You shot a few thousand portraits over the last two years. Now a data broker wants to license the set to train a facial-recognition model, or a stock library is offering a bulk buyout for an AI training pipeline. Good money. But then someone asks the question that stops the deal cold: do your releases actually cover this?
Most of the time, the answer is no. The release you signed on set was written for ad campaigns and magazine spreads. It never imagined a machine reading a face into a vector and spitting out synthetic images. This is a real and growing use of a model release form AI training setup, and the standard form was not built for it.
Let's walk through where a normal release falls apart, the exact clauses you need to add, and how to collect AI-ready consent before the shoot wraps.
Why a Standard Model Release Falls Short for AI Training Data
A typical release grants the photographer the right to use a subject's likeness in "advertising, promotion, and publication." That language was drafted for a world where images were consumed by human eyes. It assumes a fixed, visible use: a billboard, a website banner, a print catalog.
AI training breaks every one of those assumptions. Here is the mismatch.
The subject's face becomes an input, not an image. When a photo trains a model, the likeness gets abstracted into data. It can influence outputs that look nothing like the original photo, or that recombine the subject's features into new synthetic faces. A grant to "use the photograph" does not clearly cover a grant to "extract biometric or facial-geometry data from the photograph."
The use is downstream and open-ended. Your standard release covers what you do with the image. It rarely covers what a third party does with a derived dataset years later. AI training is almost always a chain: you sell to a broker, the broker licenses to a model developer, the developer ships a product. If your release does not name that chain, each handoff gets legally thin.
The word "likeness" may not reach biometric data. Some state biometric-privacy statutes treat facial geometry as a separate protected category from a plain photograph. A release that says nothing about biometric identifiers leaves a gap a buyer's lawyer will find in about thirty seconds.
This is a different problem from the one covered in our commercial vs. editorial breakdown. Commercial use is still a human-facing use. AI training data is a machine-facing use, and the consent language has to say so out loud.
The 5 Licensing Clauses Your AI Training Release Must Include
If you plan to monetize images as training data, these clauses need to be in the release the subject signs. Not added later. Not implied. Written in. Think of them as the core of any model release form AI training buyers will accept.
1. Explicit AI and machine-learning grant. Name the use directly. The release should state that the subject consents to the images being used to train, develop, test, and improve machine-learning models, artificial intelligence systems, and algorithms. Vague words like "any purpose" do not do this reliably. Buyers want to see the specific grant.
2. Biometric and facial-data consent. Say that the grant includes extraction and processing of biometric identifiers and facial-geometry data derived from the images. This is the clause that addresses biometric-privacy statutes head-on. If you skip it, a facial-recognition buyer cannot rely on your release.
3. Sublicensing and transfer rights. AI datasets change hands. The release must allow you to sell, license, assign, or sublicense the rights to third parties, including dataset brokers, stock libraries, and model developers. Without a transfer clause, your consent stops at you and cannot follow the data down the chain.
4. Synthetic-output and derivative-works clause. Training produces new images that may resemble the subject or be built partly from their features. Get consent for outputs, derivatives, and synthetic media generated from the training data. This is the clause most standard forms miss completely, because no one drafting a 2015 release was thinking about synthetic faces.
5. Scope, term, and revocation terms. State how long the grant lasts and whether it is revocable. Training data is hard to "un-train" once a model is built, so buyers want an irrevocable, perpetual grant for the training use specifically. Be honest with the subject about that. A revocation clause that promises removal you cannot deliver is worse than no clause at all.
You can build a base form with a model release form generator and then confirm these five items are present before you ever pitch a dataset. If a clause is missing, the release is not AI-ready, full stop.
How California and New York Right-of-Publicity Law Raise the Stakes
Two states set the tone for the whole country, and both raise the bar for AI training use.
California protects the right of publicity under Civil Code § 3344, which covers the use of a person's name, voice, signature, photograph, or likeness for purposes of advertising or selling. Selling a dataset built from someone's face is squarely a commercial use of their likeness. On top of that, California's biometric and privacy landscape, including the CCPA and its treatment of biometric information, means facial data pulled from your images can be regulated as personal information in its own right. A general California release that does not mention biometric processing leaves that flank open. Our California model release breakdown for influencer shoots covers the § 3344 basics, and AI training sits on top of all of them with an extra layer of biometric exposure.
New York protects the right of privacy under Civil Rights Law §§ 50 and 51, which require written consent to use a living person's name, portrait, or picture for advertising or trade. New York also passed protections addressing digital replicas and the use of a person's likeness in synthetic media. That matters directly for the synthetic-output clause above. A New York subject whose face trains a model that generates lookalike images has a stronger claim than they would have had a decade ago. The New York §§ 50, 51 checklist is the starting point, and the synthetic-media angle is the reason AI use needs its own named consent.
