.

Disclosure and the record behind it

Clifford Chance read disclosure as mandatory where AI tools have been used "to summarise, analyse or generate material that informs evidence or opinion (or its admissibility)", made at the start of the document. Piper Alderman list the elements as the tool, how and for what purposes it was used, and where in the document. Affidavits must still reflect the deponent's own knowledge, and an expert report the expert's own opinion and reasoning.

A review team rarely signs an affidavit, but its work often feeds one. As an illustration, a chronology built from the documents may end up informing a witness's evidence, and whoever writes the disclosure then needs an accurate account of the AI's part.

In Claira, that account is kept as the review runs. AI activity is each case's record of how AI was used in it: every scan, prompt version, Insight, Case Context change, Agent Mode chat and download, with the person, the time and the model for each. You open it from the mode menu under More, then AI activity, and it downloads as a PDF or as an Excel workbook listing every scanned document with its outcome. The record never contains AI answers about documents, field values or document content. Download it before a case's data is deleted, since it cannot be recovered afterwards.

Prompt History adds the detail behind each run: every version of a prompt, comments, a case-wide Final mark and exports that show what changed.

Turning the practice note into a review protocol

Read this way, GPN-AI gives a litigation support lead the outline of a written protocol for AI-assisted review in Federal Court matters. Before material goes in, record your assessment of the tool against the obligations on the material and keep it on the file. During the review, test prompts on documents counsel has already decided, mark the version the team settles on as final, and sample the AI's calls before they feed a list of documents or a privilege claim. At the end, download the AI activity record, so the person responsible can write any disclosure from it.

Our Casebook walkthrough of the C7 discovery in Seven Network v News shows that kind of review on a Federal Court record, from early case assessment through bulk scans to a chronology.

Where to start

GPN-AI leaves room for AI-assisted review, provided the obligations on the material are respected, the output is verified by the person responsible, and the use can be explained when the Court asks. A practical first step is to write the protocol above for one live matter and check it against the record your tool actually keeps. If you would like to see AI activity, Prompt History and a bulk scan on your own documents, you can book a session with our team.

Fragen zu diesem Beitrag?