Estimated reading time: 12 minutes
A courtroom confrontation in Utah, a new EU deadline, and three overlooked rulings point to where the debate over digital evidence is really heading.
Altered images cause harm long before anyone determines who must bear the costs. Digital evidence is now causing problems in connection with cyber insurance claims, and the courts are already dealing with it. This situation was something the prosecutors in the preliminary hearing of Tyler Robinson experienced on July 6. Robinson is accused of having killed Charlie Kirk at Utah Valley University. The prosecutors attempted to have a compiled surveillance video admitted as evidence, but Fourth District Judge Tony Graf refused to do so.
No one disputed that someone had altered the footage by adding annotations and close-up sections. The prosecutors openly made those modifications for the hearing. Nevertheless, Graf blocked the video and agreed with the defense on one issue: it was not the original file, and he had requested the unaltered version. Because the court broke for the evening, prosecutors had time to produce the revised footage and present it to the judge again the following day.
It was just a brief point during the extended hearing. The hearing lasted five days and included a great deal more content: DNA evidence, alleged confessions, and witness testimony. Graf has not yet decided whether the case will go to trial. The decision will be made at the following hearing, on September 1. Nevertheless, the footage exchange made an impression on Mike Nelson, VP and Field CTO at DigiCert.
“I can’t offer legal conclusions or speak to how liability would be assigned in specific cases,” said Nelson. He added, “What I can say is that as digital content becomes easier to manipulate, the ability to understand its origin and history is becoming increasingly important for anyone making decisions based on it. Provenance technologies like C2PA are designed to provide that context.”
The Fix DigiCert Happens to Sell
Nelson’s interpretation of the incident is limited and may well be correct. Editing was never really the issue, whether in court or elsewhere. The challenge is proving what happened to a file after it was captured.
C2PA, the Coalition for Content Provenance and Authenticity, addresses that problem through an open standard that attaches a signed record to a file, documenting where it originated and what happened to it.
The standard does not enable you to determine whether the content is true, it only provides information about its history.
Current file metadata has a significant limitation: it stops tracking the file once someone captures it. For example, if you edit a photo on your phone, the metadata may still show the original time and date but provide no record of the changes you made.
C2PA is DigiCert’s response to that limitation.
DigiCert was one of only two certificate authorities on the C2PA Trust List in early 2026. That’s the infrastructure behind every signed credential. Nelson isn’t describing this market from the outside.
Reconstructing a Timeline
That’s context, not a disqualifier. CINI asked him the same questions anyway. Where insurance came up, his answers were more careful than a typical vendor pitch. Asked how a broken chain of authenticity could complicate claims, he pointed to what adjusters already do:
“Cyber claims often depend on reconstructing a timeline: who initiated a transaction, what communication was received, when systems became unavailable, and what changed during an incident. Cryptographically verifiable provenance can help establish the source and history of supporting files, recordings and other digital artifacts.”
Nelson went on to say, “That could be especially valuable in business-email compromise, deepfake-enabled impersonation, fraudulent payment instructions, ransomware and business-interruption claims.”
The fix, in his telling, starts well before a claim is filed. Organizations should preserve native files, access logs, hashes, and audit trails from the beginning of an incident, he said, rather than “try to reconstruct authenticity after the fact.”
Cyber Insurer Verification
On whether insurers will eventually require this kind of verification, Nelson was careful not to oversell where the market actually is:
“This will likely be risk-based, not universal. Underwriters may look at who can create and sign content, how keys and identities are protected, and whether provenance is preserved through editing and distribution.”
He was equally direct about what C2PA isn’t. It is “not a deepfake detector or a guarantee of truth,” he said. It offers cryptographic signals about source and history, nothing more. Authenticated provenance, in his words, “should not be the only test of trustworthiness.” A missing credential, on its own, doesn’t mean content is false or that anyone acted negligently.
The Deadline That Already Passed
Nelson’s underwriting prediction falls within a broader shift. He didn’t raise it, but the timing matters. On August 2, Article 50 of the European Union’s Artificial Intelligence Act became enforceable. This rule requires machine-readable labels on AI-generated and AI-altered content, including deepfakes. It applies whether or not the system counts as “high-risk” under the Act’s separate risk framework.
