AI image detection vs. cryptographic provenance

The "is this AI-generated?" question gets answered in two completely different ways. Detection tools — a growing ecosystem of commercial classifiers and watermark readers — run a model over the pixels and return a probability. Provenance tools — Orphograph, the C2PA standard, the various OpenTimestamps frontends — bind a file to a moment in time and let anyone verify the binding later. The two are often conflated in coverage, but they solve different problems and have different failure modes.

Detection: probabilistic, and getting worse

An AI image detector is a machine learning model trained on examples of AI-generated images and examples of real photos. It learns statistical patterns — micro-textures, frequency-domain artifacts, consistency of lighting — that historically distinguished generations from captures. The output is a probability: 87% likely AI-generated, 23% likely AI-generated, and so on.

Three structural problems with this approach:

• The arms race favors generators. Each new image model is trained partly to defeat known detectors. Studies have measured detector accuracy collapsing toward random as new generations of models release.
• False positives on legitimate work. Heavy post-processing, aggressive JPEG compression, and certain camera pipelines produce artifacts that detectors interpret as AI fingerprints. Real photographers get flagged.
• You can't appeal to a detector. If a platform's detector says your photo is AI-generated and de-monetizes it, you have no cryptographic argument to make. The detector is a black box.

Provenance: deterministic, and durable

A Bitcoin-anchored provenance proof works completely differently. You hash the file. The hash goes into a Merkle tree. The Merkle root goes into a Bitcoin transaction. The block has a timestamp. To verify, anyone re-hashes the file, walks the Merkle path, and looks up the Bitcoin block. The answer is yes-or-no, not a probability. Either the hash matches the chain or it doesn't.

The decisive property for AI disputes is the block timestamp. Every major image model has a published training-data cutoff. If your anchor is dated before that cutoff, the model literally could not have generated your file — it didn't exist when the model trained. No detector argument is needed. The math is the argument.

Where detection is still useful

Detection is the only tool that applies to new images you encounter in the wild, where no provenance proof exists. If you're trying to decide whether a viral tweet's image is AI-generated, you can't ask the poster for a Bitcoin anchor — you can only run a detector and weigh the probability. Detection is necessary in that context. It just isn't the right tool for protecting your own work going forward.

Where provenance is decisive

For your work — photos you take, illustrations you draw, manuscripts you write, code you commit — provenance is the strictly stronger approach. Anchor at the moment of creation or delivery, and the AI-generation accusation becomes refutable with a thirty-second cryptographic check rather than a probability score someone can dispute.

What we don't claim

Provenance does not prove authorship. It proves this file's bytes existed by this time. To prove "I made it," pair the timestamp with the rest of your normal evidence — RAW files on your own storage, drafts on your own devices, dated correspondence with clients or editors. The anchor strengthens that evidence package; it doesn't replace it.

FAQ