AI in family law, Mid-2026

On 16 April 2026 the Federal Court issued practice note GPN-AI, requiring the practitioner responsible for a document to confirm that legal authorities cited in it exist and support the propositions stated, alongside the facts and evidence relied on. NSW had already prohibited generative AI for affidavits and expert reports without leave from February 2025. Victoria, Queensland, and South Australia have parallel notes and guidelines in force. In late 2025, in Mertz & Mertz (No 3) [2025] FedCFamC1A 222, the Federal Circuit and Family Court ordered a solicitor to pay $10,000 in costs thrown away after AI-fabricated citations were filed in a list of authorities, and referred the solicitor, along with two counsel, to their state conduct regulators; the disciplinary process that referral triggers is still pending, not concluded. A community-maintained database now catalogues dozens of further incidents internationally.

The courts have therefore answered, in unambiguous terms, the question of what is required when AI is used in legal work. That answering is the substantive shift of the past twelve months and the frame in which any other question about legal AI in 2026 has to sit. The marketing from the legal AI vendors has not yet caught up to that frame, which is the operational problem this note is about.

This is for principals of family law practices and the senior solicitors who set the operational tone inside them. The deeper evidence base for everything below is in the State of AI in Mid-2026 literature review, which surveys ten capability categories across applied AI and went through three independent fact-check passes before publication. This piece pulls the threads that matter to family law specifically.

What’s stable in family law

A family law matter still depends on the practitioner understanding the client, the law, the relevant facts, and the procedural posture, and producing material in which every factual assertion can be sourced and every legal proposition can be defended. That has been true for a long time and will be true for a long time yet. The product of the work is read by opposing solicitors who are paid to find errors, by judicial officers operating under time pressure, and by clients whose lives turn on the document. The practitioner’s name appears on it. The professional obligation attaches to that name and to no other.

The procedural texture is also stable. Affidavits stand or fall on the precision of their factual recital. Financial statements require source documents. Submissions are persuasive in proportion to the quality of their authorities. Consent orders, parenting plans, and binding financial agreements need to read the way the courts expect them to read. None of that is a domain in which a system that occasionally makes things up is a substitute for a person who does not.

The shape of the regulatory environment is now also stable in the sense that it has resolved. The questions of “is AI use in legal work permitted” and “what are practitioners obliged to do about it” have moved from “open” to “answered.” The operational consequences flow from that.

What’s changing

The leading legal AI products are now serious commercial entities. Harvey raised US$200 million at an US$11 billion valuation in March 2026; the firm said in August 2025 it had reached 42 per cent of the AmLaw 100 and by the March 2026 raise described its customer base as a majority of that cohort. Thomson Reuters announced one million professional users of CoCounsel in February 2026. These are not pilots. They are commercial products being used at scale by sophisticated firms, and they have meaningful capability on legal research, document review, drafting support, and discovery. A senior practitioner who has not formed a view on them is now visibly behind a peer who has.

The independent evidence on hallucinations is no longer in doubt. The preregistered Stanford RegLab study, published in the Journal of Empirical Legal Studies in 2025, found Lexis+ AI hallucinating in 17 per cent of responses and Westlaw’s AI research product in 33 per cent, both marketed at the time with hallucination-free language. The figures are not from a vendor’s internal benchmark. They are from a peer-reviewed study by an institution with no commercial interest in the result. Anyone making purchasing or workflow decisions on the basis of vendor accuracy claims, without reference to this study, is taking on more risk than they realise.

The Australian courts have made verification a professional obligation. GPN-AI is unambiguous on this; so are the practice notes and guidelines in NSW, Victoria, Queensland, and South Australia. The position is no longer “verify your AI output because it is good practice.” It is “verify your AI output because the court has told you to.” The Mertz & Mertz costs order and referral in 2025 showed the courts and regulators are prepared to act on that obligation, even though the disciplinary process it triggered is still working through.

AUSTRAC Tranche 2 is a parallel obligation, not a related one, and it lands on some of the same desks. From 1 July 2026, the AML/CTF regime expands to cover services AUSTRAC treats as “designated services,” and legal practices are among the professions brought inside it. Whether and how that captures a given family law practice depends on the exact services it provides, which is a question for the practice’s own AML advice rather than a call this piece is positioned to make. What is clear is that the technology-and-compliance load on a principal is stacking, and the firms that have built deliberate internal infrastructure to support it will navigate the period more cleanly than the firms that treat each new obligation as a separate fire to be put out.

