Your client's chatbot transcript is discoverable. That is the least of it.
Five predictions for New South Wales practice over the next two years, with the evidence for each and the best argument against. The author discloses a commercial interest in legal AI.
Three weeks ago the Supreme Court appointed a receiver to sell a profitable defence contractor because two of its three owners asked ChatGPT how to get rid of the third. Every document they produced became an exhibit. One meeting became eight hearing days and a company sold out from under its owners. In the matter of Lanmar Pty Ltd (No 2) [2026] NSWSC 800 rewards reading in full. The lesson is narrower than "be careful with AI": your clients are already using these tools on the decisions that end in litigation, and they are not telling you.
One. Our first serious fight about privilege
The Americans got there in February and immediately disagreed with each other.
In United States v Heppner, Rakoff J held that a defendant's exchanges with a consumer AI tool were neither privileged nor work product. The tool is not a lawyer, and its terms disclosed that inputs train the model, so there was no reasonable expectation of confidentiality. He left one door open: had counsel directed the use, the tool might have been the lawyer's agent. The same day, in Warner v Gilbarco, a self-represented plaintiff's ChatGPT queries were protected work product, because work product is waived only by disclosure to an opponent.
Nothing here yet. When it comes it runs through ss 118, 119 and 122 of the Evidence Act 1995 (NSW), rule 9 of the conduct rules, and APP 8 if the tool sits offshore. The technology will not decide it. The deployment terms will: consumer or enterprise, and whether the provider trains on what you type. Practice Note SC Gen 23 has warned about this since February last year.
Which is the practical case for a model on your own hardware. If it never leaves the building, the argument never starts.
Two. Legal tech keeps struggling in the middle of the market
Not because the products are bad, but because the capability is bundled into a subscription the firm already pays for.
The Victorian regulator's census makes the point : among Victorian lawyers using AI, general-purpose tools like ChatGPT, Claude and Gemini beat purpose-built legal products. Those are lawyers picking a twenty-dollar-a-month subscription over software built for their own industry. Bundled, not free, and the pricing gap is the mechanism.
The vendors have conceded the central point. Harvey abandoned its proprietary legal model after frontier models beat it on Harvey's own benchmark, then published the benchmark as open source. When the category leader stops competing on the model, it has told you where the value went.
Two places the prediction fails. At the top of the market may be untrue: Harvey is estimated at around three hundred million dollars in recurring revenue with half the Am Law 100 on the platform. And no chatbot runs a trust account or holds the matter file, so LEAP, Smokeball and Actionstep are safe, as are Westlaw and Lexis for as long as they own the authority you are permitted to cite.
What is exposed is the middle: the point solution whose real pitch is that it writes the prompt for you, which is what most Australian small firms were sold last year. The replacement is not another vendor. Legal-OSS indexes around 290 open-source legal projects, with categories for local models, plugins and connectors. One calls itself a Cowork alternative for lawyers. Another is a case law server that plugs straight into Claude. A thin wrapper cannot beat a free plugin doing the same job inside software the firm already licenses.
Three. The judiciary asks us for help, and not in a speech
Bell CJ has largely asked. To the Australian Bar Association last August he counted at least 23 further Australian cases of unsatisfactory generative AI use since Nash v DPP (WA) in 2023, said the difficulty is most prevalent among unrepresented litigants, and warned the verification burden falls hardest on the courts lowest in the hierarchy, which are already swamped. At his December 2024 briefing he had said of the digital divide that the court's remit did not legitimately extend that far.
The courts are absorbing a problem they did not create and cannot fund, and they are telling us whose problem they think it is. My call is expanded duty schemes, "triage tech", firmer pro bono expectations, and at last a serious look at unbundled services.
Some judges disagree, fairly. American judges say AI-assisted filings are easier to decide, because a litigant who could not frame a claim now can. The complaint is about volume, not quality. Treat the widely quoted 18 per cent figure for AI-generated filings carefully: it comes from running 1,600 sampled complaints through a commercial detector, and those produce false positives.
Four. The generational handover decides adoption, not the business case
Women have been the majority of Australian solicitors since 2018 and are 56 per cent of the profession. A senior, male-skewed cohort of principals is leaving, their practices are hard to sell, and succession keeps going into the too-hard basket.
Adoption tracks career stage and little else. The 2025 Victorian census found 36.7 per cent of respondents using AI in practice, rising to 48.7 per cent of the under-30s and falling with age and years since admission. Barristers were on 14.6 per cent. As the cohort that adopts least retires and the cohort that adopts most inherits the files, use stops being a decision and becomes a condition.
Two things cut against it. The gender gap is real and has stopped closing: Harvard Business School work across more than 100 countries puts adoption at 47.8 per cent of men against 39.3 per cent of women, stable since early 2025 and unexplained. And the people adopting are not the people signing cheques, with women holding roughly 31 per cent of equity partnerships. If uptake shifts at the bottom while capital decisions stay at the top, firms keep buying the wrong tools.
Five. The precedent folder becomes a knowledge graph, and the law is against you
Lawyers have always moved firms with their precedents, mostly lawfully, because Wright v Gasweld protects a departing employee's general skill and knowledge even where it does not protect the firm's confidential information. In two years what a good practitioner carries will not be a folder. It will be a retrieval system over their own work, a set of connectors, and a library of instructions they spent a year refining.
Here is what everyone gets backwards. You are worse off than you were with the folder. An index built on firm systems, in firm time, over client files is very likely the firm's confidential information and probably its copyright under s 35(6) of the Copyright Act 1968. Client material does not stop being client material once it is converted into numbers. A folder of precedents was mostly yours. A graph of your matters is mostly not.
No Australian authority yet, and American disputes on foot about departing employees and their former employers' systems. So the question gets answered by whatever your employment contracts say when the first one lands. They say nothing. That is cheap to fix this year.
I would rather be argued with than agreed with. If I have called one of these wrong, say so.
Jacob Carswell-Doherty is the principal of Jake McKinley, a Sydney litigation firm, and the founder of Gleg, which builds AI tools for Australian legal work.