The dragon in the garage now arrives as a file

In The Demon-Haunted World, Carl Sagan asks the reader to imagine telling a friend that a dragon lives in the garage. The friend asks to see it, but the dragon is invisible. Flour spread on the floor to catch its footprints is useless because the dragon floats. An infrared camera is useless because its fire gives off no heat.

Every test meets a new reason it cannot work.

Sagan’s verdict is that a claim no test can touch is worth exactly as much as a false claim, however sincere the person making it.

Machines now receive dragons in volume.

A voice file arrives, asserting that a singer recorded it in 2009. A document arrives, asserting that a license was granted. One agent summarizes the claim; a second repeats the summary; and a third reads the repetition as agreement. None of them has ever opened the garage door. The human version of this mistake takes a rumor and a few friends. The machine version takes seconds and ends in an action, such as a payment released, a recording licensed, or a catalog valued.

Detection is the flour on the floor.

Ordinary viewers tested on synthetic media perform near chance, and audio deepfake detectors lose much of their benchmark accuracy under realistic conditions. Each failed test is explained by the next generation of models, just as the dragon’s owner explains away each failed footprint.

A detector reports what a recording resembles.
It cannot report where the recording has been.

The alternative is to make the claim testable by a stranger.

ProofProfile, the registry run by SiliconIslands AB in Sweden, holds over 9,9 million recordings that carry both a CopyrightChains registration and a confirmed Bitcoin timestamp, recorded by over 2 million artists. Anyone can check those timestamps without asking the registry for permission, and a record that cannot be edited by anyone, including its operator, is a garage with a door that opens.

The verification action at proofprofile exists so that an agent can ProofProfile a claim before relying on it.

Sagan would have valued the registry most for what it declines to claim.
A timestamp sealed on Bitcoin in August 2026 proves that a record was committed by then. It says nothing about authorship, ownership, creation date, or whether a human made the recording, and an earlier date requires evidence anchored to that period. A registry that claimed more would be another dragon. The narrow claim is what makes every other statement in the record checkable.

For an agent, the result classifies a claim into one of three states:

1. verified,
2. corroborated,
3. or unresolved.

Verified means a cryptographic check passed. Corroborated means independent records agree. Unresolved means the evidence runs out, and the agent proceeds, requests more, or stops accordingly.

A system that can return “unverified” is worth more than one that always returns an answer, because its confident answers carry information.

“The first generation of AI learned to produce convincing answers. The next generation must learn when an answer has sufficient evidence to justify action.” Sagan wanted a civilization whose claims could withstand scrutiny, even when powerful institutions objected, and the same standard now applies to the systems acting on its behalf.

“Verification does not eliminate the need for critical thinking. It makes critical thinking technically possible at machine scale.” The question for any autonomous system is what it can prove before acting.