Proving your work is human-made: certificates of authorship in the AI era

"Human-made" is becoming a product attribute

For most of the history of selling handmade work, human origin was the default assumption, not a selling point. You did not label a painting "made by a person" for the same reason you did not label water "wet". That default is gone.

The clearest picture comes from the handmade marketplaces. Craft Industry Alliance documented the problem in mid-2024: in one search for a crochet pattern, three of the top eight results were AI-generated, recognizable by their too-sharp photos, impossible shadows, and the total absence of a finished-project picture. The same report notes the platform removed four times as many listings under its handmade policy in 2023 as the year before, and that work masquerading as handmade risks colliding with consumer protection law, not just buyer disappointment. Sellers of real work describe the practical effect simply: generated listings are produced in hours, priced below anything a human process can match, and they crowd the search results the real work used to live in.

Where a default collapses, a label market appears. BBC News reported in March 2026 that at least eight competing initiatives are racing to establish a globally recognized "human-made" mark, with the explicit ambition of becoming the Fair Trade logo of the AI era. The people behind them say the quiet part plainly: AI-generated volume is putting an economic premium on human-made work. The photography press has covered the same wave; one certification launch in 2025 gave away its first ten thousand certificates free to seed adoption.

So the commercial fact is settled: "human-made" is now an attribute buyers will look for, filter by, and eventually pay extra for, in exactly the way "organic" and "fair trade" became attributes. The open question is where the claim should live. And that is where most of the current answers go wrong.

What a badge on your website actually proves

The emerging human-made certification services come in two rough shapes. At one end are self-declared badges: you complete a survey, upload a sample of work, sign a declaration, and receive a mark to display on your site, sometimes in minutes. The coverage of these schemes is candid about the mechanism; the launch covered in the photography press acknowledges in its own standards that it relies on the applicant's honesty, and does not require the unedited originals that would let anyone check. At the other end are per-piece review services, which ask for creation evidence, progress shots, timelapses, working files, and have a person look at it before issuing a stamp for that one work. That is more rigorous, and closer to the right idea.

We are not going to sneer at any of this. The instinct behind these initiatives is correct: labeling matters, the premium is real, and a shared mark would help buyers. But there is a structural problem with the badge model that no amount of rigor fixes, and it is worth being precise about it.

A badge is a statement about a shop. A buyer is holding a piece.

A mark in your site footer says, at best, "this studio's practice was reviewed at some point and found to be human." It says nothing about the specific object that just arrived in the mail: this print, this pendant, this pot. It cannot, because it is not attached to it. Worse, a badge is an image, and images copy. The same PNG that certifies your shop can sit in the footer of a print-on-demand storefront running entirely on generated art by tonight, and the buyer comparing the two sites sees the same pixels. A mark that anyone can display is a mark that discriminates nothing, which is the exact failure that makes buyers skeptical in the first place.

The unit of trust is wrong. Buyers do not buy shops. They buy pieces, one at a time, and they resell pieces one at a time. Whatever carries the human-made claim has to travel with the piece.

Detection will not save you either

The other reflex answer is detection: run the image through a classifier and let software rule on whether it is generated. If detection worked reliably, this whole post would be unnecessary. It does not.

The most honest summary comes from the academic world, which has been burned by these tools at scale. The University of San Diego's legal research center maintains a guide on AI detection tools that summarizes multiple studies finding detectors "neither accurate nor reliable", producing high rates of both false positives and false negatives. One widely used checker claimed a false positive rate under one percent; later journalist testing put it far higher. The false positives are the part that should worry a working artist most: detectors flag real human work as generated, and there is no appeal process built into an algorithm. Building your livelihood on the hope that a classifier will vindicate you is building on sand that shifts with every model release.

A more serious infrastructure effort is Content Credentials, the C2PA standard backed by more than five hundred companies across cameras, creative software, and the AI labs themselves. It attaches tamper-evident metadata to digital files, recording how a file was created and what edits it went through. Where your toolchain supports it, adopt it; for photographers and digital artists it is genuinely useful. But be clear about what it proves. Content Credentials are explicitly designed to let good actors show their process, not to catch bad ones. The absence of credentials proves nothing, since most files have none. And the credentials record how a file was assembled, not whether hands or a model did the creative work inside the recorded steps.

There is also a boundary problem that matters enormously for makers of physical work: Content Credentials live on files. A cast pendant has no file. A thrown pot, a woven rug, a hand-bound book: no file. The provenance question for physical objects cannot be answered by file metadata, because the object is not a file.

So detection is unreliable, and cryptographic provenance is honest but partial. Neither answers the question your buyer is actually asking, which is not "does this image trip a classifier" but "did the person selling me this thing actually make it, and will they stand behind that?"

Authorship is a per-piece claim

Notice the shape of the question. "Did a person make this" is not a question about your brand, your website, or your general practice. It is a question about one object, asked at the moment of purchase and asked again, harder, at resale. Claims of that shape already have a home. It is the certificate of authenticity, which has always existed to make per-piece claims from a named maker: this piece exists, I made it, and this is what it is.

