The Brief
Issue 12 · 24 August 2026
AI does not transfer your responsibility. Your regulator has now put that in writing.
The SRA's first formal warning notice on AI misuse. An Illinois court raises its fine rate on purpose. Australia binds every court user, not just lawyers.
On 17 August the Solicitors Regulation Authority published a formal warning notice on the misuse of AI. Not a discussion paper, not a set of compliance tips. A warning notice, which is the instrument the SRA uses when it intends to act on what it finds.
It names three failures. Hallucinated citations in court filings. Inadequate supervision of AI-assisted work. And client information entered into public AI tools, with the privilege and data-protection consequences that follow. The notice states that AI use "does not reduce, remove or replace" the professional, regulatory and legal responsibility a solicitor holds for the quality and accuracy of what they deliver. It also states that failing to have regard to it puts you at risk of disciplinary action.
Read the three failures again. None of them is a failure of the model. Each is a failure of the process wrapped around it.
A court prices the deterrent instead of marking the breach
On 28 July the Illinois Appellate Court, First District sanctioned an attorney $15,000 over a response brief carrying false case citations and false statutory quotations produced with a commercial AI subscription. The court set the figure at $1,500 for each of the ten, deliberately above its usual rate, and referred the attorney to the state disciplinary body.
The reasoning is the part to read. The court said it had no choice but to increase fines for AI-hallucinated citations until those fines carry a significant deterrent effect. Courts have been sanctioning this conduct for two years. What is new is a bench openly treating the penalty as a price that has so far been set too low to change anyone's behaviour.
Australia binds the court user, not the profession
Australia's Federal Circuit and Family Court has had a Practice Direction on the use of AI in force since 29 May. Its scope is the interesting part: it addresses court users generally rather than admitted practitioners only, which pulls litigants in person and engaged experts inside the same rules.
Two requirements carry. Always verify AI-generated content, including legal authorities. And do not put confidential or privileged material into public AI tools without an express assurance that the system has appropriate safeguards, with a firmer prohibition again for anything under a discovery order, a suppression order or a subpoena. Affidavits are expected to reflect the witness's own words and knowledge.
Our read
Three jurisdictions, three instruments, one list. Citations nobody checked. Work nobody supervised. Client material typed into a tool that was never private. A regulator, an appeal court and a trial court arrived at the same three items without coordinating, which is usually what it looks like when a problem is structural rather than local. Two of the three are recoverable. A fabricated citation can be caught in review. An unsupervised draft can be re-read before it goes out. The third is not. Client material sent to a public model has already left the building, and no policy written afterwards calls it back. That asymmetry is the whole argument for deciding where your documents are allowed to go before anyone starts typing, rather than auditing where they went.
Get the next issue in your inbox