AI evidence is under federal review in the US. But who proves what the AI did?

Two questions, one gap

The Federal Judicial Center's Advisory Committee on Evidence Rules is looking at AI evidence along two tracks ^[1]. One is about reliability: when an AI draws a conclusion, how does a court decide whether to trust it? The other is about authenticity: when something appears to have been generated by AI, how do you prove it was?

Think of it like a receipt and a till. The receipt says what you bought, and you can check it against the till roll. But if the shop's system has no roll, or if the roll can be edited quietly afterwards, your receipt and their word are all you have.

That is roughly where AI evidence sits now. Federal Rule of Evidence 702 governs expert testimony, and AI conclusions can be treated as such. Rule 901 requires authentication: the proponent must show that evidence is what it claims to be ^[1]. Both rules matter, but neither tells you who has to prove what the AI system did on the day in question, or how.

The burden question

In a civil case, the party offering the evidence generally has the burden of showing it is authentic ^[2]. In a criminal case, the prosecution must prove authenticity beyond reasonable doubt. That much is settled.

The difficulty is the next step. Say a lender's AI declines a loan application and the applicant disputes the decision six months later. The lender has to show that the AI's output is genuine. But what if the lender's log of what the AI did that day can be edited by anyone with administrator access? The applicant can question the log, and unless the lender kept an independent record at the time, the argument becomes one person's word against another's.

Article 12 of the EU AI Act, which applies to high-risk systems from 2 December 2027, requires automatic logging of events over the system's lifetime ^[3]. It says what to log. It does not say how to protect the logs, or who has to prove they have not been altered. That question lands with the organisation when the output is challenged.

What this means for you

If you run AI systems whose decisions, answers or generations could be questioned later, here are three questions worth asking now:

• Could you show exactly what your AI system output on a given day, and which version was running?
• If someone disputes a record, do you have something other than your own word to back it up?
• Who in your organisation can edit the logs, and would a change show up?

The US federal review is about courtroom procedure, but the underlying problem is practical. When AI evidence is challenged, the side without a trustworthy record starts on the back foot.

How a digital witness fits here

This is a case where a digital witness fits well. The thing to seal is each decision, answer or generation as it happens: the input reference, the model version, the output and, if there was one, the reviewer. Sealing takes a fingerprint of the record, not the record itself, so the content stays with you.

Say the lender's AI declined the loan on 14 March and the applicant questions it in September. The lender sealed the decision record when it was written. To check it, the lender calculates the fingerprint again and compares it with the sealed one. If they match, the record is exactly what was written in March. Change a single character and it will not match.

Anyone the lender shares the record with, including the applicant's lawyer or a court, can do the same check. The sealed fingerprint is on WitniumChain, a chain where any later change would show, even one made by us. So nobody has to take the lender's word for it.

What it does not do: it does not show that the decision was fair, lawful or correct, or that the log is complete. It shows that this record has not changed since it was sealed. The burden of proving the decision itself was sound still rests with the lender, but at least the argument is not about whether the record has been rewritten.

How it would work

1. The AI writes a decision to the log.
2. Your system sends a fingerprint of that record to WitniumChain through the REST API, the SDK or the MCP server and gets a witness ID back.
3. When the decision is questioned, you calculate the fingerprint again and compare it with the sealed one.

The first 2,000 witnesses a month are free.

What to watch

The Federal Judicial Center's review is ongoing, with no date set for final guidance ^[1]. Meanwhile, the EU's high-risk logging requirement in Article 12 takes effect on 2 December 2027 for systems under Annex III ^[3]. Neither rule changes who has the burden of proof, but both assume you have a log you can defend.

Sources

1. Roth, Andrea, "AI Evidence Is Under Federal Review. But Who Is Accountable for How AI Is Used?", SSRN, 2024, https://papers.ssrn.com/sol3/Delivery.cfm/7585978.pdf?abstractid=7585978&mirid=1
2. Federal Rules of Evidence, Rules 702 and 901, Legal Information Institute, Cornell Law School, https://www.law.cornell.edu/rules/fre
3. Article 12 of Regulation (EU) 2024/1689 (AI Act), AI Act Service Desk, European Commission, https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-12