On Thursday, September 17, OpenAI launched Astra for Law: GPT-6 Astra bolted to an index of U.S. case law, statutes, regulations and “other legal materials,” with what the announcement calls specialised instructions for legal analysis and writing. The wire coverage framed it as the latest escalation in the race for law-firm customers.

Fine. But notice what the product actually eliminates. For two years, the legal profession’s AI panic has had exactly one villain — the fabricated citation, the brief citing a case that does not exist. Every standing order, every bar bulletin, every continuing-education webinar has been aimed at that moment.

Astra for Law kills it. That is the least interesting thing about the launch, and the only thing anyone wants to talk about.

The Deterrence Machine Guards a Door Nobody Uses

Consider the regime at its strongest. In Malkeet Lnu v. Blanche (2026 WL 1587554), the Ninth Circuit this year sanctioned two immigration attorneys who filed papers containing hallucinations and, per the New York State Bar Association’s write-up, insisted they had not used AI until the final minutes of oral argument. Monetary sanctions, plus a six-month suspension from practice before the appellate court. In the Southern District of Florida, attorney James Martin Paul drew what the legal-AI blog NexLaw calls the largest AI sanction in U.S. history — roughly $86,000 in ByoPlanet v. Johansson, dismissal of all four matters he was litigating, and a two-year obligation to attach the sanctions order to every filing. Federal judges have been saying outright that the money isn’t deterring anyone (Johnson v. Dunn, N.D. Ala., 2025).

Every one of those cases is a fabrication case. Every one is catchable by a first-year associate with a citation checker. And fabrication becomes structurally rare the moment the vendor stops writing case names from a probability distribution and starts retrieving them from a curated corpus — which is precisely the thing OpenAI just put on the market.

So watch what happens next. Sanctions filings will fall, and the same judges and committees who spent two years building the apparatus will take credit for a decline their apparatus did not cause. The tool did it. Which means the profession will conclude its AI problem is handled at the exact moment the problem changes shape.

An Index Is an Argument

The launch copy describes the index contents as case law, statutes, regulations “and other legal materials.” That last clause is doing an enormous amount of work. Which materials? Selected how? Refreshed on what schedule when a circuit reverses itself or a district splits? Weighted toward published opinions and away from the unpublished dispositions that quietly decide most motions?

A hallucinated citation announces itself. A well-supported brief built on a systematically incomplete corpus does not. It looks like a confident answer, and there is no sanctions hearing for a case that exists.

For three decades, firms paid Westlaw and Lexis for more than text; they paid for a vendor standing behind the text. OpenAI’s terms put the lawyer on the hook instead. The firms that can afford their own redundant retrieval layer will be fine, which is the least comforting sentence in this column.

As one paralegal who runs cite-checks at a forty-lawyer firm in Tampa put it in a Slack message between depositions: “The fake cites were loud. A brief that has the right cases and the wrong reading of them is quiet, and quiet is what I can’t check.”

The Affidavit That Certifies the Wrong Thing

The emerging remedy is disclosure. On August 11, a federal court in the Middle District of Pennsylvania (Courtney Voyton v. Joseph Voyton) admonished a self-represented plaintiff for AI misuse and ordered that future filings come with an affidavit disclosing AI use and confirming citations were verified — a template now spreading through chambers, and one that Damien Charlotin’s running database of hallucination cases is quietly documenting.

An affidavit attests to process. It cannot attest to a corpus. Tell a judge you ran your research through a private index and you have certified nothing about what the index omitted. The self-represented litigant — the person everyone claims this technology will liberate — is now best positioned to file a clean, real-cited, quietly incomplete brief, with paperwork proving diligence.

The deregulatory instinct here is to let malpractice law absorb the risk. Malpractice requires a client who lost and a causal chain a plaintiff can prove. Omission hands you neither.

None of this requires a new agency or a moratorium. It requires the vendors to publish the corpus: update cadence, selection criteria, adverse-authority testing, some audit trail a court can subpoena. A smaller ask than it sounds, and the only one that matters.

Sources