AI now drafts arbitration awards and fully finishes one in five benchmark legal tasks, and in the same week California sued over AI agents it says nobody was properly watching, while ChatGPT's fingerprints turned up in hundreds of court and SEC filings.
Astra for Law launched in September for a few invited firms. On November 17 OpenAI made an updated version, built on GPT-6.2 Astra, available to any ChatGPT Enterprise customer on request, and through its API. On the benchmark OpenAI leads with, the model fully completed 20.8% of assignments.
The progress is real on OpenAI's numbers: higher scores than its base model, far fewer invented authorities per assignment (0.03, against 0.59 for Claude Opus 5.5 on the same count), and less than half the output tokens of the rival it compares against. But every figure is the vendor's, and the scoring is strict for a reason. A brief that gets four of five things right still needs a lawyer to find the fifth.
A model that fully finishes one assignment in five is a strong drafter, not finished work. The distance between 20.8% and done is filled by your templates, your positions, your escalation rules and a lawyer who signs off. That is deployment work, and a licence to a better model does not do it. The work your lawyers should not be doing stays on their desks either way.
The American Arbitration Association has extended its AI Arbitrator from construction cases to documents-only business-to-business disputes in any industry, LawSites reported on November 17. The system was trained on real awards from more than 20 industries. It assesses the claims and drafts the award, and a specially trained human arbitrator reviews, revises and issues it.
For in-house teams this lands on both sides of the contract. Arbitration clauses in supplier and customer agreements may now lead to an AI-drafted award, and nothing published yet says how many awards it has drafted, how often the human arbitrator changed them, or how parties opt in. Those are the questions to put to the AAA before agreeing to the process.
The AAA has put the line in a sensible place: AI drafts against a defined record, a trained person decides, and that person's name goes on the result. That is the supervised-agent model applied to adjudication. If an arbitration body will let an agent draft a binding decision under review, the case for sending counterparty paper and NDAs to outside counsel at premium rates gets harder to make.
Attorney General Rob Bonta sued OpenAI on November 16 over a July incident in which hundreds of autonomous agents, running on experimental OpenAI models, attacked the AI platform Hugging Face for more than two days. The state's theory does not turn on what the agents intended. It turns on whether anyone was properly supervising them.
The state wants an injunction, disgorgement of profits and fines of $2,500 per violation. Public nuisance is an untested route for AI, and OpenAI will fight it. The claim that matters for legal teams is the second one: that a company misled the public about how closely it watched its agents. Any vendor selling agents into your business makes a version of that promise.
When an agent does work on your behalf, can anyone show you what it did, which rules it applied, and who approved the result before it left? If the answer is a vendor's assurance, California is now testing what that assurance is worth in court.
ChatGPT adds a tracking tag to links it hands out. A study published on November 17 searched public archives for it and opened every match. It found the tag in company securities filings and federal court papers, including two federal court opinions.
The tag does not show that text was AI-written or that a source is wrong, as the study's author concedes. What it shows is that a link went from a chatbot answer into a signed document without anyone opening it closely enough to clean it. About one in five tagged SEC links pointed to a market-research vendor, which suggests ChatGPT was used to size markets in documents investors rely on.
Does your disclosure process know where AI was used in a filing, and who checked each source? And do your engagement letters ask outside counsel the same question about what they file in your name?
Each story this week put AI output in front of a decision: an award, a filing, a court brief, an agent on the open internet. Where a named person reviewed the work, it was usable. Where nobody did, it became the story.
Expensive people are still doing a steady flow of work beneath their cost, and a better model does not take it off them. What does is an agent built to your own rules, doing the whole piece of work, with a lawyer accountable for what leaves the system.
Flank deploys agents that do legal work. They run on your templates, positions and escalation rules, they finish the work, and your lawyers sign off the few points they raise.