What is worth mentioning is the timing. The Digital Omnibus package of the Act went through the EU’s legislative process in June and moved the compliance deadline for high-risk systems to December 2027. However, Article 50 was not included in this delay and therefore remained on schedule.
Systems that are already on the market have until 2 December 2026 to meet the marking requirement. All the other provisions came into effect this month. Fines amount to €15 million or 3% of the company’s worldwide annual revenue. The rule is not only applicable to companies based in the EU. Any provider or deployer who serves users in the EU is subject to it, no matter where they are located.
The US Side of The Pond
For CINI’s readers in the United States, that final point is the most important. A U.S. based company with no office in the European Union can still qualify as a “deployer” under Article 50 simply by publishing content generated or edited with AI to an EU audience. Now provenance and disclosure have turned into a compliance issue in Europe, with a strict deadline set. The discussion in the United States is still mainly taking place within advisory committees and in vendor pitch decks. Knowing about these deadlines will enable US and non-EU companies to get ready for the upcoming legal requirements.
Agentic AI and Cyber Insurance: The Authorization Gap – PODCAST
Our panel tackled why the biggest risk in agentic AI is not the model. It is the access you give it. Featuring Julia Garcia-Trombley (CertX), Jeremy Epstein (Mayflower Specialty), Rich Gatz, Esq., FIP (Arch Insurance Group Inc.), and Tristan Morris (SplitSecure).
What the Case Law Actually Shows
Nelson’s prediction rests on where the legal ground is moving. So CINI looked at what’s already on the record. Three data points stand out. None of them come from DigiCert, and together they complicate the story a little.
Puloka went against the side offering the AI enhancement, and it wasn’t the prosecution.
The earliest reported US ruling on AI-altered video didn’t involve a deepfake. Nobody accused anyone of fabrication. In State of Washington v. Puloka, a King County judge excluded video the defense had enhanced with AI software. The goal was to sharpen a blurry ten-second phone clip of a 2021 shooting. The court’s objection wasn’t that the enhancement misrepresented events. It was that the relevant technical community hadn’t validated the method. The judge said forensic video analysts were the right community to ask, not commercial filmmakers. Forensic analysts hadn’t accepted the software. That’s a useful correction to a common assumption: that AI-altered evidence tends to help whoever makes the accusation. In this first case on record, it was the defense trying to get AI-enhanced footage admitted. The court kept it out.
Mendones is the closer match to an actual insurance claim.
The more useful case for CINI’s audience happened in a California housing dispute, not a criminal court. In Mendones v. Cushman & Wakefield, self-represented plaintiffs filed for summary judgment. They backed the motion with video they said showed a witness testifying. Judge Victoria Kolakowski noticed something off.
The witness’s expression barely moved. Her voice didn’t track naturally with her words. The same gestures repeated. Kolakowski concluded the footage was AI-generated. It stood in for a real witness who did appear elsewhere in the case, credibly. What actually broke the case open wasn’t visual inspection alone. Metadata embedded in the video traced it to an iPhone 6. That model didn’t support the plaintiffs’ claim about how the footage was captured. That’s a real mismatch: a file’s own history versus the story built around it. It’s exactly what provenance tools are meant to catch. Here, though, it was the file’s own leftover metadata that caught it, not a signed credential. The court found the plaintiffs had submitted fabricated evidence.
It dismissed the case with prejudice, the harshest sanction short of a criminal referral. This is a civil case. It’s the closest thing on record to an actual insurance claim scenario a party fabricated supporting media. Details the file was never supposed to carry gave it away.
The federal rule everyone cites isn’t in force yet.
The fix most often mentioned for all of this is a new Federal Rule of Evidence, 707. It would govern machine-generated evidence. A companion amendment to Rule 901 specifically targets deepfake claims. Both remain proposals. Neither advanced to final approval at the Advisory Committee on Evidence Rules’ May 2026 meeting.
Public comments closed in February. At the Standing Committee’s June 2026 meeting, only one evidence rule change moved forward. It was an unrelated update to Rule 609. The December 2027 effective date that keeps circulating on panels and in vendor materials isn’t accurate right now. That means the courts decided both Puloka and Mendones under existing rules.
Neither case required an AI-specific evidence rule. Judges haven’t been waiting for new rules before excluding unreliable media. They’ve been doing it on a case-by-case basis, under the authority they already have.
