Your AI Prompts Are Evidence Now

In May, a federal court ordered an expert witness to hand over the exact prompts she typed into an AI tool. Not a summary of her process. Not a description of her methodology. The prompts themselves. If you use AI in professional work, that order is worth five minutes of your attention, because the logic behind it applies to far more people than expert witnesses.

I work in litigation support. Part of my job is producing intelligence reports that need to survive scrutiny from opposing counsel. Over the past couple of months I have watched courts work out, case by case, what happens when AI shows up in the evidence chain. The pattern that has emerged is clear enough to state in one sentence: what happens to your AI prompts depends on who wrote them, for what purpose, and on what platform.

The expert who called them search terms

The May ruling came out of an environmental case in Connecticut, Conservation Law Foundation v. Shell Oil. The plaintiff's expert used an AI tool to sift a massive document production down to a working set. When the defense asked for the prompts she used, her side argued that prompts are not discoverable, that a prior discovery agreement shielded them, and that no prompts existed anyway. Only search terms.

The court rejected all three arguments. The reasoning was not exotic. An expert's methodology has always been fair game in discovery. If you used a formula, you disclose the formula. If you used code, you disclose the code. The court simply treated prompts the same way: the instructions that shaped how the evidence was filtered are part of the method, so they are within scope. What sank the "only search terms" claim was the expert's own research assistant, whose earlier declaration mentioned prompts. He also had not preserved a complete log of them, which opened a second front about destroyed evidence.

The executive who asked a chatbot how to dodge $250 million

Expert witnesses are not the cautionary tale that sticks with most people. This one is. In March, the Delaware Court of Chancery decided a dispute over a $250 million earnout owed to the founders of a video game studio. The buyer's CEO had asked ChatGPT how to avoid paying it. The chatbot eventually produced a multi-step pressure strategy. He named it Project X and his company ran the play.

The court's opinion quotes his ChatGPT conversations at length, because nothing about them was privileged. They were ordinary business records, no different from email or Slack. He also deleted his chat logs along the way, which courts tend to notice. The buyer lost. The fired CEO was reinstated and the earnout window was extended.

A month earlier in New York, a criminal defendant learned a related lesson. Before trial, he had used a consumer AI tool to outline his defense strategy, then shared the 31 resulting documents with his lawyers. The court ruled that none of it was privileged. A chatbot is not an attorney. The platform's own privacy policy said user data could be shared with third parties, so there was no reasonable expectation of confidentiality. And forwarding the documents to counsel afterward could not retroactively protect them. Prosecutors got to read his defense playbook.

Who wrote the prompt decides what happens to it

It is not all one-directional. Courts have also protected AI prompts, and the line they are drawing is consistent. When attorneys craft prompts as part of a litigation strategy, those prompts reveal the attorney's thinking, and courts have treated them as protected work product. One federal court put the other half plainly: AI programs are tools, not persons. Merely using one does not hand your notes to the other side.

But the protection is narrow and easy to lose. In a copyright fight between music publishers and an AI company, the publishers' attorney-crafted prompts were protected right up until some were shared with an expert witness. That sharing waived the protection. In the same case, the AI company was ordered to produce a sample of five million prompt-output pairs from its own systems. Five million. Prompt logs are not a niche discovery item anymore. They are the discovery item.

Courts are even writing AI rules into protective orders before disputes start. One federal court in Colorado now requires that any AI tool touching confidential discovery material come with contractual guarantees: no training on the data, no disclosure to third parties, and deletion on request. A court in Kansas went further and banned consumer AI tools from touching discovery materials at all, reasoning that once data goes into a model, you cannot reliably get it back out.

The standards bodies have not caught up

Here is the part that should concern anyone whose work product can end up in front of a judge. The courts are moving fast. The professional standards bodies are not. The American Bar Association issued ethics guidance on AI in 2024. The AICPA published guidelines for AI use in forensic and valuation work in late 2025, and to its credit, that document specifically recommends keeping prompts and outputs in the engagement file. But it is explicitly non-authoritative. The fraud examination field has no formal AI documentation standard. The major investigations standard published in 2025 contains no AI-specific provisions either.

So the documentation requirement is being written by cross-examination instead. Sanctions for unverified AI work started at $5,000 in the first famous fake-citations case in 2023. Since then, courts have escalated to disqualifying attorneys from cases, referring them to bar regulators in every state where they hold a license, and in one appeals court, dismissing a case outright. One judge, rejecting a monetary fine as too soft, wrote that if fines and public embarrassment worked, there would not be so many cases to cite.

What I do about it, and what you should consider

In my intelligence work, every report has to survive scrutiny from people who are paid to take it apart. So our AI use is documented as methodology, the same way any other analytical step is. When the question comes, and it does come, we can answer it. That is the whole game.

If AI touches work that could ever land in a dispute, four habits cover most of the risk:

• Record the tool and the model, by name and version, in the work file. Not "we used AI."
• Keep the prompts and outputs as you go. Reconstructing them later is how the Shell expert ended up facing a destroyed-evidence argument.
• Know what platform you are on. Consumer tools with permissive data policies offer no confidentiality. Enterprise deployments with contractual protections fare better in court, repeatedly.
• Assume the deposition question is coming. "What AI tools did you use, and what did you ask them?" is now a standard line of inquiry. Be the professional who can answer it calmly.

One last note. While writing this post, I ran a small test: I asked four different AI tools to research this exact topic. Between them, they handed me one case under three different names, three wrong dates for a single ruling, two citation formats that do not exist, one legal test I cannot find anywhere, and a sanctions figure inflated two different ways. On a topic about AI fabrications ending up in court, the tools fabricated. Every claim in this post was verified against court documents and primary coverage before it was published. The claims that did not survive are not here.

That is the takeaway in miniature. AI is genuinely useful in serious work. It is also a witness with a shaky memory, and now the court can call it. Keep your receipts.