The assignment clause is there. The model flagged it. You move on.
Except the guideline required an unconditional assignment with no consent requirement, and the model, having correctly identified the presence of an assignment provision, said nothing about whether consent was required. It found what it was looking for, and missed the only thing that mattered.
This is an argument Habeas has been making since launch. A recent benchmark puts data behind it.
LegalOn's 2026 Contract Review Benchmark tested eleven AI models across 3,282 head-to-head contract reviews, measured against twenty-one precision-critical guidelines. The same pattern repeated across the general-purpose models: they identified the concept but failed on the threshold. Finding an assignment clause is not enough when the guideline requires an unconditional assignment with no consent requirement. The benchmark gives the category something it has lacked, independently produced, in-window evidence that the gap between "broadly right" and "legally correct" is real and measurable.
For anyone who has watched how these tools behave in Australian matters, it confirms something familiar: general reasoning capability and jurisdiction-specific accuracy are different problems.
Concept identification is something the frontier models have largely solved. The failure lives one level down, in what the standard precisely requires and whether the clause meets it. The practitioner in the assignment scenario never reached that second question. The model answered the first and stopped, with no signal that it had treated them as different things.
Follow that practitioner one step further. The matter extends to a commercial software agreement, and they ask the model to assess a warranty disclaimer. The statutory consumer guarantees in the Australian Consumer Law apply to consumer contracts by force of law and cannot be excluded, however the clause is drafted. [VERIFY: confirm the exact section span before citing it numerically, e.g. "sections 51 to 64A". The guarantees sit in Part 3-2 Division 1 of the ACL; check the precise range and the non-exclusion provision (around s 64) against the Act before publishing.] A model trained primarily on US commercial precedent will frequently accept a well-constructed disclaimer as doing its job, because in American law it often does. The practitioner gets a clean summary, the advice goes out, and the gap travels out with it.
The gap is wider still for the doctrine underneath. Statutory unconscionability under section 21 of the ACL has been shaped by a line of Full Federal Court authority [VERIFY: name at least one actual decision here rather than "a line of authority", e.g. confirm and cite the relevant Full Court case establishing that s 21 can operate systemically against a class without proof any individual was under special disadvantage. Confirm the case name and citation before publishing.], including decisions that the provision can operate systemically, directed at a class of consumers, without requiring that any individual was under special disadvantage at the moment of contracting. That is a meaningful departure from the equitable doctrine the High Court applied in Commercial Bank of Australia v Amadio (1983) 151 CLR 447, and it lives in local decisions a model trained on globally distributed legal text will not have absorbed with precision. A model can read the section and still misread what Australian courts have made of it.
High MMLU scores, bar exam pass rates, and general legal reasoning indices measure performance across a broad distribution of tasks. They do not measure whether a model will correctly identify that a notice obligation under Australian securities law runs to a different window than its US equivalent, or whether a restraint clause will survive scrutiny as Australian courts have developed the common law. The way this surfaces in practice is at the advice stage: the research looks complete, the clause has been identified, a framework offered, and then the partner asks which Australian court last considered the question and what it held. The model's thoroughness was working at the topic level, and the threshold question underneath it was never reached.
This is the gap Habeas was built to close. The platform searches over 300,000 Australian cases and pieces of legislation from a closed dataset of legitimate Australian legal sources, so every result traces to the source document and you can see whether the authority resolves the Australian question or an analogous foreign one. One General Counsel puts the difference plainly: "The Australian-law focus and the depth of nuance in the answers is a major differentiator. It materially changes my confidence and speed when forming legal views." Research that once took much of a morning can be done in a fraction of it, with the citations there to check.
The practitioner still has to read the clause, know the standard, and sign the advice. The assignment scenario this piece opened with cannot be solved by any research tool alone. What a research foundation should deliver is clarity about which Australian court said what, and when, traceable to the source, so the practitioner carrying the professional risk is working from something verifiable rather than something that merely sounds plausible. The LegalOn benchmark shows that general-purpose models cannot reliably deliver that. A system built from the ground up on Australian primary law has a different answer to give.
Book a demo at habeas.ai.
The legal research in this article was conducted and every citation verified using Habeas, the Australian legal AI research platform.
Hero image: Dimitri Karastelev on Unsplash
