Legal AI vendors have spent the past two years competing on the strength of their marketing copy. "Predicts litigation outcomes." "Automates legal research." "Hallucination-free." The ACCC has now given those claims a number: up to $100 million per contravention for the most serious misleading conduct, under the doubled penalty regime.
The regulator published its compliance and enforcement priorities for 2026-27, and artificial intelligence sits prominently on the list. The ACCC has explicitly named misleading claims about AI capabilities as a focus area for the coming year. Enforcement priorities are not guarantees of action; they are reliable signals of where the regulator is directing its attention and its resources.
Both elements matter together. An enforcement priority without teeth is a press release. A doubled penalty regime without enforcement focus is background noise. Together, they constitute a materially different regulatory environment from anything Australian AI vendors have operated in before.
The term has been in circulation for a few years, mostly to describe companies overstating the role of machine learning in their products. In the legal sector, the problem takes a specific form.
Vendors claim their tools predict litigation outcomes. They describe accuracy rates without specifying what was measured, by whom, or against what baseline. They use "hallucination-free" without explaining the technical architecture behind the claim. They imply that their products reduce professional liability rather than potentially adding to it.
Some of these claims are untrue. Others are technically accurate in narrow conditions that bear little resemblance to how a practitioner uses the product day to day. Both categories are capable of misleading a firm that relies on them in a procurement decision, and both sit within the kind of conduct the ACCC's new priorities are designed to reach. When regulators publish enforcement priorities, they have usually already started gathering evidence.
The ACCC's conduct in adjacent categories gives that observation weight. Its greenwashing enforcement over the past two years moved from priority to penalty faster than many companies anticipated, and in several cases the targets were not fringe operators but established market participants who assumed the language of aspiration was not the language of law. The legal AI sector would be unwise to read enforcement priorities as a longer runway than they are.
This is where the issue becomes a practitioner's problem rather than a vendor's.
A firm that purchases legal AI on the basis of misleading capability claims has been misled. The Australian Consumer Law provides remedies for that. But the firm also carries the downstream risk of deploying a tool that cannot do what the vendor claimed. An AI system sold as producing reliable legal research, which then hallucinates authorities, creates a professional risk that sits with the lawyers who used it. The vendor's misleading marketing does not transfer that risk back to the vendor.
There is a further layer that procurement decisions often miss. Many vendor contracts contain limitation clauses that sit in direct tension with the marketing copy: "no warranty as to accuracy," "outputs should be independently verified," "the platform is not a substitute for legal advice." A firm that bought on the strength of "hallucination-free" and then signed a contract disclaiming accuracy is exposed on two fronts. The marketing may constitute misleading conduct by the vendor; the contract may leave the firm without a remedy when the output fails.
The practical question for any firm that has adopted AI tools recently, or is evaluating them now, is whether the vendor's claims are defensible. "Predicts outcomes" is almost certainly not. No AI system can reliably predict litigation outcomes across Australian jurisdictions; the variables are too many and the training data too sparse. A vendor making that claim is either describing something much narrower than the language implies, or making a claim that does not survive scrutiny.
"Automates legal research" is similarly elastic. If it means the tool returns a first sweep of relevant authorities faster than a manual search, that is a real and useful capability. If it implies the practitioner can skip review and file what the AI produces, it is a misrepresentation and a professional risk simultaneously. The Federal Court's Generative AI Practice Note, which came into force in April 2026, makes this specific: the signing lawyer personally confirms that cited authorities exist and support the stated proposition, regardless of how the draft was produced. Verification by the practitioner is the substantive task, not an optional quality check appended to a workflow that has otherwise done the work.
The ACCC's focus on this space gives procurement decisions a new dimension. Asking a vendor to substantiate its claims is sound due diligence and a reasonable response to a regulatory environment that has identified the category as one where misleading conduct is occurring.
We have been deliberately restrained in how we describe what Habeas does. That restraint was a brand choice from the start. The ACCC's priorities have made it a compliance posture as well.
Habeas searches over 300,000 Australian cases and pieces of legislation, returns results traceable to their source documents, and is built on a closed dataset of legitimate Australian legal sources. Foundational research processes that used to take a full morning can now be completed in minutes. That is the genuine capability; it is also the bounded claim.
We do not claim Habeas predicts outcomes. We do not claim it replaces practitioner judgment. One General Counsel who uses the platform put it plainly: "It feels like having a law firm in your pocket. Not something you blindly bet the house on, but a powerful first-line legal intelligence tool." That framing captures the relationship between the tool and professional responsibility more accurately than anything we could have written ourselves: a first-line intelligence layer, with the analysis still squarely the practitioner's work.
Restraint in how we describe the product reflects something real about where legal AI sits right now. Research assistance, fast and traceable and grounded in Australian primary sources, is a genuine value. Outcome prediction is a capability no current system can claim with any integrity. Describing the former accurately while declining to assert the latter is what defensible AI marketing looks like, and it is precisely what the ACCC's new priorities will increasingly reward.
If your firm is using legal AI tools, or evaluating them, the ACCC's announcement is a prompt for a concrete set of questions.
When a vendor claims their system produces accurate legal research, ask what "accurate" means. Is it measured against a benchmark? Which one? What was the error rate in testing, and what happens when the system operates outside the conditions it was tested on?
Ask what the practitioner is still expected to do when a vendor claims their system automates research. If the honest answer is "verify everything before you rely on it," the automation claim is significantly overstated.
A claim of outcome prediction warrants a request for methodology. In most cases, the honest answer either will not be available or will reveal that "prediction" describes something far narrower than the word implies, perhaps a win-rate estimate on a particular issue type in a particular jurisdiction under specific conditions.
Ask to see the contract's limitation clauses alongside the marketing claims. The distance between "hallucination-free" in a brochure and "no warranty as to accuracy" in the terms is informative. A vendor whose contract contradicts its own marketing has given you the most honest summary of its product buried in the fine print.
These questions are straightforward, and a well-advised procurement process should have been asking them already. The ACCC's enforcement priorities give them urgency. A vendor that can answer them clearly is describing a product you can deploy with confidence. One that cannot is now operating inside the regulatory environment's stated focus area, and your firm's reliance on the vendor's claims is documented in a contract and a purchasing decision.
The legal AI market will sort itself over the next year as the ACCC moves from signalling to action. Vendors whose marketing survives scrutiny will differentiate from those whose claims were always aspirational. The practitioners who ask hard questions now are the ones whose procurement decisions will hold up. The regulator has made the cost of ignoring that concrete.
Habeas is free to try. If you'd like to see what traceable, citation-grounded Australian legal research looks like, visit habeas.ai or book a demo.
If you want to try for yourself or get in contact, book a demo with us here. We also offer the capacity for self-serve individuals to sign up, and subscribe or register a free trial at app.habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
