Deepfakes are popping up in the news all over the place, from their use in cybercrime and fraud to political propaganda, entertainment, and even non-consensual intimate images, which have triggered recent legislation. But they’re not just the subject of the law — they’re also a hot evidentiary issue in the courtroom.
What’s a Deepfake?
First of all, what’s a deepfake? The term’s a portmanteau of the words “deep learning” and “fake.” Deepfakes use machine learning — often including generative adversarial networks (GANs) — to create synthetic content.
A GAN has two parts. The first part creates images, and the second part distinguishes between those images and reality. Over time, the second part trains the first so that the images it generates become better and better. Eventually, they become virtually indistinguishable from reality. The image on this blog is a "deepfake."
Deepfakes carry with them two big evidentiary risks:
- Fabricated Proof: The first risk is that fake evidence might be passed off as proof that something happened when it actually didn’t. (Think of a deepfake video that looks like a Ring camera recorded your car accident.)
- The "Liar's Dividend": The second risk is the exact opposite — where judges and juries discredit your genuine, authentic evidence by claiming that it’s just a clever deepfake.
Courts Are Using Old Rules for New Technology
One of the tricky things about deepfake evidence is that it’s so new. The Federal Rules of Evidence were created long before artificial intelligence was on everybody’s mind. In fact, they were enacted into law on January 2, 1975, by President Gerald Ford. The first Apple computer didn’t even go on sale until the following year. So, the Rules are very much written with traditional media in mind, not AI-generated material.
Courts are in the process of creating authentication frameworks that help protect the public from AI-generated evidence, but old rules still do most of the work. Article IX covers authentication and identification.
- Rule 901(a) and (b) require that the person who presents evidence in court must produce sufficient proof that the item is what they claim it is. Examples of this include testimony from the witness who took a picture. But this flexible standard works for AI, too — it can include circumstantial details in the recording, metadata, device info, or expert testimony about how the file was maintained.
- Rules 902(13) and (14) govern the self-authentication of electronic evidence in some circumstances, such as certifications of electronic processes or data copied from a device. Crucially, while a hash value can show that a copy matches a source file, it doesn't automatically prove the source file itself depicts real-life events.
Drawing Lines on AI "Enhancement"
What to do about AI-generated enhancements is more than theoretical. In the 2024 murder trial State v. Puloka, a Washington state court tackled the gray area of AI-enhanced video. The defense tried to introduce cellphone footage that had been run through machine-learning software to “sharpen” blurry frames.
The judge excluded it. Under Rule 702 (which governs expert testimony), the court ruled that the AI’s method of predicting and inventing missing pixels wasn’t generally accepted in the forensic video community. The decision highlights a key distinction: while tweaking brightness or contrast is fine, using generative AI to fill in the blanks creates unproven predictions rather than authentic evidence. This makes it even more important to know when AI has been used in any capacity.
Recent Cases Are Charting a Course
So how do attorneys protect their authentic video and photographic evidence? As attorneys James Koukios and Tara McGrath recently highlighted in Law360, courts appear to be taking a moderate approach to deepfake challenges, and this blog draws on some of the same recent cases to explore what that means.
In January of this year, the Iowa Court of Appeals decided State v. Amyda. In that case, the Iowa Court of Appeals upheld the admission of a digital video in a sexual assault case because of the victim’s testimony. She recognized her body, clothing, room, and other important details in the clip. The court concluded that this testimony satisfied Rule 901 and that the defendant’s unsupported deepfake claim did not create a genuine question about authenticity.
Other cases are reinforcing this standard. In 2024, in a case called Mooney v. State, the Supreme Court of Maryland held that video evidence can be authenticated through a combination of witness testimony and circumstantial evidence, clarifying that a witness isn’t required to have personal knowledge of every single event depicted in a video to authenticate it.
However, when courts identify a real problem with the chain of custody or source files, they apply much closer scrutiny. In In the Matter of M.S., the New York Court of Appeals considered whether video clips should have been admitted when the original camera feed was unavailable, a significant part of the chain of custody was unexplained, and physical devices weren’t offered at trial.
The court held that the foundation was inadequate, noting that the prevalence of deepfakes makes it risky to rely solely on matching circumstantial details. Because fabricated media can easily incorporate real details from real people and places, proving that the source history and chain of custody are intact is becoming essential.
Advisory Committee on Evidence Rules Is Working on a Draft of Rule 901(c)
The Advisory Committee on Evidence Rules, which advises the Judicial Conference of the United States, has considered a working draft of Rule 901(c) of the Federal Rules of Evidence. This proposed rule directly targets evidence alleged to be created or altered by generative AI.
If enacted, Rule 901(c) would establish a clear two-step framework:
- Threshold Showing: The opponent would first need to offer evidence supporting a reasonable claim that the item was created or altered using AI.
- Burden Shift: Once that threshold is met, the burden shifts back to the proponent to establish that the item is more likely than not authentic.
It’s a rule that would codify the balance found in the case law above — but it’s not official yet. At its last meeting in May, the Committee recommended further study instead of approving 901(c) for publication.
When the Law Is Evolving, Attorneys Need to Be Cautious
The law is changing to keep up with technology. In the meantime, to avoid having digital evidence excluded as a potential deepfake, attorneys likely need to consider the issue throughout litigation:
- Start immediately: Document the who, what, where, why, and when of all digital files immediately upon receipt.
- Hold On to Native Files: Cases can no longer rely solely on compressed screenshots or photos. Retain original devices, preserve native media, and be ready to explain metadata and transfer history.
- Scrutinize the Source: When assessing opposing counsel's exhibits, look for genuine red flags — missing source files, suspicious timing, unexplained gaps in custody, or signs of compression and editing.
Ultimately, courts are still figuring out the best ways to distinguish genuine evidence from synthetic media. Litigators who focus on solid chain-of-custody practices and traditional foundational proof will be best equipped to protect their evidence in the courtroom.
