Two things are happening in federal courtrooms at once. Courts are deciding how to handle evidence that a machine produced — and, separately, how to handle a party who claims that perfectly real evidence was fabricated by one. Both questions are being answered right now, through amendments to the Federal Rules of Evidence, and both will land on ordinary businesses long before they land on anyone famous.
If your company keeps surveillance footage, access logs, recorded calls, or any report that an AI tool helped produce, this is about your records.
Proposed Rule 707: machine output gets held to expert standards
There has been a quiet gap in evidence law. When a human expert testifies, their methodology is subject to scrutiny under Rule 702 and the Daubert standard — opposing counsel can probe whether the method is reliable and whether it was applied correctly. But when a machine produced the conclusion and no expert sponsored it, that scrutiny had no obvious hook.
Proposed Federal Rule of Evidence 707 closes that gap. It would require machine-generated evidence offered without a sponsoring expert to meet the same standards as expert testimony: that the output is based on sufficient facts or data, produced using reliable principles and methods, and reflects a reliable application of those methods to the case at hand.
The rule does not restrict AI in court. It applies the same reliability test to a machine that has always applied to a human expert.
On the effect of proposed Rule 707
The practical translation: an algorithm that reaches a conclusion is now treated as an expert witness, and it can be cross-examined like one. If you cannot explain what the system did, what data it used, and why its method is sound, its output may not survive a challenge.
Where the rule actually stands
This is a proposal working its way through federal rulemaking, not current law. That distinction matters, and it is worth being precise about the timeline rather than implying the rule is already in force.
| Stage | Timing | Status |
|---|---|---|
| Advisory Committee votes 8–1 to publish for public comment | May 2025 | Complete |
| Public comment period closes | 16 February 2026 | Complete |
| Advisory Committee final report | June 2026 | Expected |
| Judicial Conference review | Autumn 2026 | Projected |
| Transmission to the Supreme Court | By May 2027 | Projected |
| Earliest possible effective date | 1 December 2027 | Projected |
The first two rows are confirmed. Everything below them follows the standard Rules Enabling Act sequence and is not yet scheduled — the process can slow down, and the text can change before it is adopted. What is not in doubt is the direction of travel.
The other half: proposed Rule 901(c) and challenging a deepfake
Rule 707 governs evidence a machine made. A separate proposal addresses evidence a machine may have faked.
Proposed Rule 901(c) would create a structured path for challenging electronic evidence as fabricated. The mechanism is a burden shift, and the threshold is deliberately meaningful:
- The party challenging the evidence must show it is more likely than not fabricated or altered — the preponderance standard, not mere speculation.
- If they clear that bar, the burden moves to the party offering the evidence, who must then show its probative value outweighs its prejudicial effect.
A companion amendment to Rule 901(b)(9) would require a party offering AI-generated material to describe the software or program used and show that it produced valid and reliable results in that specific instance — not merely that the tool is generally reputable.
The liar’s dividend, and why it is the bigger business risk
Most coverage of deepfakes focuses on fake evidence being admitted. For an ordinary business, the likelier and more damaging scenario runs the other direction: your genuine evidence gets dismissed because deepfakes exist.
Legal scholars call this the liar’s dividend. Once everyone knows convincing fakes are cheap to produce, simply asserting that something is fake becomes a viable tactic — and the mere possibility of fabrication does some of the work regardless of whether the challenge succeeds.
Consider what that does to an ordinary dispute. You terminate an employee for misconduct captured on camera and they claim the footage was manipulated. You submit call recordings in a contract dispute and opposing counsel suggests the audio was synthesized. You produce access logs showing when a former contractor entered a system, and the response is that logs are trivially edited.
Rule 901(c)’s preponderance threshold is designed precisely to stop this from becoming a free move — a challenger has to actually show fabrication is more likely than not. But you still have to be able to answer. The businesses that will struggle are the ones that never thought about how they would prove their own records are authentic.
