While much ink has been spilled on the troubling incidents of counsel using AI-hallucinated cases in their legal submissions, there is also reason to be concerned about the use of evidence created by generative artificial intelligence (“AI”).

Digital evidence like photos, videos, and voice recordings which are tendered as evidence of facts (for example, photographs of damage) are often admitted without rigorous scrutiny. If someone can testify that they took that picture or video, the triers of fact tend to believe what they see.

However, with the growing advancements in AI and the shockingly realistic nature of deepfakes, seeing is no longer believing, and there is a risk of parties submitting AI-generated evidence without disclosing it to the court.

What happened when AI-generated evidence reached a U.S. court?

Recently in the Superior Court of California (the “Court”) in Mendones v Cushman and Wakefield Inc [Mendones],  the plaintiffs’ case was dismissed for intentionally submitting false testimony to the Court. The plaintiffs submitted various exhibits of audio and video testimony, along with screenshots of messages, that the Court suspected of having been altered or created by AI. The use of AI was flagged because of the lack of expressiveness in the facial expressions and cadence in the speech of the individual depicted in the videos. As a result, the Court issued an order that the plaintiffs produce all the metadata associated with the evidence in issue. In addition, the Court ordered in-person appearances of the individual who took the images and videos, and the individuals identified in the images.

Upon review of the additional evidence ordered, and, specifically, the metadata in the images which were not consistent with the photos and seemed to be fabricated, the Court found that some of the evidence had been AI-generated. With respect to the balance of the evidence, the Court did not have the time, funding, or technical expertise to determine the authenticity of the evidence or to conduct a forensic analysis of the suspect evidence. However, the Court found that the plaintiffs had submitted at least two exhibits created by AI. The Court held that the use of deepfakes significantly undermines the Court’s ability to administer justice, significantly erodes the public’s confidence in the judicial system, and burdens under-resourced and overworked courts with the time-consuming task of assessing whether evidence presented to it was fake. As a result, the Court dismissed the plaintiffs’ action.

Why the absence of Canadian deepfake cases is not reassuring

While there appear to be no reported decisions yet in Canada of parties attempting to use a deepfake as evidence, there are an increasing number of attempts by parties to use generative AI content to bolster their claims in other ways.[1] It is likely only a matter of time before someone attempts to use deepfake evidence.

In addition, the Mendones case happened over a year ago and the quality of deepfakes has only improved, making detection from visual or sound differences even more difficult. The fact that there are no more recent cases where deepfakes were addressed by the courts is far from evidence that it has not occurred or is not occurring. It could be the opposite – that people are using more sophisticated technology, and it is going undetected.

Canada’s authentication framework: Built before deepfakes

The test for admissibility of digital evidence has not changed with digital advancements. Section 31.1. of the Canada Evidence Act, RSC, 1985, c C-5, provides for authentication of electronic documents as follows: any person seeking to admit an electronic document as evidence has the burden of proving its authenticity by evidence capable of supporting a finding that the electronic document is that which it is purported to be. This may not need to change in the modern AI world, but the burden of proving authentication may require more than traditional practice, and the reliability that evidence is what it purports to be must be more carefully scrutinized.

This was noted by the Ontario Courts in R. v. Medow, 2025 ONCJ 661 (CanLII). In that case, the accused asserted that the video evidence presented by the Crown was digitally altered and asked the Court to consider whether it was a deepfake. The court said the following around authentication and weight:

“I also agree that even if a court diligently evaluates digital evidence, a deepfake video could be “authenticated” by a witness who might honestly but mistakenly believe it to be an accurate depiction of the events it purports to represent. Alternatively, a witness could intentionally deceive a court by claiming a video was authentic when, in fact, they knew it was not. I agree that the existence of deepfakes presents a potentially serious concern to the integrity of our justice system. However, the admission of digital evidence does not mean that its ultimate reliability should be presumed. By contrast, a rigorous analysis of the ultimate reliability of any type of digital evidence is always required.”