A COA in 2026 needs to carry authorship explicitly, not by implication. The base document is the same one we walked through field by field in what goes on a certificate of authenticity. Five of its fields do the authorship work:

• Medium and process. The field that states what hands did. This is where "human-made" stops being a slogan and becomes a description: what was carved, thrown, painted, set, stitched, and with what. The next section is entirely about wording this field.
• Creation date. An anchor in time. It places the piece in your catalogue and your life, and for older work it can place it before the generative era entirely, which is its own kind of evidence.
• Edition number. Authorship scoped to a unit. "I made this run of 25, and this is number 4" is a claim about a specific object, not a design. The conventions are in our edition numbering guide.
• The signed attestation. The named human act. A specific person or studio puts its name to the process claim, per piece. This is the field the badge model is missing: not "we are a human studio" in general, but "I made this one, and here is my name on it."
• A verification URL. The field that makes the other four checkable. A certificate that lives at a stable address on your domain can be pulled up by a skeptical buyer, a marketplace moderator, or, increasingly, an AI shopping agent comparing claims it can verify. A claim with a resolvable URL behind it is a different species of claim from a sentence in a listing, to machines and careful humans alike.

Put together, that is a certificate of authorship: a per-piece, named, dated, checkable statement of human making. Nothing about it is new technology. What is new is the economic reason to write it carefully.

Writing an attestation that means something

The word "handmade" is close to worthless now. It has been diluted by a decade of loose marketplace usage, and generated listings type it for free. What still works is process language, because process language is specific, checkable, and costly to fake convincingly.

The test for every sentence in your attestation: could it sit unchanged on a mass-produced or generated listing without being an outright lie? If yes, it is not doing any work. Rewrite it until copying it would require lying.

The right column follows a four-part formula, and you can write yours from it directly:

1. What your hands did, in verbs. Carved, threw, pulled, soldered, stitched, painted. Verbs are the hardest thing to fake because they commit you to a sequence of physical facts.
2. What tools were involved, by type. A wheel, a burin, a raster painting app, a film camera. You are describing a workshop, not endorsing products.
3. What AI did not do, stated flatly. One sentence: "no generative AI was used in the creation of this piece", or the scoped version from the next section if that is the truth.
4. What evidence exists. Progress photos, the layered file, a timelapse, the wax, the sketches. You do not publish it all. You state that it exists and can be produced.

That last point deserves a beat, because it is the quiet engine of the whole approach. Stating that evidence is retained makes your claim falsifiable. If you lied, the lie could be exposed by asking for the evidence, and you have named yourself and your business on the document. That is exactly why the honest version is credible: it accepts a risk that a fraud would not accept. A vague claim risks nothing and is therefore worth nothing. A specific attestation risks your name, and is therefore worth something.

Two cautions carried over from the general COA rules. Keep every sentence factual and first-hand; the attestation is not the place for poetry about your artistic journey. And date it, because an undated attestation invites the question of whether it was written for the piece or pasted onto it.

When AI touched part of the workflow

Here is the part most badge programs cannot handle and most makers quietly worry about: real studios in 2026 are rarely absolutist. You may have used AI to explore reference imagery, to mock up how a pendant sits on a collarbone, to rough out a color study, to draft the listing copy you rewrote. Does that disqualify the work from being human-made?

No, but it changes what you are allowed to claim, and the difference is where brands are won and lost. Two rules cover nearly every case:

Rule one: describe the finished object precisely. The authorship claim attaches to the piece: what it is made of and what process produced it. If no generated output is present in the finished work, say so in those words. "The piece itself contains no generative output" is a strong, precise, defensible claim even when AI appeared earlier in the pipeline.

Rule two: disclose where AI touched the creative pipeline. One sentence is enough. "AI-generated reference imagery was used during design exploration." "A generated mockup was used to test the composition; the final painting was made from my own studies." This feels like it costs you something. In practice it buys you more than it costs, for a reason that is easy to state: anyone can type "100% human-made, AI-free". It is free, so it signals nothing. A precise disclosure that admits something is expensive to fake, because a fraud gains nothing by volunteering nuance. Honesty reads as credible precisely because it is against interest. That is the moat, and it is one a generated-listing operation cannot cross by copying your words.

The inverse risk is just as concrete. An absolute "AI-free" claim on a workflow that used AI anywhere is a misdescription sitting in public, waiting. One leaked prompt history, one sharp-eyed customer, one platform audit, and the brand built on the claim detonates along with it. Consumer protection frameworks generally treat listing statements as binding claims, and the marketplace reporting above already frames undisclosed AI work as a consumer protection problem, not a taste problem. Overclaiming is not a marketing style. It is a liability with a fuse.

So draw the line where the truth is: the finished object, described exactly, plus a scoped disclosure of anything generative upstream. Makers who do this now will look prescient in three years, because disclosure requirements are only moving in one direction.

