When AI Becomes Evidence in Court

Imagine a recording in which a director appears to approve an unlawful payment. The voice sounds familiar. The words are clear. The file arrives without the device that recorded it, the full conversation or a reliable account of how it was obtained.

The first question is whether the recording represents a real event. The second is whether the words mean what the party offering them says they mean. Artificial intelligence can complicate both questions, even when nobody deliberately fabricates evidence.

Separate the source from the processing

AI can generate an entire image or recording. It can also transcribe real speech, translate a message, enhance poor audio or classify a face in existing footage. These uses create different evidential problems.

A transcription error may change a name or omit a negative. Audio enhancement may produce an apparent word that the original cannot support. A face matching system may generate a candidate for investigation without establishing identity. A court needs to know which operation took place before it can assess the result.

Under Rule 901 of the US Federal Rules of Evidence, authentication requires a sufficient basis for finding that the item is what its proponent claims. Evidence about a process or system can form part of that foundation. The rule offers a useful starting point for asking how a digital item was created and why it should be trusted.

Preserve the material needed to test the claim

The original file, its metadata and the collection history should be preserved where available. If a party uses an AI tool, the record should also identify the tool, version, settings, input and output relevant to the disputed result. A later rerun may behave differently, particularly after a service has changed its model.

For generative systems, instructions and repeated attempts can be especially important. The final output may have been selected from several inconsistent answers. Presenting only that output conceals a fact that could matter to its weight.

Digital fingerprints can help show that a file has remained unchanged after a particular point. They cannot establish that the event depicted actually occurred. A perfectly preserved fabrication remains a fabrication. The integrity of storage and the truth of the content require separate analysis.

Rule 902 provides certification routes for certain electronic records and copied data in US federal proceedings, subject to its conditions. Such routes address authentication. Other objections, including hearsay, relevance and the reliability of an expert conclusion, may remain.

A score needs an explanation

An assertion that a system is highly accurate tells a court little unless the measure is explained. Accuracy across a general test set may conceal a much weaker performance on short recordings, compressed footage or people absent from the training data. The relevant issue is performance under conditions resembling the disputed evidence.

Expert analysis should identify the method, the validation material and the limits of the conclusion. In US federal proceedings, Rule 702 requires the proponent to establish the conditions for expert testimony, including reliable methods and their reliable application. A product's commercial reputation does not satisfy that inquiry by itself.

The opposing party needs sufficient access to challenge the actual proposition advanced. A claim that every part of a system is commercially confidential can leave a decisive inference effectively untested. Courts may need to consider protected disclosure, independent testing, limits on the evidence or other measures permitted by their procedural law.

Avoid the reverse mistake

The existence of deepfakes also makes it easier to dismiss genuine evidence. A bare suggestion that anything can be fabricated should not displace a properly established chain of evidence. Objections should identify a concrete weakness, such as missing source material, an unexplained edit or an unsupported attribution.

Courts already distinguish between admission and weight. AI makes that distinction more consequential. A digital item may cross the threshold for admission while still supporting only a narrow inference. The judgment should reflect the proposition actually proved, rather than the apparent confidence of the technology that produced it.