An actor looking at a phone screen showing a grid of generated portrait images
Home  ›  Academy  ›  Acting Headshots  ›  AI Headshots
Acting headshots

Nobody had read what these companies actually promise. We read three of them. Two say in writing that they will not train on your photographs and will delete them within thirty days — and all three make it very hard for you to do anything if that turns out not to be true.

So the honest answer is more interesting than the scare story. The privacy risk is real but it varies enormously by vendor, and the casting argument does not depend on it at all. One of the three contradicts its own privacy policy in its own terms of service.

Also on this page: what actually happened to the famous lawsuit, which is not what anyone reports; what changed about Illinois biometric damages in 2024; why a British reader is better protected on paper and not necessarily in practice; and how to get a real headshot this weekend for nothing.

The short answer
No union or casting platform has a rule against AI headshots, and that is not the same as permission. The problem is not that casting cannot tell. It is that they eventually open a door and you walk through it. On privacy, read the vendor’s actual terms rather than trusting either the marketing or the outrage, because they differ wildly. And if you are here because you are broke, skip to the last section.

What these companies actually promise

Everyone writes about AI headshots. As far as we can tell, nobody had opened the terms of service. We read three, and the results genuinely surprised us in both directions.

The vendorWhat their own documents sayWhat we make of it
HeadshotPro
“Neither your User Content (input photos) nor your AI Headshots (outputs) will be used to train or retrain our generative models or third-party models.” Input photos deleted automatically thirty days after your gallery is generated.
A flat prohibition, not a hedge. It rules out the exact thing people assume is happening. No mention of biometric data anywhere, though.
Aragon AI
Names uploaded photographs as potentially biometric data, requires a consent checkbox, and states: “We do not sell, lease, trade, or otherwise profit from face data.” Deletion within thirty days of delivery.
The most careful document of the three by some distance. It also contractually bars its own suppliers from training generalised models on face data.
Secta Labs
The licence covers “training Secta’s AI”, and their own notes say “we keep a copy of the user model and uploaded images for internal use purposes”. Ten to twenty photographs required.
This is the one that matches the scare story. And it contradicts its own privacy policy, which says originals are deleted from all storage they have access to.
What we were expecting, and what we found instead. We went in expecting to confirm that these products quietly harvest faces, and the research we were working from was built on that assumption. Two of the three vendors we read say the opposite in writing, in unambiguous language. That does not mean they are trustworthy — a promise in a terms-of-service document is a promise, not an audit — but it does mean the blanket version of the accusation is not supported by the vendors’ own paperwork. The third one supports it completely and cannot keep its own story straight across two documents. So the useful instruction is not “never upload your face”. It is: read the terms of the specific company, because they are not the same.

One absence is worth naming across all three. None of them mentions the Illinois biometric law by name, and only one treats face data as a distinct legal category at all. For an American consumer product processing facial images at scale, that is a notable gap in itself.

A person reading a long document on a laptop screen late at night

The clause that matters more than the promises

All three vendors share something that gets no attention at all, and it is the thing that decides what a promise is actually worth.

Every one of the three carries a binding arbitration clause and a class-action waiver. If a company breaks its own retention promise, you do not go to court and you do not join anyone else. You go to arbitration, individually, over a thirty-dollar product.

That is not a theoretical concern. It is exactly what happened to the case everybody cites about this.

What actually happened to the Lensa lawsuit.

Lensa’s Magic Avatars — which required at least eight selfies and returned generated portraits — was the subject of a class action filed in February 2023 in the Northern District of California.

On 8 August 2023 the judge granted the company’s motion to compel arbitration. The users had agreed to it when they signed up.

After that, there is no public record. Arbitration is confidential by design, so nobody outside it knows whether it settled, failed or is still going.

So the case cannot honestly be described as won, lost or pending. It was moved somewhere nobody can see, which is what these clauses are for.

Read the two findings together and you get the real shape of the transaction. The promises in the terms are often better than expected. The mechanism for enforcing them is worse than expected. That is a more useful thing to know than either half on its own.

