Can you use an AI model's output as evidence in a US federal court, and what would proposed Federal Rule of Evidence 707 require you to show about how it was produced?

What should I check, step by step?

1. Decide early whether you'll offer AI output through an expert. Under the revised draft, an expert would ordinarily be required, and Rule 702 already applies when a testifying expert relies on AI.

2. Record each run you might rely on at the time it happens: the model and dated version, a hash of the exact input and output, the settings, and when it ran. A screenshot or a re-run months later won't show what that version produced that day.
3. Keep the validation behind the tool: tests in circumstances similar to your case, error rates and who ran them. The draft Note lists validation in similar circumstances and independent evaluation among the things a court would usually consider.
4. Know what the system keeps and deletes. The draft Note says courts should review what records the system keeps, what is deleted, and whether the output can be reproduced or audited.
5. Be ready to give the other side enough to test it. The draft Note says courts should consider whether the opponent has received information about how and why the process works, similar to what a human expert would disclose.
6. Plan for notice. The revised draft makes AI evidence admissible only if the proponent gives all parties reasonable notice, so the reliability fight happens before trial.
7. Don't assume a certificate is enough. A Rule 902(13) certification can authenticate a record generated by an electronic process, but the draft Note says it doesn't establish reliability.
8. Check whether reliability matters to your point. The draft Note says the rule wouldn't apply where the evidence isn't offered for its accuracy, for example to show that a driver relied on an AI's directions.
9. If you used AI to enhance or edit a photo, video or recording, expect a challenge. Courts have excluded AI-enhanced video and AI-modified photos where nobody showed the tool was reliable.
10. Follow your court's local rules and standing orders on AI, and talk to a litigator about your case.

Is AI-generated output admissible as evidence in federal court today?

No Federal Rule of Evidence addresses AI by name. When an expert relies on AI to reach an opinion, Rule 702 applies, and the court decides whether the expert's testimony rests on sufficient facts or data and reliable methods reliably applied. The problem, as the Advisory Committee's May 2026 report puts it, is that Rule 702 isn't clearly applicable when AI output is admitted without any expert, either directly (for example under Rule 902(13)) or through a lay witness who merely pressed the button.

• The committee's reporter, quoting a draft article by Professor Jeffrey Bellin, describes the gap this way: authentication only asks whether an item is what the proponent claims it is, and the problem with a typical AI statement is not that it is forged but that it may be unreliable. The hearsay rule doesn't apply because an AI isn't a person and can't be cross-examined.
• Courts have still excluded AI-altered evidence under existing rules. In Washington v. Puloka (2024), a trial court applying an expert-reliability standard excluded a bystander video enhanced with a Topaz Labs AI tool, because the tool was largely untested and the experts couldn't rule out that it had added pixels that didn't represent what happened. In Linkugel v. City of Los Angeles (Cal. Ct. App. 2026), photos of a sinkhole that the plaintiff acknowledged were modified by generative AI were held inadmissible because no showing was made, through an expert or otherwise, that the AI use was reliable.

What would proposed Federal Rule of Evidence 707 require for AI evidence?

The version published for comment in August 2025 said that if machine-generated evidence is offered without an expert witness and would be subject to Rule 702 if testified to by a witness, it must satisfy Rule 702(a)-(d). It didn't apply to the output of simple scientific instruments. Commenters split: some said it was too broad and could reach routine evidence such as geolocation data, others that it was too narrow because it only covers evidence the proponent admits is AI-generated, and the Justice Department said Rule 702 already covers the problem.

• The revised working draft in the committee's October 2026 agenda book is titled 'Evidence Produced By Artificial Intelligence and Presented at Trial Without an Expert'. It requires the four Rule 702-style showings, adds that ordinarily the proponent must provide an expert to explain how the AI system reliably produced the evidence, allows other proof in exceptional circumstances, requires reasonable notice to all parties, excludes facts that can be judicially noticed, and defines AI as a machine-based system that can, for a given set of human-defined objectives, make predictions, recommendations or decisions influencing real or virtual environments. The reporter notes that the committee hasn't approved all of its language.
• The draft Committee Note lists what a Rule 707 analysis would usually involve: whether the inputs, including training data, are sufficient and representative for the case; whether the process has been validated in sufficiently similar circumstances; what records the system keeps, what is deleted, and whether the output can be reproduced or audited; whether it has been evaluated independently of the developer; and whether the opponent has had enough information about how and why it works. It adds that if a process can't be explained, a court should in most cases find reliability not established, unless an expert can show, for example through validation studies, that it has a low rate of error, and that courts should consider a limiting instruction that machine-generated evidence is subject to error.

Has Federal Rule of Evidence 707 on AI evidence been adopted?

The Advisory Committee voted in May 2025 to publish Rule 707, and the comment period ran from August 15, 2025 to February 16, 2026. At its May 7, 2026 meeting it decided not to recommend the published rule, tentatively approved the revised draft, and chose not to republish it for comment yet, preferring to hear from technology and AI experts first. Its October 15, 2026 meeting in Boston held expert panels on Rule 707 and on a draft Rule 901(c) on deepfakes. Neither rule has been adopted.

• Federal evidence rules move slowly by design. At the May 2026 meeting the committee chair noted that the deepfake rule couldn't take effect before December 1, 2028 even if it had been published for comment that summer. The draft Rule 901(c) would make a party claiming that evidence was fabricated by generative AI first present evidence sufficient to support that finding; the proponent would then have to show the item is more likely than not authentic. A Federal Judicial Center survey drew responses from 931 federal judges, 15 of whom reported having encountered a deepfake challenge.

How do you prove an AI exhibit matches what the model returned?

Whatever form the rule takes, the questions are about one specific run: which model and version produced the output, what it was given, with what settings, and whether anyone can check that the exhibit is what the system actually returned. Hosted models change and providers delete logs, so the evidence often has to come from your own records.

• A signed record with hashes of the input and output, made when the run happened, lets you show the exhibit hasn't changed since and which version you recorded. It doesn't show the output is reliable or correct; that still takes validation evidence and usually an expert.