Legal: the bill for getting AI wrong just became law

Legal AI hit its accountability moment in June 2026, when a US federal appeals court made the duty to check an AI's work binding law. On 3 June the Ninth Circuit published a precedential ruling sanctioning two lawyers for filing briefs full of cases that did not exist, and it fixed the lawyer's responsibility at one precise point: the moment they sign and file. For a small firm, the lesson is blunt. The tool can be cheap, but getting it wrong is expensive, and the bill now sits with the lawyer who signed.

When Era Haus last looked at this industry in AI tools got cheap, and an AI-first firm arrived, the story was supply: serious legal AI stopped being a premium purchase and dropped to roughly $20 a seat, while a venture-backed firm began selling AI-run legal work to funds. A month on, the other side of that bargain has arrived. The work got cheap to produce; the responsibility for it got more expensive, and the courts have started naming the price.

What the courts just decided

The development to lead on is days old and, unlike earlier warnings, binding. On 3 June 2026 the Ninth Circuit, the US federal appeals court covering the western states, published its opinion in Lnu v. Blanche. Two lawyers had filed briefs citing cases that did not exist and quotations that were never written, then denied AI was involved. The court fined each $2,500, suspended both from practising before it for six months, and ordered that for two years every lawyer at their firm attach a sworn statement to each filing: whether AI was used, which tool, and confirmation a human checked every citation (Reason, June 2026).

Publishing it was the point. Trial courts had sanctioned lawyers for AI fabrications dozens of times, but those rulings bound only the parties in the room. An appeals court opinion marked for publication becomes law for every court beneath it, and this one located the lawyer's duty at the act of signing, not at the moment the model wrote the draft.

Five days later the same point landed again. On 8 June a federal judge in Mississippi cancelled a trial and removed every lawyer on the case, two on each side, after both teams filed work containing AI-invented sources. Two were barred from the district's courts for two years, and the rest were fined and disqualified (Bloomberg Law, June 2026). The judge accepted they had acted carelessly rather than in bad faith, and removed them anyway. Carelessness was enough.

It is not only a US story

This pattern reaches well beyond the US. In the United Kingdom, the High Court restated this spring that a solicitor cannot hand legal research to a chatbot, after a junior lawyer at the UK firm Pinsent Masons filed letters citing an insolvency rule the AI had invented (UK Human Rights Blog, June 2026). A public database tracking these episodes now counts more than 700 worldwide. Wherever a lawyer signs a filing, the same duty is being restated: the machine can draft, but only a person can vouch for it.

Meanwhile the tools moved deeper

None of this slowed the products. On 1 June 2026 OpenAI formally opened a legal division and hired the co-founder of the contract-software company Ironclad to run it, weeks after Anthropic wired its AI into the research databases and document tools lawyers already use (Artificial Lawyer, June 2026). The two largest AI companies are now building for lawyers directly, and the newest tools do not only answer questions; they run multi-step tasks on their own between human checkpoints. The capability keeps getting cheaper and more autonomous at the same moment the courts are making the human reviewer personally answerable for it.

What it means for you

For a small firm or a solo practitioner, the squeeze is now visible from both sides. The cheap tool that lets you produce a brief in an afternoon is the same tool that can quietly invent a case, and the cost of that invention is no longer a private embarrassment. It is a published sanction and a story your next client can find online. The work moved to the machine; the liability stayed where it always was, with the name on the signature.

That changes what you are actually selling. First-pass research and drafting, the work AI now does in minutes, was never the thing a client paid a lawyer to guarantee. What they pay for is the judgement that the answer is right and the willingness to stand behind it. That is the case Era Haus made in defensibility in the AI era: when the tool becomes cheap and common, the durable advantage is the trust and the accountability around it, which do not copy.

What to do about it

Three grounded moves. First, write a short AI-use policy and require that every citation in anything leaving the office is checked against its actual source by a named person before filing. An appeals court has now made that duty explicitly yours.

Second, expect the disclosure itself to become routine. Courts are beginning to require a signed statement on whether AI was used and confirmation that a human reviewed the output. Build that habit now, while it reads as a competitive signal, rather than later when a judge orders it.

Third, decide how you price this. If AI lets you produce a first draft far faster, a client will eventually ask why the bill looks the same. Moving routine work toward a fixed fee, verification included, protects you better than billing fewer hours at the old rate.

What to watch rather than act on yet: the agentic tools that run a whole task on their own. They are genuinely useful and genuinely cheaper, but every output still arrives under your signature. Until you have watched one work on low-stakes matters and know how often it is wrong, keep a person between the machine and the court.

The pattern underneath

The shape matches what we have traced across other fields: the capability gets cheap, the routine work moves to software, and value moves to the judgement and responsibility the software cannot take on. Legal's version is the sharpest, because the regulator and the judge are already in the room. The firm using these tools to produce more, faster, gets the small win. The one treating responsibility for the work as the actual product, and pricing and staffing around that, is working on the real one.