Harrisburg moved: HB 1925 passed the House, 144 to 59.

On September 30, Pennsylvania's House passed HB 1925 and sent it to the Senate. The bill would make insurers, hospitals and clinicians disclose how they use AI, and it would require a human decision-maker to have the final say when AI is involved in coverage decisions.

For your office: nothing in it touches a CPA, law or engineering firm. Still, the direction is plain enough for any practice: disclose the tool and keep a licensed human deciding. An assistant that drafts while a person signs already fits.

The IRS put AI in Circular 230 terms.

On June 24, the IRS Office of Professional Responsibility told tax practitioners what it expects when AI touches client work: handle client data only in secure, enterprise-approved tools; verify every fact, citation and calculation yourself; keep the procedures documented and disclose to the client the AI activities performed.

For your office: "secure, enterprise-approved" is easiest to show when the tool is a machine you own and client data never leaves it. The verification duty stays yours either way; nothing signs for you.

Pennsylvania wrote down what a trustworthy AI tool is.

The state Supreme Court's interim policy for court personnel, in force since December, defines a "Secured AI System": it protects confidential inputs, does not retain user-entered data or documents, does not train models on that data and bars transfer to third parties. Anything less, the policy says, should be treated as public.

For your office: the policy binds the courts, not you. But it is the state's clearest official definition of an acceptable tool, and a model running on your own hardware clears it by design: there is no outside party to retain anything.

Judges now ask who checked the citations.

Butler, Lancaster and Westmoreland counties require a statement with any filing AI helped draft: name the tool, name the portions, certify the human check. A number of federal judges in the state's Eastern and Middle districts ask for the same.

For your office: the common thread is receipts. Written proof that a person verified the work is what courts weigh; a policy on a shelf has not saved anyone.

2,149 decisions and counting.

A public database tracking court decisions that involve AI-invented citations counted 2,149 worldwide on October 5. Of those, 854 involved lawyers, and solo and small firms make up the largest share of them.

For your office: the problem has never been using AI. It is signing what nobody verified. A tool that answers with page cites makes the checking fast; nothing makes it optional.

These notes summarize public legislative and court actions as of October 7, 2026. They are a vendor's reading, not legal advice.

If you want a map of where your firm's client data goes today, that is what the AI Exposure Review is for.