· Sam Sperling, Founder
Nobody sends the memo
The standard of care for using AI in a filing is being written right now, one sanctions order at a time, across districts that do not talk to each other. There is no circular. There is only the record.
On 13 March 2026 the Sixth Circuit sanctioned two lawyers in the consolidated appeals of Whiting v. City of Athens, Tennessee. Their briefs contained over two dozen fake citations and misrepresentations of fact — the court called that a conservative estimate. Each attorney was ordered to pay $15,000 into the court registry, both were made jointly liable for the appellees' full fees on appeal, and the court assessed double costs.
The reasoning is the part worth reading twice. A fake opinion, the court held, is not existing law, so citing one cannot be a non-frivolous ground for extending, modifying or reversing existing law. That is not a rule about technology. It is Rule 11 arriving at a conclusion that was always available, applied to a new way of getting things wrong.
Three months earlier, a magistrate judge in the District of Oregon imposed roughly $110,000 across two attorneys over fifteen nonexistent cases and eight fabricated quotations, aggravated by an attempt to refile without verifying the corrections. The judge called it a notorious outlier in both degree and volume.
Outliers become baselines. That is what a developing standard of care looks like from the inside.
The standard is being set in orders
Here is the structural problem. The duty of competence and the duty of candour to the tribunal are not being amended by anyone's rules committee, at least not first. They are being applied, case by case, by individual judges in individual districts, in orders that bind you and are announced to nobody.
Some of it does surface as rules — the Florida Supreme Court now requires the signer of any document filed in any Florida court to represent that the authorities cited exist and are accurately cited. Some of it surfaces as legislation; California SB 574 is being argued over as we write. But most of it is a scatter of sanctions orders across districts that do not talk to each other, and the composite picture exists nowhere official.
By the time this reaches a CLE slide deck, it has been binding on you for a year.
What one week actually looks like
We built a job that polls the legal AI trade press and measures what comes back. Over a recent seven-day window it pulled 150 items from 17 outlets. Sixty-seven of those survived a keyword screen for anything touching AI and legal work.
Sixty-seven in a week. Nobody reads that, and nobody should. But the alternative most lawyers have settled on — reading none of it and finding out at the sanctions hearing — is worse.
The month those numbers came from was not unusual. Google shipped Gemini Enterprise for Legal, joining OpenAI and Anthropic with a legal-specific offering. Thomson Reuters launched its own LLM trained on Westlaw and Practical Law. Elevate bought Lupl. BigHand, Anaqua and Legora all made acquisitions inside a fortnight — enough that the trade press started asking whether this is the great legal tech consolidation.
If your firm is buying any of this, the vendor list is being redrawn under you. If your firm is not, the other side's is.
So we built the boring version
The river is a chronological feed of legal AI news. Date, time, outlet, byline, headline, linking out. Tagged by topic so you can read the courts thread and ignore the funding thread, or the reverse.
It is deliberately not curated by importance. Somebody's ranking of what matters this week is a worse instrument than the order things happened in, because the ranking encodes their priorities and you have your own. A partner watching for malpractice exposure and a COO watching for vendor consolidation need different threads out of the same stream.
Four rules hold it up, and they are printed on the page:
Editorial only. Vendor press releases stay out, even when a real outlet carries them under its own masthead. A press release is the vendor's account of the vendor's product, and there is already an excellent place to read those.
Bylines as published. An outlet that runs a piece unsigned gets no byline rather than an invented one. You should be able to see who is telling you something.
Nothing summarised from a headline alone. Every row links to the report it came from. Where an outlet is paywalled we say so and do not paraphrase a body nobody can read.
No invented timestamps. Where the outlet publishes a date but no time, the time column stays blank. It looks like an oversight. It is the opposite.
What it is not
It is not a destination. Every headline is a link out to somebody else's reporting, which means the river's entire job is to lose the click as efficiently as possible. We are not going to summarise Bob Ambrogi at you. Go read Bob Ambrogi.
It is not analysis, and it is not advice. A feed cannot tell you whether your verification workflow satisfies your jurisdiction's version of the competence rule. It can tell you that the question is live and where it is being answered.
Where we are not neutral, and where this is thin
We sell software in this market, which gives us an interest in you paying attention to it. Worth knowing when you read anything we publish about it.
The source list is also narrower than we would like. Eleven outlets are approved, six more are on trial, and the coverage skews US and UK because that is where the accessible feeds are. Law360 and Bloomberg Law both do strong reporting here and neither publishes a feed we can read, so they reach us only as headlines.
The most conspicuous gap is the one that matters most for the beat this post is about. Damien Charlotin's running database of court decisions involving AI-fabricated citations is the authoritative primary source — the trade press cites it rather than the other way round — and it publishes nothing machine-readable. It sits on our watchlist as a thing a person has to go and look at. We have not solved that, and we are not going to pretend otherwise.
The actual argument
Not that legal AI is exciting. Some of it is, most of it is procurement.
The argument is narrower: the line between competent use and sanctionable use of these tools is being drawn right now, in public, in documents you can read, by judges who are not going to send you a memo when they draw it. Following that is not technology interest. It is the same reason you read advance sheets.
The river is here. It is free, there is nothing to sign up for to read it, and if you would rather have it arrive than go and get it, the subscribe box on that page sends one email a week and nothing else.
Nothing here is legal advice, and a news feed is not a substitute for checking the rules of the court you are actually in. Sanctions figures and case details are as reported by the outlets linked above.