How Courts Are Ruling on Fake AI-Generated Evidence

Risk Digest lane

This record answers a working filing question: if an image, video, or text exhibit is AI-contaminated, or is accused of being AI-contaminated, what has actually happened in court so far? The filing consequences now include dismissal, sanctions, exclusion fights, emergency protective-order consequences, and the cost of proving that a challenged exhibit is real or fake.

Not legal advice. Last verified: August 26, 2026, UTC. Review status: source-status checked, with primary-order gaps marked rather than filled in. The records below distinguish confirmed court action from secondary reporting because a bad exhibit does not become less dangerous merely because the docket is hard to reconstruct.

The current evidence picture is uneven. A reported Alameda County case produced dismissal and sanctions after self-represented plaintiffs filed AI-generated video testimony as authentic evidence; a Florida protective-order incident produced an arrest and jail time before fabricated text messages were reportedly exposed; and other courts have split between excluding AI-enhanced video and admitting or refusing to exclude challenged material under familiar authentication rules.

Record grid by court response

Terminating sanctions: the Mendones filing problem

Mendones v. Cushman & Wakefield is the record that should make risk teams stop treating fake AI evidence as a seminar topic. The reported filing was not merely an online clip, a social-media rumor, or a demonstrative exhibit labeled as synthetic. It was AI-generated video testimony submitted as authentic evidence by self-represented plaintiffs in Alameda County Superior Court. The reported result was dismissal of the case and sanctions against the filers on September 9, 2025.[1][2]

That is the cleanest sanction-and-dismissal signal in the current set, but it should not be made cleaner than the materials allow. The available sources for this digest are secondary summaries from NCSC and CU Boulder Today, not the sanctions order itself. The judge’s name, the court’s exact authentication analysis, and the precise sanction terms are therefore not stated here as confirmed from the primary docket. Before the case is cited in a brief, policy memo, CLE deck, or vendor pitch, the primary sanctions decision should be pulled and checked against the docket.

Even with that limitation, the filing posture matters. A court did not have to adopt a comprehensive AI-evidence rule to impose a case-ending consequence. The reported misconduct sat in familiar litigation territory: a party submitted evidence as real, the evidence was determined to be synthetic, and the court used existing authority to end the case and impose sanctions. The emergency falls first on the people who must unwind the filing: the opposing lawyer billing time to investigate provenance, the clerk classifying an abnormal evidentiary submission, and the judge deciding whether the record can be trusted enough to move forward.

For a filing lawyer, the lesson is not that every AI-associated exhibit is fatal. It is that an exhibit represented as authentic must have a provenance file before it reaches the courthouse. If the only support is a client’s assurance, a downloaded file, or a video with no original capture path, the risk is no longer theoretical.

Fabricated texts and the lag between filing and correction

The Florida incident reported by NCSC is not treated here as an evidentiary holding. It is a harm record. According to that report, a woman spent two days in jail after an ex-partner fabricated AI text messages that triggered a protective-order arrest; the charges were dropped only after about eight months.[1]

The important part is the lag. Fabricated messages can produce immediate process consequences before anyone has time to reconstruct the device history, subpoena account records, or test whether the screenshots came from an original messaging environment. By the time the forgery question is answered, the person targeted by the false record may already have been arrested, jailed, charged, or forced to spend months clearing the file.

That does not make the incident a general rule about protective-order evidence. It does make it a concrete warning about screenshots and message exports. Text evidence needs the same intake discipline as images and video: original device preservation where possible, account-level corroboration, timestamp consistency, and a clear record of who produced the file and when.

Admissibility has not moved in one direction

Washington v. Puloka is the exclusion-side record in the current set. Secondary sources describe the court excluding AI-enhanced video because of reliability concerns.[3][4] That is a different risk signal from Mendones. The exhibit was not reported here as a fabricated party submission used to obtain sanctions; the problem was whether the enhanced video was reliable enough for admission.

Huang v. Tesla and USA v. Khalilian sit on the other side of the admissibility contrast. The available secondary summaries treat them as examples in which courts admitted evidence or refused to exclude evidence when the deepfake theory was thin.[3][4] The record does not support the broader claim that courts are now rejecting AI-contested exhibits as a class. It supports a narrower and more useful conclusion: when a party cannot make a concrete showing of fabrication or unreliable methodology, an AI accusation may not carry the day.