Evidence hygiene: what to do before you need it
Authentication is far easier to establish in advance than to reconstruct under deposition. None of the following is expensive; all of it is difficult to retrofit.
- Preserve originals, unedited. Keep the original file with its metadata intact and work from copies. A re-encoded, re-compressed, or re-exported video has lost most of what would corroborate it.
- Keep hashes. Record a cryptographic hash of important files when they are created. It is a few lines of automation and it makes tampering demonstrable rather than arguable.
- Document chain of custody. Who collected it, when, from which system, and where it has been stored since. This is routine in forensics and almost absent in ordinary business record-keeping.
- Log the system, not just the event. Camera model, firmware, clock synchronisation, retention settings. Authentication questions are usually about the system’s reliability, not the individual file.
- Label AI involvement honestly. If a tool transcribed, summarised, enhanced, or flagged something, record that it did and which tool it was. Under proposed Rule 901(b)(9), undisclosed AI involvement discovered later is far worse than disclosed AI involvement declared up front.
- Do not “clean up” evidence. Sharpening a frame or reducing audio noise to make something clearer is the fastest way to make it inadmissible. Enhance a copy, never the original, and disclose it.
Who this reaches first
Litigation and insurance will feel it soonest. Cyber-insurance claims already turn on log evidence, and an insurer disputing a claim now has a structured route to question whether those logs are what you say they are. Employment disputes, contract claims, and any matter resting on recorded communications follow closely behind.
There is also a compliance angle worth flagging. If your organisation uses AI in any process that could later produce evidence — fraud scoring, anomaly detection, automated incident reports — the reliability of that system is now a legal exposure and not just an engineering concern. That is a different conversation than the one most companies are having about AI agent security, though the two are related: both come down to whether you can account for what your systems did.
Frequently asked questions
Is Rule 707 currently law?
No. It is a proposal that completed its public comment period on 16 February 2026, with an Advisory Committee report expected in June 2026. Under the standard rulemaking timeline it could not take effect before 1 December 2027, and the text may change before then. Courts can already address AI evidence and deepfakes under existing rules in the meantime.
Does this apply in Florida state court?
The Federal Rules of Evidence govern federal court. Florida has its own evidence code, which closely tracks the federal rules in many respects but is amended separately. States frequently follow federal developments in this area, so the direction is likely similar — but the timing will differ, and you should not assume a federal rule applies to a state matter.
Our security camera footage — is it still usable as evidence?
Yes, and nothing in these proposals changes that. What changes is how carefully you may need to establish authenticity if it is contested. Preserve the original export rather than a phone recording of a monitor, note the system and its clock settings, and avoid editing. Those three habits resolve most challenges before they start.
We used AI to summarise an incident report. Is that a problem?
Not inherently — but disclose it and keep the underlying source material. The risk is not that AI touched the document; it is that AI touched it, nobody said so, and opposing counsel found out. Retain what the summary was drawn from so the conclusion can be traced back to primary evidence.
If you are already in a dispute
Delana Technologies provides digital forensics and expert witness testimony for law firms, insurers and businesses — breach investigation, evidence authentication, and technical declarations that hold up under examination. If authenticity of digital evidence is at issue in a matter, or you want your records reviewed before it becomes one, call 239.414.5126 or contact us.
This article is general information about developments in the rules of evidence and is not legal advice. Delana Technologies is a cybersecurity and digital forensics firm, not a law firm — consult qualified counsel about any specific matter.
Sources: Advisory Committee on Evidence Rules, proposed Rule 707 and proposed Rule 901(c); Eckert Seamans, “Proposed Federal Rule of Evidence 707: What Businesses and Their Lawyers Need to Know”; Barnes & Thornburg, “New Evidence Rule 707 Would Set Standards for AI-Generated Courtroom Evidence”; Maryland State Bar Association, “Amendments to the Federal Rules of Evidence Under Consideration To Address AI and Deepfakes.”