//Rule 707
How to Make Your AI Investigation Survive Rule 707
A new rule for machine-generated evidence is being written, and it just got harder to predict. Here is where Rule 707 actually stands, and how to make your AI-assisted work defensible now, no matter how the rule lands.
Status as of July 2026
Proposed FRE 707 is not law and not scheduled. At its May 2026 meeting the Advisory Committee on Evidence Rules declined to advance it, revised the text, and deferred it for further study together with the deepfake problem. The “effective December 2027” date still repeated across the web is outdated.
The rule is a moving target. The reason to prepare now is not the deadline. It is that the standard it points to, Rule 702, is already the law.
The AI-assisted findings that hold up, today and under any version of 707, are the ones with sourced inputs, a documented method, human verification, a chain of custody, and disclosure already in place. That is the Evidence-Grade AI Standard. This is general information, not legal advice.
//The rule
What Rule 707 would require
As drafted, when machine-generated evidence is offered without a human expert to sponsor it, and it would be treated as expert opinion if a human said it, a court could admit it only if it meets the Rule 702 reliability test: it helps the fact-finder, it rests on sufficient facts and data, it derives from reliable principles and methods, and those methods were reliably applied to the facts of the case. The proposal carves out simple instruments. Everything with judgment baked into it, the risk scores, the classifiers, the enhanced media, the analytics, would be in scope.
The text is being revised, so the final wording may shift, and the deepfake question is now being studied alongside it. But the direction is not in doubt, because 707 does not invent a new standard. It extends the reliability test courts already apply to human experts. That test is live today under Rule 702 and Daubert, which is why the smart move does not depend on 707 passing at all.
The practical consequence is discovery. If your AI output is in the case, the other side can already demand to know how it was produced: what data, what model, what prompts, what verification. The party that cannot answer loses the evidence. The party that documented it up front does not.
//The preparation
How to be ready
Readiness is not a filing you make later, and it does not wait on the rule being finalized. It is a record you build as the work is done. Six things carry an AI-assisted finding through a reliability challenge, under Rule 702 today and under whatever 707 becomes, and they map directly to the Evidence-Grade AI Standard.
Sourced, lawful inputs
Every input the finding depends on is identified, preserved, and lawfully obtained under the rules that govern the work without AI.
A documented, repeatable method
The method is written down so another examiner could follow it, with the specific system and version recorded.
Independent human verification
A qualified human verifies every output against source evidence before it becomes a finding. Unverified AI output is a lead, not a conclusion.
Chain of custody and provenance
A record of what data, what system and version, what steps, who verified, and when, preserved to survive challenge.
Disclosure, calibrated
How AI was used is stated plainly, and no claim overstates what the method can support.
Reproducibility
An independent examiner, given the record, could follow it and reach the same result.
The full framework is the Evidence-Grade AI Standard. Validating a specific workflow against it, so it holds up under Rule 702 today and under whatever 707 becomes, is the AI Investigation Audit.
//Why me
Who prepares you for this
The seat Rule 707 creates is narrow. Law firms write about the rule but do not build the systems. Detection vendors sell a score no one can defend. Compliance shops audit for governance, not for the courtroom. Preparing an AI-assisted investigation for 707 takes someone who is a licensed investigator, a certified forensic examiner, and the person who builds the AI, all at once. That is the work I do: I validate the workflow, document the reliability record, and hand you findings and an affidavit that hold up, so the rule is something you are ready for rather than something that surprises you.
//Questions people ask
Rule 707, answered straight.
What is Federal Rule of Evidence 707?
Proposed FRE 707 is a draft rule that would hold machine-generated evidence, offered without a sponsoring human expert, to the same reliability standard the rules already apply to expert testimony under Rule 702. In plain terms, if an algorithm's output would be treated as expert opinion when a human said it, the output would have to clear the same reliability bar. Simple instruments like thermometers and scales would be exempt. It is a proposal still being written, not current law. This is general information, not legal advice.
Did Rule 707 pass? Is it in effect yet?
No. As of mid-2026, Rule 707 is a proposal still in drafting, not law. A great deal of commentary online still says it takes effect December 1, 2027 — that is now outdated. At its May 2026 meeting the Advisory Committee on Evidence Rules declined to advance the rule, revised the text, and held it for further study together with the separate problem of deepfakes. Nothing is in effect. The reliability standard it points to, Rule 702, already governs expert evidence today.
When will Rule 707 take effect?
The timeline is open. After a public comment period that closed February 16, 2026, the Advisory Committee did not vote the rule forward at its May 2026 meeting; it revised the proposal and deferred it for more study. Any December 2027 date that once circulated is no longer realistic, and it is possible the rule changes substantially or does not move forward at all. That uncertainty is the point: the reliability discipline below holds up no matter what the final rule says, because it tracks Rule 702, which is already the law.
Does Rule 707 cover deepfakes?
Not directly, and that gap is now part of the story. Rule 707 as drafted addresses evidence a party admits is machine-generated. It does not resolve disputed-authenticity deepfakes, where the fight is whether the media is real at all. At its May 2026 meeting the committee folded the deepfake problem into the same workstream as 707 for further study, so the two are now moving together. Either way, a disputed deepfake still needs a qualified examiner. Both are work I do.
How do I make my AI-assisted evidence defensible while Rule 707 is unsettled?
Build the reliability record before the case needs it, on principles that survive whatever the final rule says: sufficient and lawfully sourced inputs, a documented and repeatable method, the specific system and version recorded, independent human verification of every output, a defensible chain of custody, and disclosure of how AI was used. That is the Evidence-Grade AI Standard, and validating a workflow against it is what the AI Investigation Audit does. Because it tracks Rule 702, it is defensible today, not just whenever 707 lands.
Who needs to prepare for this?
Anyone who offers an algorithmic output as evidence: risk scores, classification results, AI-enhanced media, predictive-coding or eDiscovery output, and forensic analytics. Law firms, insurers and SIU teams, and investigation and forensics shops all fall in scope. If AI touched the evidence, the other side can already demand you prove the machine was reliable under Rule 702, regardless of whether 707 is ever adopted.