The part that is not an allegation

There is one detail about that app that gets repeated as a lawsuit claim. It is better sourced than that, and it is the single most relevant fact for anyone about to upload photographs of their face.

In December 2022, journalists testing the app found that it would generate sexualised and nude images from photographs of fully clothed people. A reporter at MIT Technology Review put in ordinary photographs of herself and got back a hundred avatars, of which sixteen were topless and fourteen more were in sexualised clothing she had not asked for. She also reported the effect was worse for images reflecting her Asian heritage. TechCrunch published a similar test the week before.

Why we are telling you it this way round. Our research presented this as an allegation in a legal complaint, and it is in the complaint — but the complaint is itself citing the journalism. Reporting by named journalists who ran the test and published the results is a stronger thing than an untested claim by a plaintiff, so we are citing the reporting and mentioning the lawsuit second. It matters because the lawsuit went to arbitration and proved nothing, while the journalism stands on its own. This distinction is the whole reason we grade sources, and it is the difference between a fact and a talking point.

The application to you is direct. You are being invited to upload ten or fifteen photographs of your face to a generative image system. What comes back is decided by a model, not by the policy on the website, and at least one popular product has demonstrably produced outputs that its own stated policy prohibited.

Why a photograph is not the same as your face

There is a real legal distinction underneath all of this, and it is precise enough to be useful rather than alarming.

Illinois has the only American state law that lets an individual sue directly over this. Under it, a photograph is explicitly excluded from the definition of biometric data. A scan of face geometry is not. So the picture itself is not the issue. What a system extracts from it is.

An AI headshot generator cannot function without performing that extraction. That is the hinge on which every one of these cases turns.

And an update our research did not have, which changes the numbers people quote. The statutory damages are a thousand dollars for a negligent violation and five thousand for an intentional one, with no need to prove you were harmed. Until recently a separate violation was counted every single time your data was scanned, which produced some spectacular arithmetic. An amendment effective in August 2024 changed that: repeated collection of the same identifier from the same person by the same method is now a single violation with a single recovery. A federal appeals court confirmed in April 2026 that the change applies retroactively. So any article telling you each scan is worth thousands is working from the old position. We were about to be one of them.

Everywhere else in America, the picture is thinner than it looks. Texas and Washington have biometric laws that only the state can enforce. California’s much-discussed private right of action applies to data breaches, not to a company collecting your face with inadequate consent. So the lawsuits you read about are Illinois lawsuits, and the practices are not Illinois practices.

If you are in Britain

Our research had nothing at all on this, which was a real omission for a site with a lot of British readers. The answer is better than the American one on paper, and slower in practice.

Under UK data protection law, biometric data used to identify a person is a special category requiring a higher standard of protection. The same distinction applies as in Illinois: an ordinary photograph is not automatically in that category, and facial geometry extracted from one for identification purposes is.

Processing it lawfully generally needs explicit consent, which means a clear affirmative act. Consent buried in general terms and conditions does not meet that standard. And you have rights an American outside Illinois simply does not have: access, erasure, objection, and a free complaint to the regulator.

The reality check, because rights on paper are not the same as redress. The regulator fined a facial recognition company seven and a half million pounds in 2022. A tribunal overturned it in 2023 on the grounds that the company was outside UK jurisdiction. A higher tribunal reversed that in October 2025. In December 2025 the company was given permission to appeal again. More than four years on, the case is still not finally decided and no money has been collected. So: better rights, a free route to complain, and no evidence at all that a British reader gets a fast or certain answer. We could also find no case of the regulator investigating this specific category of selfie-based app.
A casting director at a desk comparing a printed headshot with a person sitting opposite

The argument that does not depend on any of this

Everything above is about privacy. The reason not to use an AI headshot is simpler, and it survives every improvement the technology will make.

Alpha Tyler teaches acting at SCAD and was previously a casting director for Tyler Perry Studios. Speaking to a student publication in April 2026, he put it plainly. “I gotta tell you, I don’t like these AI headshots. Casting directors will still see the actor in person, if given an opportunity to audition, so we will know exactly what you look like.”