The practical takeaway is simple. In both states, "we had a release" is not the same as "we had a release that covered training data." The grant has to reach the specific use. A buyer operating in California or New York will demand that the paperwork match the statute.
None of this is legal advice, and no release guarantees an outcome in any particular court. These are templates based on standard industry practice. If a specific dataset deal is on the table, that is a conversation for your own attorney.
Dataset Sale vs. Dataset License: Why the Distinction Changes Your Release
How you monetize the images changes what your release needs to say. The two common paths are a sale and a license, and they are not the same.
A dataset sale is a buyout. You transfer the images and the rights to a buyer, often permanently, and you walk away. Bulk sales to a stock library or a data broker usually work this way. Because you are handing over control, your release must include full transfer and sublicensing rights. The buyer will resell or relicense the set, so the consent has to survive multiple handoffs. If your release only granted rights "to the photographer," the buyer cannot legally use what they bought.
A dataset license is a rental. You keep ownership and grant limited rights for a defined use, term, and sometimes a defined model. Synthetic-data pipelines and some facial-recognition training projects license this way, because they want a specific scope and are willing to pay per use. Here your release needs precise scope language that matches what you are actually licensing, plus the term and revocation terms from clause five. You cannot license a broader right than the subject granted you.
The trap is licensing more than you own. If the subject consented to "editorial use" and you license the image for facial-recognition training, you have granted a right you never had. That exposure lands on you, not the buyer. This is the same principle behind knowing where a property release is needed versus a model release: you can only license the rights the paperwork actually secured. For datasets, the chain is longer and the buyers read the fine print harder.
How to Collect AI-Ready Releases On Set Before the Shoot Wraps
The cleanest fix is to get the right consent at the moment the subject is standing in front of you. Chasing signatures months later, once a dataset deal appears, is a losing game. People change numbers, change minds, and stop replying.
Here is the sixty-second flow that actually works on set.
Pick a release that already includes the five AI clauses. Fill in the subject's name, the date, and the shoot details. Add the specific AI-training scope so the subject sees exactly what they are agreeing to. Hand them the device, let them read it, and collect the signature before they leave the set. Save the signed PDF to the shoot folder with the images.
This is the job SignedShoot is built for. It produces model and property release documents as PDF and .docx files, built right in the browser, so subject details are never uploaded to a server. You generate the form, get it signed on set, and keep the documentation with the shoot. The free preview is a complete, usable watermarked PDF, so you can see the exact language before you commit.
Doing this on set matters more for AI use than for any other case. With a normal campaign, you can sometimes re-shoot or swap a subject. With a training dataset, once a face is in the set and the deal is done, there is no clean way to pull one signature back out. Get it right at capture. If you want the fuller argument for on-set forms over paper you file later, our comparison of a generator versus a free template lays it out.
For shoots with groups, remember that a single release rarely covers a crowd. A background full of unsigned faces in a training set is its own liability. A crowd release form handles the general-audience scenario, and minors always need the parental consent covered by a minor model release form. Do not fold either into a standard adult release and assume it holds.
Quick Checklist: Is Your Existing Release AI-Training Ready?
Pull your current release and run it against this list. If you cannot check every box, the form is not ready to sell as training data.
- Does it name machine learning, AI, or algorithm training as a specific use, not just "any purpose"?
- Does it grant consent to extract and process biometric or facial-geometry data from the images?
- Does it allow you to sell, license, assign, and sublicense the rights to third parties?
- Does it cover synthetic outputs and derivative works generated from the training data?
- Does it state a clear term and whether the grant is revocable for the training use?
- Does it match the right-of-publicity and privacy rules of the states where your subjects live, including California § 3344 and New York §§ 50, 51?
- Was it signed by every identifiable person, with a separate signed minor model release form for anyone under 18?
- Is the signed copy saved with the shoot, not sitting in a pile of unsigned paper?
Miss even one, and a buyer's counsel can flag the set as high-risk. That either kills the deal or drops the price.
The good news is that this is a paperwork problem, and paperwork problems have clean fixes. Build a release that names the AI use, get it signed while the subject is still on set, and keep the file with the images. If you want to see how the whole approach fits together, SignedShoot's story explains why documenting rights before the shoot wraps beats chasing them later. Do that, and when the dataset offer lands, your answer to "are your releases AI-ready?" is a plain yes.
This article is general information, not legal advice, and no release guarantees a result in any court or jurisdiction. For a specific deal, talk to your own attorney.