What Does It Mean To Own A Face, And Who Decides Who Gets To?

Data ownership is a phrase the AI industry uses constantly, but rarely actually defines. For most kinds of data, that doesn’t matter much – either because the stakes are relatively low, or because existing industries, or industry bodies, have been able to use the rights-assigning tools already available to them as a wedge to assert ownership in the new era. 

A spreadsheet of customer transactions, or a collection of news articles. A library of stock photos or a set of case law. There are owners, licenses, and receipts for all these things, and where the traditional model clashes with what’s happening on the ground (as it is in publishing) a new frontier is going to be tested and found in court. 

But there’s an asset class nobody has designed the new framework for, and that the old tools have rapidly become unfit to police: faces.

I spent fifteen years modelling, so I thought I had a reasonable grasp of how a face moves through the commercial world. I had a vested interest in understanding contracts, usage rights, expirations, and territories, because my own identity was the currency that was being governed by them. Two years ago, as generative AI hit the up-ramp, I realised that confidence was being called into question, and that the circumstances had changed.

Suddenly, my likeness started appearing in places I had never agreed to, doing things I had never done. There was no one to call. No takedown form I could fill in. No agency or chartered body ready to intervene on my behalf. It wasn’t even clear who might be responsible: whether an AI image model had ingested enough images of me during training that an extremely close proxy for my face was appearing in new generations, or whether someone using a generative model had uploaded one or more of my prior shoots as a starting point. 

This felt personal, as you can imagine. But the experience wasn’t, and isn’t, unique to me. There are models across the world staring down the same shock right now. Some of them are even starting to make headlines for it, by allegedly taking action in the form of lawsuits. 

Because modelling is such an intrinsic part of fashion, this is all happening here first. But the same structural problem is also going to be coming for basically every other consumer-facing sector, from cosmetics to food to consumer electronics. Because all of these sectors are adopting generative AI workflows around the same time, and pointing them at the same purposes (primarily in-house and then downstream content creation) they are all straying into the same uncharted territory: using, inadvertently or not, other people’s likenesses as fuel for AI workflows.

So what does it actually mean to own and govern data when that data is a person? That would have sounded  like a philosophical question a few years ago, but it’s a practical one today. 

For brands, retailers, and the technology vendors selling to them, it’s also becoming a procurement question, a legal question, and a matter of earning and retaining audience trust – all in an untested space where no-one is directly qualified to provide the answers.

Four Rights Meet The Face Test 

In every other domain, ownership of data can be organised into four primary operational rights: the right to use, the right to exclude, the right to monetise, and the right to revoke. Own a house, and you can live in it, keep others out, rent it, and sell it. Own a song, and you can perform it, license it, and pull it from a streaming service if you like. The rights aren’t always perfectly enforceable, but they do exist, and the parameters for understanding them yourself – and therefore making them intelligible to others – have been tested and re-tested.

Run the same tests on a person’s face inside an AI training dataset, or brought into an AI workspace by a user either blithely grabbing inspiration from a web search and not realising the implications, or purposefully stretching the agreed-upon use of photography that was properly licensed, and we hit the barriers.

The right to use. You can still take a selfie. You can still sit for a photographer. From an input perspective, this right feels secure. But once your face is inside an AI foundation model, or brought into a generative workspace, you can’t exclusively use the model’s version of yourself; anyone can. Your individual right has been lost.

The right to exclude. Just as there is no system to prevent someone from taking a campaign shot, importing it into an AI workspace, and asking one or more generative models to place the person in a new outfit, or a new context, there is no general mechanism for an individual to keep their face out of a training dataset. Once a likeness has been published online, or has found its way into other, narrower, corpuses of data, it is out there for models to train on. Realistically that training has already happened. By the time a model is trained, though, the original images aren’t really considered an asset anymore. The weights (numerical representations) are. And weights don’t have a “remove this face” button any more than they have an easy lever that AI labs can pull to make them universally aligned or unimpeachably accurate.

The right to monetise. Anyone can license their likeness in theory. In practice, monetisation assumes a counterparty who recognises you as the sole owner, or who deals with your representative as a party with authority to license it. When your face is one of millions inside a foundation model, no one is looking at you. They’re looking at outputs. The economic value of generating consumer-facing imagery is real, large, and growing. The share flowing back to the people whose faces helped create it, or whose likenesses get imported by users at runtime,is structurally zero. 

