ChatGPT put a capable language model in front of every lawyer in the country. It drafts a clause in seconds, summarises a long email, and answers a plain-English question about a legal concept without complaint. For a lot of everyday tasks, that is useful.
The problem starts when the task is legal research on Australian law, and the output has to survive a judge reading it. That is a different job, and it is the job Habeas was built for.
This piece sets out where the two tools diverge across research, drafting and document review, what the Australian courts now expect, and how to think about which tool belongs in which part of your workflow.
ChatGPT is a general-purpose model that generates fluent text from patterns in its training data. Habeas is a purpose-built legal AI that runs an agentic research process over a curated corpus of Australian case law, legislation and commentary, then returns answers grounded in those sources with citations you can open and check.
A general model predicts a plausible answer. A grounded system retrieves the authorities first and reasons from them. That distinction shapes everything below.
It can produce something that reads like legal research. Whether you can rely on it is the harder question, and recent Australian cases have answered it in public.
A general model has no live, verified index of Australian judgments. When you ask it for authority on a point, it composes an answer that fits the pattern of legal writing. Sometimes the cases are real. Sometimes they are not, and the model presents the invented ones with the same confidence as the genuine ones. This is the well-documented "hallucination" problem, and in law it is not a cosmetic flaw. A fabricated citation in a submission is a false statement to the court.
That is no longer theoretical in Australia:
The pattern in each is the same. The tool produced fluent, confident output. Nobody caught the fabrication before it reached the court. The consequence landed on the practitioner, not the software.
This is the sharpest practical divide.
Ask a general chatbot for supporting authority and you get a case name and a proposition. To trust it, you have to independently locate the judgment, confirm it exists, confirm it says what the model claims, and confirm it is still good law. In effect, the model hands you unverified research and the verification burden falls entirely on you.
Habeas is built so the authority comes first. Answers are grounded in the underlying sources, with citation observability down to the subparagraph level, so you can trace a proposition back to the specific passage it rests on rather than taking the output on faith. The research still needs a lawyer's judgment applied to it. The difference is that you are checking real, retrievable authority instead of hunting for whether an authority exists at all.
For a barrister forming a position on a brief, or an in-house counsel giving fast advice a stakeholder will act on, that gap between "plausible" and "traceable" is the whole ballgame.
There is a second divide that has nothing to do with quality of output and everything to do with risk.
Consumer and free-tier general chatbots may retain what you type and use it to improve their models. Australian courts have been explicit that practitioners cannot safely enter confidential or commercially sensitive material into tools where that data may leave a controlled environment or be used for training. Material subject to the implied undertaking or produced on subpoena is squarely in scope.
The NSW Supreme Court's Practice Note SC Gen 23, in force since 3 February 2025, sets the current baseline: AI-generated citations must be independently verified, generative AI must not be used to produce the content of affidavits or witness evidence, and confidential material must stay within a controlled environment. Similar guidance now exists across Victoria, Queensland, South Australia and other jurisdictions, and the list of court protocols keeps growing.
Habeas is built for this reality. Infrastructure is Australian-hosted, firm and matter data is not used to train models, and access is controlled through SSO, role-based permissions and per-firm domain policy. The product is designed to sit inside the professional obligations, not in tension with them. A general-purpose consumer chatbot was never designed against those constraints, because it was never built for legal practice in the first place.
Both tools draft. The question is what they draft from.
A general model drafts from its training patterns. It will produce a competent generic clause or a serviceable letter, but it does not know your firm's precedents, your established positions, or the authorities that actually govern the point.
Habeas approaches drafting as a legal task. Magic Draft produces first drafts of advice, submissions, letters and clause work drawing on matter context and the underlying authorities in a single pass, and Precedent Bank brings a firm's own precedents into the loop so the output reflects house style and prior thinking rather than a generic template. For a senior lawyer, that collapses the mechanical part of the first draft and keeps the work on judgment and argument. For the firm, it captures knowledge that otherwise lives only in individual lawyers' heads.
Uploading a long document to a general chatbot and asking for a summary works for low-stakes material. It does not scale to matter work, where you need cross-document reasoning across pleadings, contracts, discovery and correspondence, with the analysis tied back to specific sources.
Habeas Document Stores ingest matter materials and support analysis at volume: building litigation chronologies, testing the strength of a case, comparing clauses across a contract portfolio, surfacing inconsistencies across witness materials, and finding authority buried inside the matter file itself. This is high-volume, high-stakes work that would otherwise consume significant junior and paralegal hours.
Yes, and pretending otherwise would be dishonest. For brainstorming a structure, rewording an internal note, explaining an unfamiliar concept in plain terms, or handling non-confidential administrative text, a general chatbot is fast and perfectly adequate. Plenty of lawyers use one that way every day.
The line to hold is this: the moment the task involves confidential material, Australian authority you will rely on, or output that goes to a court or a client, a general-purpose model is the wrong instrument. Not because it cannot write, but because it was never grounded in the law and never built against the obligations that govern the work.
That is the space Habeas occupies. Not a chatbot that happens to talk about law, but a research, drafting and strategy platform built on Australian legal sources, designed so the authority is real and the data stays where it should.
Is ChatGPT reliable for legal research in Australia?
Not on its own. It has no verified index of Australian judgments and can fabricate citations that read as genuine, a problem that has already led to sanctions and cost orders against Australian practitioners. Any legal citation it produces must be independently verified before it goes anywhere near a court.
What makes Habeas different from ChatGPT?
Habeas is grounded in a curated corpus of Australian case law, legislation and commentary, returns answers with citations you can trace to the source, is hosted in Australia, and does not use firm or matter data to train models. It is built for legal work rather than adapted to it.
Can I put confidential client information into ChatGPT?
Australian courts have warned against entering confidential or commercially sensitive material into tools where data may be retained or used for training. Practice Note SC Gen 23 and equivalent guidance in other states set out the expectations. Habeas is designed to keep that material within a controlled, Australian-hosted environment.
Does Habeas still require lawyer verification?
Yes. No AI removes professional responsibility, and the court rules make the practitioner accountable for what is filed. The difference is that Habeas gives you real, retrievable authority to check rather than leaving you to work out whether a cited case exists at all.
Curious how Habeas handles your kind of work? See the platform in action at habeas.ai or ask us for a walkthrough on a real research task.
