AI Detection vs. Provenance Verification
AI detection and provenance verification are two different approaches to answering "was this AI-generated?" — and they give very different guarantees. AI detection uses statistical or pattern-based models to guess whether unlabeled content was likely AI-generated, based on artifacts in the pixels, audio, or text. Provenance verification instead checks a cryptographic signature attached to content at the moment it was created, confirming a specific, signed claim rather than making a probabilistic guess.
How AI detection works — and where it breaks down
AI detectors are typically machine-learning classifiers trained to spot statistical fingerprints left by generative models — unusual frequency patterns, inconsistent lighting, or telltale phrasing. Because generative models keep improving and detectors are trained on past examples, this is an inherent arms race: detectors can be fooled by newer models, produce false positives on genuine human content, and offer no way to verify their own accuracy on any single input. No detector can reliably identify all AI-generated content, and treating a detector's guess as certainty is a known failure mode in this space.
How provenance verification works instead
Provenance verification doesn't guess — it checks whether a cryptographic signature, created by a specific generator at the moment of generation, is valid for a specific piece of content. If a valid signed record exists (via a token, embedded metadata, or a recovered watermark), verification confirms who signed it and when, with mathematical certainty rather than probability. The tradeoff is coverage: verification can only speak to content that was signed in the first place. Content with no signed provenance simply returns "no provenance found" — that absence does not mean the content is human-made, only that no verifiable claim exists for it.
Which one Certivu does
Certivu performs provenance verification, not AI detection. It does not scan content for AI-generation artifacts or issue a probabilistic guess about origin; it checks for a valid ML-DSA signature, a resilient watermark, or a fingerprint match tied to a record created at signing time. That is a deliberate design choice: cryptographic verification is falsifiable and precise where detection is inherently probabilistic. Certivu does not claim to detect all AI-generated content — it verifies signed provenance when present.
FAQ
Can Certivu detect if an image was made by an AI I've never heard of?
No. Certivu doesn't do general AI detection — it verifies signed provenance for content that was signed through Certivu, or a compatible C2PA-based signer, at creation time. Unsigned content, regardless of its origin, returns no provenance rather than a detection guess.
Is a 'low confidence' or 'no provenance found' result the same as saying content is human-made?
No. It means no verifiable signed claim could be found for that content — nothing more. Absence of provenance data never implies human origin; it may reflect content generated by a system that never signed it, provenance metadata that was stripped, or content that was simply never covered by a provenance system at all.
Why doesn't provenance verification need to keep up with new generative models?
Because it doesn't analyze content for AI fingerprints — it checks a signature made by the generator itself. A new image or video model doesn't change how verification works; it only matters whether that model's operator chooses to sign its outputs.