A recording, a video, a photo: in criminal trials they are often the key evidence, yet artificial intelligence makes them easier than ever to fabricate. Synthetic content can imitate a voice, reproduce a face and make someone say words never spoken, and a plausible file alone does not prove who is speaking or what really happened.
The problem has already surfaced in several contexts. It is known as the Liar's Dividend: when fakes circulate widely, even authentic evidence can be dismissed as false by those who have an interest in doing so. The risk is therefore not only accepting a fake, but losing trust in what is real.
For criminal lawyers the issue is practical. In court it is not enough to say that content "looks real": one must establish who created it, who spread it and whether reliable technical verification exists. Digital forensics points to source certification, meaning traceability from origin, and to chain of custody, which documents every step of the evidence. Multimedia forensic analysis is also becoming central, with the defence able to challenge an AI-generated video or audio by requesting the original source.
The regulatory framework is moving. Since 2 August the AI Act requires labels and machine-readable markers for deepfakes and synthetic content, but experts warn that text watermarking is fragile and that the absence of a label does not prove a file's human origin. The risk of AI use in court filings is real too: a New Mexico court fined a lawyer for inventing witnesses with ChatGPT.
The topic also cuts across justice systems beyond Italy. With a new step in Japan, the Supreme Court will test generative AI in civil proceedings from 2027. The open question is which evidence can still hold when every image and voice can be built at almost no cost.
No comments yet. Be the first!