A Canadian research team building a tool to help courts identify AI-generated or altered evidence has found that its algorithms could not reach a useful consensus in two real-world tests. The difficulty points to a central problem for legal evidence: files can change substantially before they reach court.
Watch Desk analysis
What happened
The team’s system combines several algorithms that vote on whether material was made or altered by AI. Maura Grossman, co-director of the project, told The Canadian Press that the system did not reach a helpful consensus in the two real-life situations where she tried it.
The open-source project is approaching the halfway point of its two-year term. Grossman says images can be compressed and stripped of metadata as they move between platforms and people, leaving a court with something quite different from the original file. The next phase will seek input from law enforcement and policymakers.
Why it matters
A detector that looks promising in a lab may be much less decisive when evidence has been shared, compressed or lost in the long journey to a courtroom. That is a serious constraint when the question is not just whether content looks synthetic, but what can be shown about its origin and integrity.
Grossman says the project was never meant to be a silver bullet: human judgement would still be needed, and the team is considering broader legal and policy responses. The report does not establish how the tool performs across a wider range of cases; the two real-world attempts described so far were inconclusive.
Our read
The useful result here is not a courtroom-ready AI lie detector. It is a clear account of why the neat lab version of that idea can fall apart when evidence has been passed around in the real world. That is worth knowing before anyone treats a detector’s verdict as the final word.
What to watch
- Whether the next phase brings law enforcement and policymakers into the project.
- How the team tests the system on evidence altered by compression, sharing or missing metadata.
- Whether any proposed tool makes its reasoning clear enough for people to challenge it.
Discussion spark: If an AI detector cannot reliably agree on evidence after it has been shared or compressed, should courts use such tools at all, or put the emphasis on preserving original files?
Sources and evidence
- Source update (30 October 2026, 19:30 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.