Stock photos and likeness licenses are not the same thing

Walk into any creative agency and ask the legal lead: "are we covered if we use that stock face in a generated ad?" The answer used to be a confident yes. Today, the same lead pauses, asks for the prompt, asks whether the model was fine-tuned, asks where the reference came from. The pause is the story. Stock and likeness are different licenses for different things, and AI workflows have made the difference visible.

What a stock license actually grants

A standard stock license. Getty, Shutterstock, Adobe Stock, grants the right to use a specific image in specific media for a defined scope. Behind that license sits a model release: a contract between the photographer and the depicted person, signed at the time of the shoot. The release names the photographer, the agency, sometimes the brand, and authorizes commercial use of the resulting images.

Crucially, the model release covers the captured photograph. It enumerates use cases the parties anticipated when signing, print, web, packaging, broadcast. Pre-2023 releases rarely contemplated that the image might be fed into a generative model as a reference, fine-tuning input, or training sample. The clause for that simply was not standard.

What changes when AI enters the workflow

Modern campaign production frequently mixes licensed stock with AI generation. A creative team licenses a headshot, uses it as a reference, then produces a derived asset that is "based on but not identical to" the original. Some workflows even use stock images as training data for fine-tuned models that produce variations of that face for the campaign.

Each of those steps may exceed what the original release authorized. "Reference" is not "creating a derivative work in a generative system." "Training data" is almost certainly not in a 2019 release. The legal question is not just whether the original image is licensed, it is whether the generated output downstream is also licensed, and the chain of permissions usually does not extend that far.

What a likeness license is meant to cover

A likeness license licenses the person, not the image. The legal object is the right to depict an identifiable individual in a defined campaign, across channels, geographies, durations, and content types. The license is signed by the depicted person directly, with full disclosure about the campaign brief.

In a likeness license, the question "can we use this in AI-generated content?" is not an interpretation question. It is a clause. Either the license covers AI-output use or it does not. Either it covers training-data use or it does not. The contract anticipated the question and answered it.

Why this matters now and not ten years ago

For a decade, stock and likeness sat at different ends of a spectrum and rarely had to be reconciled. Stock served low-stakes commodity needs; likeness served high-stakes endorsement and editorial. AI workflows have collapsed the spectrum. A team can spin up a "looks like a real person" generated face in minutes, often using stock as input. The output looks like commodity stock. The legal exposure looks like unlicensed likeness.

The exposure is asymmetric. If you are right that the use was within the original license, you save legal fees. If you are wrong, you face a personality-rights claim, a takedown of the campaign, and damages calibrated to a license you never paid for.

The pragmatic position

For any campaign that uses a face, generated, real, or hybrid, the conservative path is to license the underlying likeness explicitly, with terms that name AI output and training as separately addressed dimensions. That can mean negotiating a fresh release with the original model. Or it can mean licensing through infrastructure that handles the question by default.