A team of researchers has proposed a formula for measuring exactly how much of an AI-assisted piece of writing, art, or code actually came from the human at the keyboard.
The framework, described in a paper on arXiv, borrows a concept from information theory called mutual information. The researchers calculate how much shared information exists between a person's input (a prompt, an edit, a sketch) and the AI-assisted output, then compare that against the output's total self-information. The ratio gives a numeric score for the human's proportional contribution rather than a yes-or-no judgment. In tests across several creative domains, the researchers say the measure reliably told apart works with heavy human involvement from ones that were mostly AI-generated with light prompting.
This matters because originality has become a genuinely hard problem to define. Copyright offices and courts have spent the last few years trying to figure out how much human input is enough to make an AI-assisted work protectable, and right now that judgment is mostly a case-by-case guess. A quantitative score would not settle the legal question, but it would give publishers, platforms, and regulators a consistent way to talk about degrees of human authorship instead of arguing in the abstract.
Still, a formula is not a law. Turning a mutual-information score into an actual originality threshold - and getting courts, platforms, or copyright offices to agree on where that line sits - is a much messier fight than the math behind it.