That distinction matters at the counsel table. A party offering a challenged exhibit may only need to clear a low sufficiency gate. A party attacking it needs more than a suspicion that the image looks strange, the voice sounds off, or a generative model could have made something similar. The deeper fight turns on what the record contains: the original file, the capture device, metadata, custody history, platform records, expert method, and whether the alleged alteration changes a fact that matters.

The governing gate is still familiar evidence law

The present rule environment is not “courts reject AI evidence.” It is closer to this: courts apply familiar relevance and authentication thresholds to unfamiliar files, and the result depends heavily on the judge and the showing made. FRE 401 asks whether evidence has a tendency to make a fact more or less probable and whether that fact matters to the case. FRE 901 requires enough evidence to support a finding that the item is what the proponent claims it is. In ordinary practice, that is a low sufficiency threshold, often leaving the harder reliability dispute for weight, cross-examination, competing experts, or a separate expert-method challenge.[3][4]

Rule 702 and Daubert enter when the court is asked to rely on expert methods, including detection or enhancement methods. A detection expert who can explain the method, error concerns, limits, inputs, and audit trail is in a different position from a party offering only a detector score. A proprietary number with no transparent methodology is a weak courthouse object, even if it looks precise.

The Kennedys discussion of an “86% fake” detector output is useful for that reason only: it illustrates how a bare probability score can be mistaken for proof. It should not be treated as an accuracy benchmark, a validation study, or a general detector-performance number.[5] If the score cannot be traced to a reproducible method, known inputs, preserved files, and a reviewable analysis path, it gives the opposing lawyer an opening rather than a clean answer.

Pending reforms are a status track, not the rule today

As of Q3 2026, the reform track remains pending for purposes of this digest. The materials identify proposed FRE 901(c), California Rule of Evidence 707, and California SB 970 as proposed or pending responses to AI-generated or deepfake evidence, not as a uniform national rule already governing every filing.[1][4][6]

Until a controlling rule changes the filing burden in a particular court, the safer working assumption is record-by-record exposure. A sanctions court will ask what was filed and represented. An admissibility court will ask what the proponent can show. A judge facing a detector score will ask, or should ask, what method produced it and whether anyone can test the path from file to conclusion.

Verification steps that would have narrowed the risk

The verification answer is not one magic detector. It is a file history that someone else can audit. For image and screenshot intake, start with source provenance and original-file preservation, then compare metadata, device history, account records, and any available platform trail. The same basic discipline appears in the site’s image-claim verification workflow and in companion records involving AI evidence contamination and AI-generated imagery authentication.

For video, the companion bodycam authentication workflow is the better model than a detector-first review: identify the source system, preserve the original, track exports, compare timestamps, and leave a record another reviewer can repeat. The same source-status discipline is used in Risk Digest records such as AI facial reconstruction cold cases, where confirmed and reported facts have to stay in separate lanes.

The companion record on an Arizona sentencing court crediting AI-generated video shows the neighboring problem: not every AI-generated courtroom use is a fake-evidence sanctions case, but every use needs a posture label before anyone can assess the risk. The same is true when AI-derived material enters an investigative file.

Fake AI-generated evidence is now a documented, sanction-backed litigation risk. The harder point for 2026 is that uniform evidentiary reform has not arrived. Exposure still turns on the record in front of the judge: what the file is claimed to be, who can prove where it came from, whether the original was preserved, what the metadata and custody trail show, and whether any expert method can survive scrutiny.

References

1. AI-Generated Evidence: A Threat to Public Trust in Courts — NCSC
2. Deepfakes and AI in the courtroom: Report calls for legal reforms to address troubling trend — CU Boulder Today, November 17, 2025
3. Deepfakes, evidence & authentication — Thomson Reuters Institute
4. Deepfaked Evidence: What Case Law Tells Us About How the Rules of Authenticity Needs to Change — BTLJ, June 2025
5. 86% Fake, 100% Admissible: Rethinking Evidence in the AI Era — Kennedys Law, 2026
6. Deepfakes in the Courtroom: Problems and Solutions — ISBA, March 2025