Trust needs a receipt

Most AI transparency currently looks like a label.

“Made with AI.”

“AI-assisted.”

“Always check important information.”

Those disclosures tell us that AI was involved. They tell us very little about what it did.

A new open specification called TRACE is proposing something more demanding.

Instead of relying on a vendor's description, TRACE creates a signed record of the system that ran, the policy governing it, the kind of data it touched and the tools it called.

The record can then be checked independently.

TRACE is highly technical, still in developer preview and aimed primarily at agents handling sensitive data. It also cannot tell you whether an agent produced a good answer. It proves aspects of the process, not the quality of the result.

But the underlying principle travels well.

If an AI system produces an analysis, recommendation or action, useful transparency should help us inspect how it happened.

Which data was available?

Which rules were applied?

Which tools were used?

What was excluded?

What can the evidence actually prove?

A disclosure asks us to trust the system's description of itself.

A record gives us something to inspect.

Trust in AI may eventually need to work less like a label and more like a receipt.