Your Chats With AI Are Now Evidence. What AI Said About You Still Isn't.

2026-09-19

In the past two years, reporters have found chatbot logs cited in at least a dozen court cases: a Florida double-homicide prosecution built partly on the defendant's own ChatGPT queries, a Los Angeles arson case where federal prosecutors quoted a conversation with the same system. Reporting over the past few weeks confirmed what litigators had already worked out: nothing typed into an AI system is privileged. It is treated the same as a credit card swipe or a phone log: discoverable, subpoenable, just another electronic record for a court to weigh.

That is a real shift, and anyone advising clients on what they put into an AI system is right to take it seriously. But it exposes something odder than the shift itself. Courts have quickly built a working process for one direction of AI evidence: what a person said to an AI system. They have built almost nothing for the other direction: what an AI system said about a person, or a company, on a given day.

The asymmetry is structural, not accidental. When a prosecutor wants a defendant's chat history, there is a clear path: subpoena the company that operates the system, which retains the logs on its servers, tied to an account, timestamped by design. The record exists because the company needed it to exist, for billing, for abuse review, for the ordinary running of the service. The chain of custody runs through one company's infrastructure, and a court can compel it.

Now flip the question. A company believes an AI system told a customer, an investor, or a journalist something false and damaging about it: an invented lawsuit, a fabricated regulatory finding, a defamatory character description. Nobody has an account to subpoena. The output was not addressed to the company; it was addressed to whoever happened to ask. No AI company keeps a queryable log of what it told the public about a given organisation on a given date. Even if one did, the organisation being described has no standing to compel it the way a prosecutor can compel a defendant's own account. The one record most likely to exist is a screenshot taken by whoever first noticed the problem: easy to dispute, easy to dismiss as unrepresentative of what anyone else saw, and silent on whether the same system said something different an hour later.

This is why the current run of AI defamation cases keeps snagging on the same three questions: what exactly was said, when, and to how many people. A Delaware court let one such case proceed largely because the AI system itself had supplied a reach figure: an admission from the machine standing in for a record nobody else had kept. That is not a repeatable strategy. Most reputational harm from an AI system's output will never come with the system helpfully quantifying its own damage.

Regulators are starting to notice the liability question without addressing the evidence question underneath it. New disclosure rules force AI systems to say they are AI. New court rulings say a company cannot blame its own chatbot for what it told a customer. New liability frameworks are emerging in several jurisdictions for AI-generated falsehoods about real people. None of them answer the practical question a lawyer asks first: how do you prove, to a reproducible standard, what a system said, sourced from where, and at what precise moment, before the next update quietly rewrites the answer.

The evidentiary infrastructure being built right now runs one way, because that is the direction with a paying customer already attached. Subpoenas serve prosecutors and civil litigants who want a defendant's own words on the record. Nobody is building the equivalent infrastructure for the party being described, because until recently that party had no obvious way to ask for it, and no vendor had a commercial reason to offer it. That gap will likely close the way most evidentiary gaps around a new technology close: through a case where the absence of a record decided the outcome. The companies and counsel who do not wait for that case to be their own are the ones already keeping their own copy of the record, sealed and timestamped, on the day the AI system spoke.