Consider a GC running a regulatory check at the end of a long day. An AI tool returns a clean, confident summary of the relevant statutory obligations, the kind of output that looks exactly like something you could act on. She drafts an advice note, flags the key provisions, and briefs the board before the week closes.
Eight months earlier, the provision had been amended. The summary was fluent. The citations were plausible. None of it reflected what the law currently said.
This is not a hallucination story. The tool did not invent a statute. The citations resolved to real sources. The advice was wrong anyway, and the reason had nothing to do with accuracy rates.
Legal work has always required that you check your sources. Not because technology is untrustworthy in some novel way, but because the professional standard demands it. A counsel who submits an authority without reading it has not met their obligation, regardless of where it came from. A GC who advises on a statutory provision without tracing the current text has not done the job, regardless of how confident the summary sounded.
This means the hallucination question, while real, is somewhat beside the point. Even a tool that is right ninety-five per cent of the time still requires that practitioners treat every output as provisional and confirm every authority at its source. The five per cent is not predictably distributed; you do not know in advance which citations fell over. So you check. The accuracy rate affects how often the check will fail, but it does not affect whether the check is required.
If verification is constant, the relevant variable is the cost of doing it. How long does it take? How many steps between the output and the source?
Generic AI tools built for general use have a structural problem in legal work that accuracy statistics do not capture. When an output cites an authority, the practitioner still has to find that authority, confirm the citation resolves to the right paragraph, read the passage, and assess whether it actually supports the proposition advanced. If the tool cannot tell you precisely where to look, or if the citation is plausible-sounding but unverifiable, the verification step costs as much as running the research again. You have not saved time; you have added a layer.
The GC from the opening scene, going back through that regulatory check, faces exactly this. She has a summary. She does not have traceable paths to the sources. Each authority requires a separate search. Each amended provision requires a manual check against the current text. The morning she saved on the first pass costs her the afternoon on the second.
We think this is the distinction that gets lost when the conversation stays fixed on accuracy rates. A tool that hallucinates less is better than one that hallucinates more. But a tool with traceable citations is worth something a tool without them is not, even at identical accuracy scores. One makes verification cheap. The other makes it expensive.
Run the same regulatory check in Habeas, and what comes back is different in character. The Search Engine scans over 300,000 Australian cases and pieces of legislation in seconds, drawing from a closed corpus of legitimate Australian legal sources. Every authority is traceable. Every citation points somewhere real, and that somewhere is directly accessible. The practitioner does not hunt for the source; she is already at it.
A GC running a first-line regulatory check across multiple statutes and the relevant case law, the kind of work that used to consume a full morning, can now be completed in minutes, not because the tool eliminated the need to verify, but because verification became cheap. She confirms each citation at its source. She checks the current text. The professional obligation stays intact. What changes is the friction around meeting it.
One in-house counsel described the effect directly: the Australian-law focus and the depth of nuance "materially changes my confidence and speed when forming legal views." The confidence is not misplaced. It is earned by the structure that makes the checking fast.
When evaluating a legal AI tool, the question worth asking is not how often it gets things right. The question is what happens when you try to verify it. Can you trace each authority? Can you confirm the proposition against the source in seconds? Does the tool make the verification step easier, or does it just move the work around?
A tool you can audit is worth building a workflow on. A tool you can only trust is a liability dressed up as efficiency. Auditable research, grounded in a closed Australian corpus with traceable citations, is what the standard has always required, and what the right tool makes genuinely achievable: Habeas.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Leon Seibert on Unsplash
