Technology
The Evidentiary Admissibility of AI-Generated Evidence
Quick fact
In the United States, there is no specific law governing AI evidence, but courts apply the same rules that were written long before AI existed—and judges have already ruled in several cases that deepfakes cannot be admitted without rigorous authentication, even leading to sanctions on parties who submitted them.
Why this is interesting
You've seen deepfakes that look impossibly real. But if a deepfake video ends up as evidence in a court case, can the court even use it?
Read the full explanation
Understanding The Evidentiary Admissibility of AI-Generated Evidence
Imagine you are trying to prove someone was at a crime scene, and you have a video that appears to show them there. In the past, you would just show the video to the judge. But now, a video could be completely fabricated by AI, yet look perfectly real. The law has old rules that all evidence must be authentic, meaning it really is what it claims to be. How do you authenticate something when it could be a perfect fake? This is the core challenge. The process is like a detective checking the chain of custody: you need to show the video originates from a reliable source, was not altered, and is a true representation. For AI-generated evidence, this might involve examining the metadata, questioning the software used, and bringing in experts who can testify that the video has no signs of manipulation.
A deeper explanation
Admissibility of AI evidence rests on the same pillars as all evidence: relevance, authentication, and reliability. Under rules like the U.S. Federal Rules of Evidence, a party must prove that evidence is authentic, often through a witness who can testify it is what it is claimed to be. For AI-generated content, this becomes tricky: the algorithm's inner workings may be opaque (the 'black box' problem), and there may be no human witness with direct knowledge of the content's creation. Courts are grappling with how to apply the 'reliability' factor from cases like Daubert v. Merrell Dow Pharmaceuticals, which requires that expert testimony be based on valid, reliable methods. Some jurisdictions are leaning toward requiring a thorough explanation of the AI system, its accuracy rates, and its potential errors. Another issue is the hearsay rule: if the AI output is an assertion by an out-of-court 'declarant' (the AI itself), it might be hearsay, but if it's simply a machine calculation—like a GPS readout—it might be treated as a real-world observation, not hearsay. Courts are currently split. The trend is toward a cautious approach: a party must lay a solid foundation explaining how the evidence was generated and why it is trustworthy, or it will be excluded. This matters because if we admit fake evidence, we risk wrongful convictions; but if we exclude all AI evidence, we may lose access to valuable new tools like facial recognition, predictive policing, or digital reconstructions.