The right to revoke. Even on platforms with takedown processes, such as stock libraries, you would be removing a copy, not the root of the original use. You cannot untrain a model,pull a face out of the weights, or realistically block workspaces from allowing users to import inspiration images of unclear provenance. People ask. People sue, as we’ve already seen. Sometimes they win. But the honest technical answer to “can I get it out?” is no. This is also amplified by the fact that any individual could “pose” as you and resubmit the data without your knowledge.

All of which is to say that the traditional tools of rights enforcement are resolutely unfit for the world we now inhabit. Three of the four agreed-upon rights are being actively contested, while the fourth is fundamentally impossible due to the way pre-training works.

A  Society-Wide Problem, Not Just A Plea For Sympathy From The Famous

It’s easy to read this as a fashion model’s problem, or a champagne issue for celebrities – something that happens to people whose faces are commercially valuable in the first place, and not really a concern for everyone else.

I understand the appeal of that framing. It’s also wrong.

The reason this ownership and accountability  is surfacing in fashion first isn’t that the underlying issue is unique to fashion. It’s that fashion is the first industry where the gap between the commercial value of a face and the legal protections around it became impossible to ignore. Models, by definition, have faces brands want. That goes double for celebrities. That makes the consent gap visible faster, but it doesn’t make it non-existent or irrelevant for people whose faces aren’t currently in the market.

If your photo has ever been on the internet, and that covers most of the audience reading this, you’re already inside the same pipeline. Photos posted to social platforms, or captured at events, get scraped at scale, ingested into training data, then surfaced back into the world as outputs that resemble specific people without being attributable to any of them. When we say that AI models are trained on the internet, we mean all of it – even if the parts with famous people in them are more heavily represented.

For a brand using an AI-generated face in a campaign, a virtual fitting tool, or a synthetic model in a lookbook, the question gets uncomfortable fast: whose face is it actually built on? Today, the honest answer is usually that the brand doesn’t know, even if they pre-selected for demographics. And the people whose faces the models are built on more than likely didn’t agree to any of it – and they certainly didn’t sign any licenses or agreements.

In these kinds of situations, which are becoming more commonplace day by day, if “data ownership” is going to mean anything at all, then it has to mean providing that visibility and equipping people with that control over the recognisable parts of themselves. And just as the problem became visible in fashion first, where the claimants are big names and famous faces, any attempt to mount a solution will also need to start there. But the people that end up being cited in case law will be the standard-bearers for the rest of us.

Legislation Is Looming

As befits the scale of the challenge, and the speed with which generative images are suffusing every part of communications, regulations are crystallising faster than people might realise.

In the US, New York’s Fashion Workers Act already requires written consent for the creation or use of a model’s digital replica, with specific terms about scope and duration. New York also passed a synthetic performer disclosure law requiring explicit labeling when AI-generated performers appear in commercial content. Washington State expanded its personality rights statute to give individuals stronger control over the commercial use of their likeness in AI contexts. In Korea, the AI Basic Act sets out broad obligations for transparency and consent in AI systems that process personal data, including biometric data.

These laws aren’t identical and they aren’t well-coordinated with each other yet, but they’re converging on the same recognition: the existing vocabulary for “data” doesn’t handle faces well and the existing vocabulary for “likeness” didn’t anticipate models that can generate it at scale, at negligible cost. 

For anyone building or buying tools on the current ambiguity, the implication is direct. “We trained on publicly available data” was always a thin defense on the part of model and technology providers, and it’s set to getthinner, even if the world’s biggest tech companies have deeply vested interests in keeping the current fair-use argument going. 

From a brand point of view, “we used a third-party tool and can’t answer for what it was trained on” is also not a defense that the public is likely to accept for long. Procurement diligence on training data provenance is becoming a real category, and, hat question is being asked now, by lawyers, by procurement teams, and increasingly by the talent whose faces ended up somewhere they never agreed to be. To hedge against an uncomfortable answer, we are already seeing companies emphasising the need to be able to account for every input into the models they use.

What You Can Do About It

Data ownership in the AI era is being built in real time, as you read this report. It’s formalising In regulation and court cases, and its sharpest corners are being sandblasted into codified language the contracts, brands and platforms are starting to write. Crucially, it’s crystalling in the awareness  of the people whose faces are inside the systems, even if their likenesses aren’t, right now, their livelihoods. 

Right now, the brands approving AI campaigns, the agencies licensing synthetic models, and the vendors building the tools is working to figure out how to bake consent, rights ownership, and accountability into their workflows – and the decisions they make will determine whether the market for a human face develops into its next form organically, or whether it’s forced. One of those routes will be far more expensive than the other. For everyone.