Output and training are two different licenses

Most likeness contracts written before 2024 collapsed two questions into one. Can we use this person in advertising? Yes or no. The contract did not separately ask: can we use this person's likeness to train a generative model? In a pre-AI world, the second question rarely arose. In a generative world, it is constantly relevant, and conflating it with the first is the most common contract failure we see.

What "output" actually means

Output use is the visible, intended placement. The licensee includes the likeness in a finished asset, a banner ad, a video spot, an OOH placement. The asset is published, distributed, archived. The license covers that placement: where, when, on which channels, in which industries, for how long.

Output use, in AI workflows, is what a generator produces and what a brand publishes. If a model is prompted to generate "a woman wearing the brand product" and the output happens to resemble a real licensed person, that is output. The license covers it if the contract terms cover it, and not otherwise.

What "training" actually means

Training use is invisible to the published asset. The licensee includes the likeness in a dataset that fine-tunes or pre-trains a model. Once trained, the model has internalized statistical features of the person, their face shape, expression patterns, the way light falls on their skin. The model can then produce outputs that resemble the person, even when not specifically prompted to.

Training use is open-ended. A trained model is reused, redistributed, and may live in production for years. A face used for training appears in an unbounded number of future outputs, none of which are individually authorized at the time of training.

Why these are different licenses

Three reasons. First: scope. Output is bounded by the campaign brief; training is bounded only by the model lifecycle. The risk profile and the appropriate price are different by orders of magnitude.

Second: revocability. Output use can usually be retracted, pull the campaign, recall the assets. Training use is technically very hard to undo. Once weights have absorbed a face, removing that influence requires either retraining from scratch or a research technique called "machine unlearning" that does not yet work reliably at production scale.

Third: regulatory exposure. The EU AI Act, GDPR Article 9 on biometric data, and proposed national deepfake legislation all apply more strictly to training datasets than to individual outputs. A training-data inclusion creates compliance obligations that an individual output does not.

The contract structure that works

A modern likeness contract treats output and training as separate switches. Output: defined scope, with channels, geography, duration, exclusions, brand-size factor. Training: a separate decision, included or not, with its own price, its own term, its own revocation procedure. The default in any well-designed contract should be that training is OFF unless explicitly enabled by the depicted person, per request.

On FaceLedger, that is how it works in practice. Standard licenses cover output only. Training is a distinct license dimension that the talent decides on individually, per campaign request, with separate compensation. No one becomes a training dataset by surprise.

What to ask before you license

Before signing, before paying, before going to production: confirm in writing whether the license covers training, and if so, on what terms. If the contract is silent, assume training is not authorized. The reverse assumption is expensive.