How Do You Prove a Human Made It?
The claim "a human made this" once needed no evidence, because there was no alternative. That era is over, and the infrastructure for proving it is being built right now in several incompatible directions at once. None of it is finished. Here is what actually exists.
Why Looking at It Doesn't Work Anymore
The instinctive answer is that you can simply tell. Increasingly you cannot, and more to the point, whether you can is a terrible foundation. Every year the answer degrades, and any system resting on visual inspection has a shelf life measured in months.
Detection tools carry the same flaw in a more expensive form: they are locked in a race where every improvement in generation erodes them. Build your trust infrastructure on detection and you are rebuilding it continuously while losing ground throughout.
So the question moves from does this image look human-made to what evidence exists about how it was made. That is a records problem, not a perception problem, and records problems are the kind you can actually solve.
Approach One: Content Credentials (C2PA)
The most established effort. C2PA is an open standard that attaches signed provenance metadata to a file: what device or software made it, what edits were applied, whether generative tools were involved. Some cameras now sign images at capture. Some editing software records its own steps. Generated images increasingly carry C2PA manifests declaring themselves as generated.
What it is good at: an auditable, industry-standard trail, backed by major players, spanning the whole pipeline rather than one app.
Where it is weak: metadata can be stripped, most platforms still discard it on upload, and it currently proves what tool touched a file rather than who moved their hand. A C2PA manifest reading "made in a drawing app" cannot distinguish a person from a script.
Approach Two: Capture-Time Proof
A newer consumer category: apps that make a verifiable record at the moment something happens. Sign w Selfie is an example, dual-camera capture with server-side verification, producing a record you can check later. Its use cases are mostly everyday: marketplace listings, insurance claims, events, handoffs, proving a message was real.
What it is good at: binding a specific person to a specific moment. For a physical object or a real-world event, that is genuinely hard to fake and easy to check.
Where it is weak for artwork: a selfie next to a finished drawing proves you stood near it, not that you drew it. It also drops a face into a verification flow, which plenty of artists will not want and which carries real regulatory weight. Biometric data is special-category under GDPR, and Illinois' BIPA has produced very large settlements for companies that collected face data without correct consent. Any implementation touching faces needs a lawyer before it needs a copywriter.
Approach Three: The Process Record
This is the underexplored one and, for drawing specifically, the most interesting.
A drawing app does not merely end up with an image. It holds the whole input stream: stroke order, timing, velocity, pressure, pauses, the redraw at the two-minute mark. That is a recording of a hand moving over time, and it is a far stranger thing to forge than a picture. To fake it convincingly you would have to synthesize plausible human motor behavior across the entire session, not just produce a good final image.
It also degrades gracefully in a way images do not. Even a partial stroke log is evidence, whereas a stripped metadata field is nothing at all.
Where it is weak: it works only inside an app that records it, it proves a human-shaped input rather than which human, and nobody has standardized it, so there is no shared way to check one against another.
What None of This Solves
Honest limits, because overselling verification is its own failure mode:
- Verification proves process, not value. A verified drawing can still be bad.
- Any system can be gamed by a determined person with the right setup. The goal is raising the cost, not achieving certainty.
- Adoption is the actual bottleneck. A standard nobody implements verifies nothing, and this is where most provenance efforts quietly die.
- Nobody has agreed on a standard, so the competing approaches mostly cannot even check each other.
Anyone telling you this is solved is selling something. It is early, and being straight about that is more useful than a confident story.
Where We Sit
We have not solved this either. What we have is a platform with no generative tools and no smoothing, where the only way to make something is to draw it.