What Provenance Can’t Fix
C2PA isn’t a solved problem. Nelson didn’t pretend otherwise. Its biggest weakness is also its most common one. A screenshot, or a phone pointed at a screen, creates a brand new file. That new file has no cryptographic link to any original. Most platforms people actually use to send media still strip embedded credentials on upload. Messaging apps do it. Social platforms do it. Plenty of ordinary content systems do it too, intentionally or not. Even a fully intact, validly signed manifest only proves one thing. A specific device or piece of software made a specific claim, at a specific time. It doesn’t prove the claim was true. A camera can sign a photo of a screen showing a deepfake. That signature will verify perfectly.
That leaves the question CINI opened with unanswered. It’s worth saying plainly, rather than letting a provenance conversation quietly stand in for it. A cryptographic record can show what happened to a file after it left the camera. It cannot tell a court, an adjuster, or a jury who should have caught the problem first. It cannot say who pays when nobody did. Provenance reconstructs what happened. Liability is a separate argument, made by lawyers and decided by courts. Nothing here settles it yet: not the Robinson hearing, not the Mendones sanction, not Brussels’ new marking rule.
FAQ – Digital Evidence Cyber Insurance Claims
What was the outcome regarding the surveillance video in the case involving Charlie Kirk?
On July 6, 2026, a Utah judge refused to admit the surveillance video as evidence. The hearing concerned Tyler Robinson, who had been charged with murdering Charlie Kirk. Prosecutors had altered and magnified the video for use in court. No one disputed that it was fake. The judge simply asked for the original file before making a decision, and the prosecutors provided it; the hearing took place the following day.
What does C2PA stand for?
It stands for the Coalition for Content Provenance and Authenticity and is an open standard. It includes a signed record with a digital file, indicating where the file originated and the history of any edits made to it. The standard does not prove that the content is true or accurate; it only establishes that a particular device or piece of software made a specific, traceable statement.
Does cyber insurance currently cover deepfake fraud?
Whether cyber insurance covers deepfake fraud depends on the insurer and when the policy was last renewed. Since late 2024, a number of insurers have introduced exclusions or have made more specific the wording relating to social engineering. In some cases, they now offer separate deepfake endorsements rather than including the risk in their standard cyber or crime coverage. If you want to find out how your policy stands, get in touch with your broker. Make sure that the definitions and exclusions cover content that has been generated by or altered using AI.
Has there already been established a federal rule concerning evidence that is generated by AI?
Not yet. Two proposals are still under consideration: Federal Rule of Evidence 707, which applies to evidence produced by machines, and an amendment to Rule 901 addressing deepfakes. Neither of these has been submitted for final approval according to the committee’s latest meetings in 2026. The sometimes-quoted effective date of December 2027 is not accurate at this time.
What constitutes Article 50 of the EU AI Act, and is it applicable to companies in the United States?
Article 50 of the European Union’s Artificial Intelligence Act requires providers and deployers to make machine-readable disclosures for content generated or altered by AI, including deepfakes. The provision took effect on August 2, 2026. It applies when users in the EU can access the content, regardless of where the company operates. As a result, a U.S.-based business that publishes AI-enhanced media for an EU audience could fall within the article’s scope.
What should organizations do at this stage, prior to any of these issues being resolved?
They should preserve the native files, access logs, hashes, and audit trails from the very beginning of any incident, rather than attempting to reconstruct them later. They must ask their broker directly whether the existing policies account for AI-generated or AI-altered content. This doesn’t require anyone to wait for new evidence rules or a finalized provenance standard. Investigators can apply the same documentation practices that have always mattered in digital investigations, just earlier and more consistently.
Related Cyber Insurance Posts
- Beazley Security Launches Exposure Management To Counter AI-Assisted Attack Risk(Opens in a new browser tab)
- Cyber Insurance Goes from “Hard to Sell” to “Hard to Buy” — What Happens When We Reach Cyber Insurance 3.0? (Opens in a new browser tab)
- Bitdefender Targets Business Email Compromise With Dual-Layer Email Security Platform(Opens in a new browser tab)
- The $26 million Video Call from Hell: Munich RE(Opens in a new browser tab)