Those involved in civil litigation or trial work within the past couple of years, have likely seen something play out like this: an AI-enhanced image, a machine-generated damages model, or an algorithmic authenticity check offered as if it were self-evidently reliable. The rapid evolution of technology is beginning to outpace most litigators’ instincts for how to effectively challenge it.

Although the legal world is trying to catch up, it is similar to a novice cyclist trying to keep pace with a cheetah. However, not everything is all gloom and doom. In August of 2025, the Advisory Committee on Evidence Rules released proposed Federal Rule of Evidence 707 for public comment. The proposed rule states as follows: when machine-generated evidence is offered without an expert witness, and that evidence would be subject to Rule 702 if testified to by a witness, the court may admit the evidence only if it satisfies the requirements of Rule 702 (a)-(d). In other words, you can no longer launder an unreliable opinion through a black box and simply call it “data.” If the output functions like expert testimony, it has to clear the same bar.

To date, commentators have highlighted some key problems with this rule that are worth noting. To begin, what exactly does “machine-generated” mean? While the rule was geared towards AI-generated evidence, the term remains undefined which leaves room for unnecessary ambiguity.Another unknown is what procedure should be used when admitting machine-generated evidence under Rule 702 when there may not be a qualified expert to cross-examine regarding the proffered evidence. These unknowns will likely jumpstart more disputes and lead to skyrocketinglitigation costs in ways that would be more expensive than traditional expert testimonychallenges.

For civil litigators, the practical shift is consequential. Today, a party offering an AI-generated damages projection, a predictive coding output, or an AI-enhanced forensic accounting report usually only needs to satisfy the requirements of relevance and authentication, under Federal Rules of Evidence 401 through 403 and 901, respectively, all without facing a true reliability hearing.

However should a revised Federal Rule of Evidence 707 come into effect, we will likely see “Rule 707 hearings” become as routine as Daubert hearings, complete with disputes over training data, validation methodology, error rates, as well as the tool’s susceptibility to the same biases the Daubert factors were designed to catch in human experts, to name a few.

All of that applies for evidence that a party admits is machine generated. Yet, what about evidence whose AI origins are disputed? What about those deepfake videos, the manipulated audio recording, or the doctored document that is being offered as authentic?

While the Committee has acknowledged that is technically a Rule 901 problem, this same Committee shelved a parallel proposal to amend Rule 901 to include AI generated media in 2025 as the committee preferred to let the existing “low threshold” authentication standard absorb the problem for now. The result, at least through 2026 and maybe even beyond, is doctrinal inconsistency: similar AI-disputed evidence may be admitted in one courtroom and excluded in the next, depending less on settled law than on a given judge’s comfort with the technology.

There is also palpable structural tension for anyone advising clients on AI forensic tools. Many of the most capable detection and analysis tools are proprietary, and their creators are reluctant to expose the methodology that took years and significant financial investment to develop. A reliability hearing under Rule 707 will likely require exactly that disclosure: error rates, training data composition, and validation studies. Vendors who will not provide it, or whose tools are accessible only to favored institutional users such as law enforcement, may find their outputs excluded regardless of practical accuracy.

The changes actually made remain to be seen. Rule 707 still needs Judicial Conference endorsement, Supreme Court approval, and congressional review, putting realistic adoption towards the end of 2027 at the earliest. However, it is clear change is coming as judges are already starting to scrutinize AI-generated evidence for reliability, utilizing existing Rules 702 and 901 to analysis the offered evidence ahead of any formal amendment.

What should civil litigators take away from this? Start building your record now. Document the AI tool’s methodology, retain the underlying data, and be ready to produce an expert. Technologyis not waiting for the rulemaking process to catch up, and lately, neither are the judges.