It is 11pm. A GC at a mid-sized technology company has a board meeting at 9am and a question she cannot leave unanswered: whether the IP assignment clause in a term sheet signed eighteen months ago creates exposure under the acquisition structure they are about to present. She opens a general-purpose AI tool, pastes in the clause, pastes in the relevant commercial context, and types her question.
The answer comes back polished, confident, and structurally sound. She does not know whether the authority it cited exists. She does not know whether the platform retained the clause, the commercial context, the name of the counterparty she mentioned in passing. She does not know whether what she typed at 11pm will surface in some inference log she cannot inspect, or feed a fine-tuning pipeline that serves someone else tomorrow.
She does know the answer felt right. That is the problem.
Most general-purpose AI tools are trained on large, heterogeneous corpora scraped from the public web. They learn to predict plausible text, and they are extremely good at it. The outputs feel authoritative because they sound authoritative: well-structured, hedged in exactly the right places, carrying the cadence of something that has processed a great deal of law.
When a GC uses one of these tools to research a live matter, two things happen simultaneously. The model answers the question, and it receives everything the user supplied to ask it: the nature of the dispute, the commercial structure, the internal concern driving the query. On most consumer-grade platforms, this material does not vanish. It may be retained for inference logging, used to refine the model, or held in ways the user cannot examine or verify. The privacy terms of these tools were not designed for people whose information carries privilege.
This is where the hallucination risk and the data-handling risk meet. Both flow from the same design choice: a model trained on everything, connected to everything, with no bounded surface and no clear account of what is inside it. That boundlessness is what makes the tool feel powerful. It is also what makes it inappropriate for serious legal work.
A closed, licensed corpus changes the architecture in ways that matter structurally. When the research surface is defined and finite, the model cannot invent authorities from outside it. What is in the corpus is there; what is not, is not. You can ask where a cited case came from, and there is a real answer, traceable back to a legitimate source.
This is a security property as much as an accuracy property. The bounded corpus shrinks the attack surface. A model that cannot draw on the open web cannot reconstruct a fictitious argument from some distant fragment of training data, pattern-matching on enough real cases to make an invented one sound convincing. The confinement is what makes the output trustworthy, and confinement requires a deliberate design decision at the outset, not a patch applied after the fact. Patching accuracy does not patch the underlying risk.
The other half of the threat model matters equally. When research happens inside a platform built for that purpose, on a defined corpus, with clear data handling, the client context supplied during research stays where it was put. It does not become training data for a model serving someone else's query tomorrow morning.
The GC at 11pm deserves a different experience. One where the answer is traceable because the input is bounded. Where the case law comes from a closed dataset of legitimate Australian legal sources, not a probability distribution over the open web. Where confidence in the output is earned, not assumed.
A GC who told us that Habeas "materially changes my confidence and speed when forming legal views" was describing something concrete: the difference between knowing the research surface is inspectable and hoping a generative model happened to get this one right. Confidence built on a closed, traceable corpus and confidence built on a probabilistic guess can feel identical until something goes wrong, which is exactly when it is too late to ask which one you were relying on.
Habeas grounds every answer in that closed surface. The Search Engine scans over 300,000 Australian cases and pieces of legislation in seconds, drawing from a corpus of legitimate Australian primary law, so citations are verifiable, reasoning is followable, and the output can be defended under scrutiny. That is what makes it appropriate for the kind of work in-house counsel actually do: at 11pm, before a board meeting, when the question cannot wait and the stakes of the answer are real.
The first question to ask about any AI legal research tool is whether the surface is bounded. Everything else is secondary. If you want to see what bounded, traceable Australian legal research looks like, book a demo at Habeas.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Anastassia Anufrieva on Unsplash