Running this on a real store

Everything above is tool-agnostic; you could run it with a word processor and a filing cabinet, minus the verification URL. Here is how it maps onto Editioned, since carrying per-piece claims is the thing we build.

Each numbered piece you sell gets its own hosted certificate page on a stable URL, with a QR code and a PDF. The certificate carries the piece's title, materials, edition number, and date, and its custom fields are where the authorship attestation lives: your process description, your AI disclosure line, your evidence-retained statement, worded exactly as you wrote it. The page is token-verified, so a certificate link either resolves to the real record or it does not; a screenshot of a certificate is not a certificate.

Two lifecycle properties matter specifically for authorship claims. First, revocation: if a piece is returned, or a claim ever needs to be withdrawn, the certificate URL starts returning a 410 Gone response, a deliberate tombstone that tells anyone checking that the record was revoked on purpose. Dead claims should not keep circulating, and the certificate lifecycle is designed so they cannot. Second, transfer: when the piece is resold, the certificate and its attestation move to the new owner instead of dying in the first buyer's inbox, which is exactly when a human-made claim gets its second, harder test.

The Free plan covers 3 products and 15 certificates, which is enough to run this on real orders before deciding anything. Paid plans lift the caps and remove the watermark; there are no per-certificate fees on any tier, so a certificate for every piece costs the same as a certificate for one. Details are on the pricing page.

One disclosure of our own, in the spirit of this post: Editioned is built by Guild 79, a jewellery atelier. We write these attestations for our own cast gold pieces, with our own name on them. This is not a workflow we designed in the abstract.

What a certificate cannot do, honestly

Let us close the loop we opened at the top, because the credibility of this whole approach depends on not overclaiming for it.

A certificate of authenticity is a signed claim, not forensic proof. No field on it detects anything. If a seller is willing to lie in a signed, dated, published document, the certificate will carry the lie as smoothly as it carries the truth. Anyone selling you a certificate as "proof" in the laboratory sense is overclaiming in exactly the way this post warns against.

What the certificate changes is the structure of the lie. Without one, a false "human-made" claim is an anonymous string in a listing: free to make, free to abandon, gone when the shop rebrands next month. With one, the same lie requires a named business to sign a specific, falsifiable statement about a specific object, publish it at an address it controls, and leave it standing where any future buyer, journalist, or platform can check it, for years. Fraud does not survive contact with that structure very long, not because the document catches it, but because accountability does.

That is the honest comparison with the alternatives. Detection promises certainty and delivers coin flips. Badges promise trust and deliver a copyable image. A certificate promises only this: a name, a record, and consequences, per piece. Markets have run on that mechanism for centuries, under other names: the hallmark, the signature, the maker's mark. The certificate of authorship is the same mechanism, given a URL. For the wider system it fits into, from paper to hosted records, see the complete guide to provenance certificates.

The short version

"Human-made" is becoming a sellable attribute exactly as generated work floods every marketplace, and a label market is forming around it. But a site-wide badge proves nothing about a specific piece, detection tools misfire in both directions, and file-level provenance cannot follow a physical object. Authorship is a per-piece claim, so carry it on the per-piece instrument: a certificate of authenticity with specific process language, a creation date, an edition number, a signed attestation that states what AI did not do, a scoped disclosure of what it did, and a verification URL that resolves or does not. Not forensic proof. A name, a record, and consequences, which is what trust has always been made of.

FAQ

Can a certificate of authenticity prove my art is not AI-generated?

Not forensically. No document can. What a certificate does is convert an anonymous ambiguity into a named, per-piece, checkable claim: a specific business states how a specific piece was made, signs it, dates it, and publishes it at a URL that can be verified or revoked. Its strength is accountability, a maker staking their reputation on one object, which is more than a site-wide badge or an unreliable detector can offer a buyer.

Should I mention AI tools I used for mockups or reference images?

Yes. Describe the finished object precisely, then disclose where AI touched the creative pipeline: reference imagery, mood boards, mockups, listing copy. A line like "the piece itself contains no generative output; AI-generated reference imagery was used during design exploration" costs you almost nothing and reads as credible precisely because it admits something. Blanket "AI-free" claims on a workflow that used AI anywhere are a misdescription waiting to surface.

What is the difference between a human-made badge and a certificate of authenticity?

The unit. A badge is a statement about a shop or an artist as a whole, usually self-declared, and the image can be copied onto any site in seconds. A certificate of authenticity is a statement about one specific piece: its process, date, edition number, and a signed attestation, ideally at a verification URL unique to that piece. Buyers do not buy shops, they buy pieces, so the per-piece claim is the one that matters at the moment of purchase and again at resale.

Do Content Credentials (C2PA) prove a work is human-made?

No, and they do not claim to. Content Credentials attach tamper-evident metadata to digital files recording how content was created and edited. They are built to let good actors show their process, not to catch bad ones, and the absence of credentials proves nothing. They also only exist on files: a cast pendant or a thrown pot has no file to sign. For digital artists and photographers they are worth adopting where the toolchain supports them, as a complement to a certificate, not a replacement.