Unfortunately, as was noted by the Court in Mendones, courts are not adequately prepared to discern whether a video or image was created using AI, and do not have the resources to fund AI detection tools. Therefore, the burden will fall to counsel to carefully scrutinize digital evidence produced by both the opposing party and their own clients so that it can assist the court with the “rigorous analysis of the ultimate reliability” of the digital evidence presented.

Digital evidence is routinely submitted to courts and authenticated by simply being attached to an affidavit, typically by the party who took the photo, attesting to its authenticity, or voice recordings being authenticated by someone who knows the speaker, attesting that it is their voice. But given the sophistication of deepfakes, something more is now required. The Court in Mendones initially flagged the issue due to the cadence of the voice recordings but confirmed it with metadata. While it is becoming more common for parties to request metadata as part of the document production process, it is not mandatory. Unless and until the courts or legislature amend the standard of what is required to prove the authenticity of digital evidence, counsel should routinely request the metadata for any digital production that there is any reason to doubt.

A practical checklist for challenging digital evidence

If the digital evidence seems too good to be true, and you suspect that it may have been generated or manipulated using AI, here are some tips* to look for and raise as issues in the authentication of any digital evidence or in submissions as to the reliability of that evidence:

  • Carefully scrutinize digital evidence for visual flaws, including: distorted body parts; scrambled text or logos; backgrounds that are repetitive or blur into one another; lighting and shadows that point in different directions or do not match the object that is casting the shadow; or how the lighting, contrast, color and sharpness of the background compares to the lighting, contrast, colour and sharpness of the object of the image to see if it appears as though the object has been switched into the photograph. For videos, consider the tone of voice, cadence, gestures, and facial expressions; whether the voice, accent and choice of words match the person being depicted; and whether the movement of the mouth matches the words being spoken.
  • Conduct a Google reverse image search by uploading the file to check if the image exists elsewhere online.
  • Request and check the metadata associated with the digital evidence to view the properties to see when it was created and GPS data of where it was created to see if it matches what it purports to be. Look at things like the file format, date created, date modified, identity of device capturing video, lens used, shutter speed, and file type.
  • Subpoena the individuals identified in the images/video. If there are parties other than those tendering the evidence in the images/video, seek their testimony as to whether those images/video depict what occurred.
  • Consider the use of AI detection tools. There are commercial tools available that can be used to detect both text and images that are AI-generated, although these come at a cost and it is unclear how courts may treat the results of these commercial platforms.

*Some of the ideas reflected in this checklist were developed with the assistance of generative AI and subsequently reviewed and refined by the author.

Conclusion

The traditional practices for authenticating digital evidence are not built for an era in which convincing fabrications can be generated in minutes. Courts have acknowledged the threat with Mendones dismissing a case over it, and Medow calling for rigorous reliability analysis, but the courts do not have the resources to catch up with the technology. Therefore, the burden of identifying and challenging suspect digital evidence falls, for now, on counsel.

Litigators who are not yet incorporating metadata requests, reverse image searches, and other AI detection techniques into their document review process are exposing their clients, and potentially their own professional obligations, to significant risk.

Miller Thomson’s Commercial Litigation lawyers advise clients on the evidentiary, procedural, and strategic challenges arising from AI-generated content in litigation. Whether you are dealing with suspect digital evidence, preparing to challenge the authenticity of an opposing party’s productions, or looking to strengthen the admissibility of your own digital evidence, our team can help.

[1] Yang v. Gibbs (dba D & G Cedar Fencing), 2024 BCCRT 613 (CanLII); Westcore Industries Ltd. v. Le, 2024 BCCRT 1272 (CanLII); LaRue v Kent Institution (Warden), 2026 BCSC 1313 (CanLII); and Iida v. Meadow Lane Equine Clinic Ltd, 2026 BCCRT 626 (CanLII)