Signing What You Didn't Fully Make

An inherited figure

For two centuries, the signature at the bottom of a work has pointed to one precise subject: an individual, accountable, identifiable, whose hand made what the eye is looking at. That figure, the Romantic author, solitary and coherent, is largely a construction of nineteenth-century Europe. It shaped modern copyright, the rules of academic attribution, the very idea of intellectual property, and even the way platforms organize cultural distribution. The whole legal apparatus governing AI in the cultural industries rests, implicitly, on that figure. Take it away and there's no rights holder, no assignment, no contract. The signature isn't an ornament. It's the keystone.

The keystone was fragile long before AI. Film, advertising, publishing and architecture have run on collective authorship for decades: directors backed by hundreds of technicians, art directors supported by teams, sound engineers and editors whose names rarely make the poster. The individual signatory is often a convenient fiction that covers the collective nature of the work. Generative AI didn't invent the dispersal of the creative subject. It pushes the logic to the point where the mask stops holding.

Who holds the pen?

Several agents have a hand in a generated image. There's the prompter, who wrote the request, sharpened the intention, sorted through the outputs. There's the model, the statistical product of a training run whose parameters come out of successive optimizations. There's the alignment and fine-tuning team, who decided what the model should and shouldn't produce, which biases it would own and what it would censor. And there are the millions of anonymous authors whose works fed the training sets, many of them without explicit consent. If the signature is going to hold, it has to account for at least part of that cascade.

In practice, the person who signs an AI-assisted image is the one who sells it, shows it or publishes it. Almost always the prompter. It's a useful convention: it lets the market function and contracts exist. It's also a fragile one. It assumes the prompter's contribution is decisive enough for authorship to belong to them alone. Sometimes the assumption holds, with heavy manual retouching, complex compositing, art direction worked through several rounds. Sometimes it doesn't. A four-word prompt followed by a quick pick is not a work in the traditional sense.

Transparency as part of authorship

Since the single authorial subject has become untenable in theory, the signature needs something added to it: disclosure. Disclosing the use of AI, stating the prompter's role, naming the tools, laying out the stages of the process doesn't weaken the signature. It returns it to its real function, which was never to hide collaboration but to assign moral and legal responsibility to someone identifiable.

Several cultural platforms have started asking for this kind of disclosure. Spotify has adopted the DDEX industry standard so that labels and distributors can declare AI involvement in a track's credits. Photography competitions have created separate categories, or require transparency about the tools used. Publishers are beginning to add similar clauses. Slow and piecemeal as they are, these changes sketch a new economy of the signature: no longer the solitary gesture of a single subject, but an explicit acknowledgment of the production chain. CNIL, France's data protection authority, is also pushing for transparency standards along the same lines.

What it changes for a studio

For a studio that brings AI into its work, this reshuffle has immediate practical consequences. The signature is no longer just a name in the bottom right corner of a delivered visual. It comes with information, in the production notes or the metadata, that places the human contribution within the full chain. That documentation isn't a legal luxury. It protects the studio, it protects the client, and it builds the trust a working relationship rests on.

It takes work. It means keeping, for each project, a record of prompts, choices, retouches, successive versions. It also means knowing where the signable zone ends: the point past which human work stops being decisive enough for authorship to be claimed honestly. WIPO, through its copyright program, has published several studies that are useful to anyone who wants to bring their practice in line with the international guidelines now taking shape.

Signing with less arrogance, perhaps

What changes, in the end, isn't the possibility of signing. That remains. It's the stance of the person who signs. The Romantic author signed with a certain assurance: he had, as he saw it, done everything. The contemporary author working with a generative model signs knowing they haven't done everything, that a substantial share of what they deliver comes from elsewhere, from others, from processes whose details they'll never see. Held honestly, that knowledge tempers the signature without canceling it. It may make it more modest. Possibly more accurate.

The authorial subject hasn't disappeared. It has lost its illusion of autonomy. That isn't necessarily bad news for art: the works worth caring about are often the ones whose authors know how much they owe to others. AI, by making any other stance impossible, forces that acknowledgment. It remains to be seen which creators will make a virtue of it, and which will use it to hide more comfortably.