A practitioner pulls up a generic AI assistant the evening before a filing. The query is straightforward enough: the limitation period for a professional negligence claim in New South Wales, with the latent damage angle. What comes back is fluent, structured, and internally consistent. The Limitation Act 1969 is correctly identified. The primary period is right. The latent damage rule that follows tracks professional negligence doctrine with apparent authority, and the practitioner, under pressure, relies on it.
The problem surfaces later, when it matters. The model's latent damage reasoning tracked US doctrine rather than the provisions as Australian courts have actually construed them. The output read exactly like law. It was close enough to pass a tired review. In law, close is precisely where the liability lives.
On 29 June, Artificial Lawyer published a piece arguing that "AI sovereignty" — the drive to control your own legal AI stack — is becoming a coordinated global movement. Its examples: Mistral in France, Noxtua in Germany, Kirkland and Ellis apparently building its own GPU clusters. The framing is almost entirely European or American biglaw. The frame is wrong, and so, probably, is the proposed solution.
Start with what sovereignty seems to mean in the current debate. For a firm like Kirkland, it apparently means ownership of compute: physical infrastructure, the ability to run its own training workloads. For European jurisdictions, it means choosing foundation models built on European data under European governance, rather than models produced in California and subject to American export controls. Both positions have genuine merit. But for the practitioner in that scenario above, or a solo GC advising across employment, privacy, and contract work on a Tuesday afternoon, neither is a realistic aspiration. Nor is it the right goal.
A mid-size Sydney firm is not going to build a data centre. That reflects good resource allocation. The sovereignty conversation only becomes useful when it shifts from infrastructure questions to something more immediately tractable: what does a practitioner actually need to control to ensure their AI use is reliable, defensible, and grounded in Australian law?
This is where the Australian version of the argument turns out to be sharper than the European one. The major foundation models were trained on enormous corpora dominated by American legal material: US caselaw, US statutes, US academic commentary. When a practitioner asks a generic assistant about proportionality in a costs assessment, or the interaction between the Civil Liability Act and a common law negligence claim, the model draws on a training distribution in which Australian material is a thin signal. If the models, the chips, and the data supply chain all sit under the operational control of US companies, an Australian firm's AI capability is rented infrastructure. It can be repriced. The outputs can shift as the underlying model is updated, fine-tuned, or redirected. That is ordinary commercial logic, and few firms have seriously stress-tested it.
The practitioner in our opening had no way to know any of that from the output. It looked authoritative. Sovereignty, in this context, has a concrete meaning: ownership of the ground truth and the ability to audit how any given output was derived from it.
Those two things are what most Australian practitioners are actually missing when they reach for a general-purpose AI tool. The ground truth question is whether the legal sources informing an answer are Australian primary law, or a vast undifferentiated corpus that happens to include some. The audit trail question is whether the practitioner can trace an output back to a specific authority and verify that it says what the model claims. Both carry professional weight, and both bear directly on the verification obligations the Federal Court's GPN-AI, published in April this year, now places on AI-assisted work in federal proceedings. The Court assumes a practitioner can identify what tool was used, how, and for what purpose. A tool whose reasoning is opaque by design cannot satisfy that assumption, however fluent it is.
There is a further wrinkle. The GPN-AI places verification responsibility squarely on the practitioner who signs the document, not on the vendor who built the tool. A generic model's opacity is not a shared risk; it transfers entirely to the firm. That compounds the jurisdictional problem.
Habeas scans over 300,000 Australian cases and pieces of legislation, drawn from a closed dataset of legitimate Australian legal sources. Every result is traceable to the source document. Foundational research that used to take a full morning can now be completed in minutes, and the citations that come back are ones a practitioner can open and read, not outputs that require a separate verification pass before they can safely be relied upon. That architecture is a sovereignty position. The ground truth is Australian law, maintained as a closed and curated dataset. The audit trail exists because every output points to primary sources. Matter documents uploaded through Document Stores stay in a closed environment; they do not train a shared model or become accessible to third parties. The Research Assistants return structured, cited analysis rather than fluent assertions that leave the verification risk entirely with the practitioner.
For the barrister running a brief the night before a hearing, or the GC who needs to explain to a court how an AI-assisted filing was verified, those specifics are the whole argument. The Artificial Lawyer piece frames sovereignty as something large firms achieve through infrastructure investment and something national governments achieve through industrial policy. That is accurate as far as it goes. But it leaves a gap where most legal work actually happens.
European firms building on Mistral, or US biglaw constructing its own compute layer, are investing significantly to own infrastructure that remains opaque at the source level. The model is theirs, but if it returns an unsupported proposition about Victorian contract law, owning the GPU cluster does not help a practitioner defend that answer in court. Owning the machine is a different thing from owning the ground truth.
For most Australian practitioners, that distinction is the whole argument. They will not build their own models and they should not have to. What they can insist on is that the tools they do use run on Australian primary law, keep matter data onshore, and show their sources with enough specificity that the firm is not simply trusting a foreign black box. That version of sovereignty is available now, and it does not require a policy push from Canberra or a GPU cluster in Alexandria. It requires the right tool, built for the right jurisdiction. If you want to see how that works in practice, Habeas is 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: Homa Appliances on Unsplash
