Deepfakes and Litigation: Authentication, Detection, and the Deepfake Defense

Synthetic media technology has advanced to the point where video, audio, and images depicting events that never occurred can be produced with consumer-grade hardware and freely available software. This development has created a category of evidentiary challenge that courts, practitioners, and legislators are only beginning to address: the deepfake problem in litigation.

What Is a Deepfake?

The term "deepfake" originally referred to a specific technique using deep learning models to swap one person's face onto another person's body in video. The term has since expanded to encompass a broader range of AI-generated synthetic media, including face swaps, voice clones, full-body synthesis, and text-to-video generation. What these techniques share is that they use machine learning models trained on large datasets of real media to generate outputs that are designed to be indistinguishable from authentic recordings.

The technical sophistication of these tools varies considerably. At one end of the spectrum are consumer applications that can produce convincing face swaps in minutes. At the other end are professional-grade synthesis systems capable of producing synthetic video that is difficult to distinguish from authentic footage even under careful examination. The evidentiary significance of a given piece of synthetic media depends in part on where it falls on this spectrum.

The Current State of Deepfake Detection

Deepfake detection is an active area of technical research, and the state of the art changes rapidly. Detection methods fall into several broad categories. Physiological signal analysis examines whether the video contains consistent biological signals such as pulse, blinking patterns, and micro-expressions that are characteristic of real human subjects. Artifact analysis looks for technical inconsistencies introduced by the synthesis process, such as unnatural blending at face boundaries, inconsistent lighting, or temporal artifacts between frames. Provenance analysis examines the metadata and file structure of the media for signs of manipulation or re-encoding.

Each of these methods has limitations. Physiological signal analysis is less reliable for low-resolution or compressed video. Artifact analysis depends on the specific synthesis method used, and artifacts that are characteristic of one generation of tools may not appear in outputs from newer systems. Provenance analysis can be defeated by careful re-encoding or metadata stripping.

An important limitation of all current detection methods is that they were developed and validated on datasets of known synthetic media. A detection model trained on outputs from specific synthesis tools may not generalize reliably to outputs from different tools or to novel synthesis techniques. This means that a negative detection result does not definitively establish that media is authentic, and a positive detection result does not definitively establish that it is synthetic.

The Deepfake Defense in Criminal Cases

The deepfake defense is the argument, raised by criminal defendants, that video or audio evidence purportedly depicting the defendant is AI-generated or AI-manipulated. This defense has been raised in cases involving surveillance footage, recorded conversations, and social media content.

Courts have generally been skeptical of deepfake defenses that are not supported by specific technical analysis. In several cases, courts have excluded or given little weight to expert testimony that merely described the general capabilities of deepfake technology without analyzing the specific evidence at issue. The emerging standard appears to require that a defendant raising a deepfake defense present expert testimony that specifically analyzes the challenged evidence and identifies concrete technical indicators of manipulation.

For prosecutors, the deepfake defense has created a new burden: affirmatively establishing the authenticity of video and audio evidence that would previously have been admitted without significant challenge. This requires technical analysis that goes beyond the traditional chain of custody documentation and addresses the specific technical characteristics of the evidence.

Authentication Standards for Video and Audio Evidence

Under Federal Rule of Evidence 901(b)(9), evidence may be authenticated by evidence describing a process or system and showing that it produces an accurate result. For video and audio evidence, this has traditionally been satisfied by testimony about the recording equipment and conditions. The deepfake problem complicates this standard because the question is no longer only whether the recording equipment was functioning correctly but whether the recorded content has been subsequently manipulated.

Several courts have begun to require more robust authentication for video and audio evidence in cases where manipulation is alleged or plausible. This includes documentation of the chain of custody from the recording device through to the evidence presented at trial, technical analysis of the file metadata and encoding history, and in some cases expert testimony about the technical characteristics of the evidence.

Civil Litigation and Synthetic Media

Deepfake issues arise in civil litigation in several contexts beyond the criminal deepfake defense. In defamation and harassment cases, plaintiffs may seek to introduce synthetic media as evidence of the harm caused, or defendants may challenge the authenticity of media offered by plaintiffs. In intellectual property cases, the question of whether a video or image was AI-generated may be relevant to copyright ownership and infringement analysis. In employment and discrimination cases, synthetic media may be used to fabricate or challenge evidence of workplace conduct.

In each of these contexts, the evidentiary framework is the same: authentication is required, and when authenticity is genuinely disputed, expert testimony is typically necessary to address the technical issues.

Practical Guidance for Attorneys

Attorneys handling cases where synthetic media may be at issue should take several practical steps. First, preserve the original media files in their native format, including all metadata. Re-encoding or format conversion can destroy the technical information that detection analysis depends on. Second, obtain the complete chain of custody documentation for any video or audio evidence, including documentation of every system through which the files passed from capture to production.

Third, retain a qualified technical expert early. Detection analysis is most reliable when conducted on original, uncompressed files with complete metadata. An expert retained late in the case may be working with degraded evidence that limits the reliability of any conclusions. Fourth, be prepared to address the limitations of detection analysis in your expert's testimony. Courts and opposing counsel will probe the limits of what the analysis can and cannot establish, and an expert who can clearly articulate those limits is more credible than one who overstates the certainty of their conclusions.