AI content authenticity

State of the art (2026)

Two methods are now used together. C2PA Content Credentials are signed metadata that record where a file came from. Watermarks such as Google's SynthID are embedded in the content itself. On 31 July 2026 OpenAI extended SynthID to the audio it supports, added audio checks to its public verification tool and announced API access for verification. A detector can only find the signals it supports, so a file with no signal may still be AI-made. What to watch: how many generators take part, whether the signals survive editing and re-uploading, error rates, and whether tools from different companies read each other's signals.

Core thesis

As AI-generated images, audio and video become harder to tell from the real thing, a reliable record of where content came from grows more valuable. C2PA Content Credentials bind a signed record of origin and edits to a file, and show if that record has been tampered with. They say where content came from, and C2PA itself says that does not make the content true or accurate. Credentials can be stripped from a file. C2PA's durable-credential methods can help recover them, but only where the tools involved take part. Watermarks can also be attacked, and how well each one holds up depends on the model and the attack.