Knowledge synthesis at the firm level is now technically feasible in a way it was not a year ago. Current frontier models can hold an organisation’s accumulated material in working memory at a scale that meaningfully changes what a junior practitioner can do with a question. A boutique firm with twenty years of file notes, submissions, draft consent orders, and matter outcomes can build a tool that lets its solicitors ask the firm’s accumulated experience a question in plain English and receive an answer grounded in the source documents. The best independent evidence still shows leading grounded summarisation systems hallucinating in roughly 13 per cent of responses on this task, which is the same fact pattern as everywhere else in legal AI: the capability is real, the verification step is non-negotiable, and the value lives in the documents the firm already has.

Where it lands

AI fits the practice where verification is built into the routine rather than left to the individual practitioner’s discretion, and where the final document still carries a verified human signature. Several places fit that shape.

Initial research and authority retrieval, with the practitioner reading the underlying decision before citing it. First-cut drafting of routine documents (initiating applications, response materials, financial statement schedules, correspondence) with the practitioner editing rather than approving wholesale. Document review across large discovery sets, with sampling controls and a human pass on anything load-bearing. Client communication for routine progress updates and document requests, where the firm’s voice and the matter’s posture are templated and consistent. Internal knowledge retrieval, where the firm’s prior matters, file notes, and methodology become legible to the practitioners working today. Translation between legal and lay language for client-facing material, with the practitioner verifying the legal precision has survived the translation.

What unites the list is what the firm retains. The practitioner’s name on the document. The verification of every authority before citation. A workflow in which AI output never reaches the court without a human pass.

Where it doesn’t yet land

And several places where the marketing has run ahead of what the courts have just made enforceable.

Any court-facing material in which authorities have not been personally verified against their primary source; GPN-AI is unambiguous on this and the list of practitioners referred to conduct regulators over AI-fabricated material is growing. Any affidavit or expert report material in NSW prepared with generative AI assistance without leave of the court; that is now a direct breach of the Supreme Court’s practice note. Any reliance on a legal AI product’s internal accuracy claims without reference to the Stanford RegLab evidence; the 17 per cent and 33 per cent numbers from a peer-reviewed study are the more credible baseline.

Anything involving children’s matters or family violence material where the system’s error profile has not been examined. The dominant failure mode of grounded summarisation across the literature is omission rather than fabrication, and an omitted fact in a family violence context is not a recoverable error. Any vendor whose published evidence rests on a single secondary source or an unspecified internal benchmark; that is a procurement red flag rather than a technology view. Anything that would otherwise be a candid file note about strategy, opposing counsel, or client conduct entered into a tool whose data handling and retention posture has not been verified at the firm level. The lit review walks through the broader pattern of perception-versus-measurement across office AI; in legal work the consequences of getting that gap wrong are more direct than in most.

What this means for the operator’s posture

What the courts have written into practice notes, a firm has to operationalise. Treat AI as a layer that thickens the firm’s accumulated experience and accelerates the assembly of inputs, not as a layer that produces court-facing material without verification. Keep the practitioner’s name on every document, and keep the workflow consistent with that. Build the verification step into the routine, not into the individual practitioner’s discretion. Make the firm’s own material more useful to its own people before reaching for tools that promise to replace any part of the work the courts and the regulators have identified as professional.

Be specific, internally, about which tools are in use, on what kinds of matters, with what verification routine attached. Be willing to make that specificity visible to clients and, if asked, to the court. The firms that come through the next three years in the strongest position will be the ones that built the verification routine early, made it legible to everyone in the firm, and used it as a mark of seriousness rather than a constraint to be minimised.

The frontier in legal AI is moving quickly enough that a senior practitioner does not need to be the first to adopt anything. It is moving slowly enough that anyone who waits eighteen months without forming a view will be visibly behind a peer who spent those eighteen months running small structured pilots and learning what to trust and what to verify a second time.

The question worth working through is not “which legal AI product should we license.” It is “what is the verification routine that sits around any AI-assisted output in this firm, how would we describe it to a Federal Court judge, and what does our accumulated experience look like once every solicitor in the firm can ask it a question.” If those are the right questions, we are useful to talk to. We work with professional services practices on exactly this.