Notice what that argument does not require. It does not require anybody to detect anything. The generators will get better and the visual tells will go away, and it changes nothing, because the failure happens when you walk through the door rather than when somebody squints at a photograph.

What a headshot is for
The casting director Jane Anderson, in a Spotlight round-up: a headshot should be “a reference to the person who’s going to walk in the room.” That is the entire job description.
The five-year-old photograph problem
Frank Moiselle, in the same piece: “there’s no point having a headshot that is years old or [where you look] too young or touched up in such a way that really it doesn’t represent you.” An old photo fails the test. A generated one fails it in the same direction and adds smoothing on top.
So the answer to the most common rationalisationAn accurate AI image is not better than an inaccurate old one. They are the same failure. A current, honest, imperfect photograph beats both, and the last section explains how to get one for nothing.
An honest note about the sourcing on this section. Alpha Tyler is the only named, current casting professional we could find on the record specifically about AI-generated headshots. We looked hard for a second and there isn’t one. There is plenty of casting comment about AI in casting generally, and about digital replicas in production, which are different subjects. That absence is worth reporting rather than papering over, and it means this page rests one voice deep on its central professional claim. We would rather say so.

Is it actually against the rules?

No, and that remains the most misunderstood thing about this subject. We rechecked every platform today rather than trusting our own earlier table.

SpotlightPhoto guidance requires an image that is natural and an accurate representation of how you look, and bans screen grabs and glamour shots. No mention of AI anywhere. It does require a photographer credit before you can upload.
No rule
Actors Access, Casting Networks, Backstage, MandyNothing published on AI-generated profile images at any of the four, as at today.
No rule
SAG-AFTRA and EquityThe union AI framework covers digital replicas of recorded work in production. It does not cover the photograph on your profile, and no separate headshot guidance exists.
No rule
But something did moveThe International Casting Directors Association published a statement on AI in casting in May 2026, updated in July, supported by Spotlight. It does not ban AI headshots. It does reject unauthorised use of actors’ biometric data and likenesses.
Watch this

So the position holds and it is worth stating carefully. The absence of a rule is not permission. It reflects how recent all of this is rather than any considered decision that the practice is fine, and the industry conversation has visibly started moving even though nobody’s submission rules have changed yet.

A person standing facing a large window while a friend photographs them with a phone from a few feet back

If you are here because you cannot afford a photographer

This is why most people are reading this page, and answering it in a closing line would be a failure of nerve. Here is the whole answer.

The comparison is not what the advertising implies. An AI headshot is not the cheap option. It costs thirty dollars and a set of photographs of your face handed to a company under terms you have not read. A window and a friend cost nothing.

1
Find a window and wait for the right day

A north-facing window on an overcast day is the best free light there is. Stand facing it, close enough that it fills your face softly, with the room darker behind you.

2
Put the camera at chest height and step back

Phone cameras are genuinely fine now, whatever a vendor tells you. Back the phone up and zoom in rather than holding it close, which keeps your features in proportion.

3
Get somebody to talk to you while they shoot

You need a person, not a timer. The difference between a photograph of a face and a photograph of somebody thinking is the person standing behind the camera.

4
Judge it on one question

Are the eyes sharp, and does it look like the person who will walk into the room? That is the entire standard. Not flattering. Accurate.

And there are cheap routes that are still real photography. Photography students building a portfolio, drama school graduates doing the same, skill swaps with another actor, and group sessions that split a studio rate several ways. None of these is as good as an expensive photographer, and all of them clear the bar that matters.

Grading the sources

This page carries legal material, so the standard is higher than usual. Two things in our research were wrong in ways that mattered, and one name on our own existing page could not be verified at all.

The claim, and where it came fromWhat we found
What three vendors’ terms actually sayTheir own published documents
Primary
Read directly and quoted verbatim. As far as we can tell nobody had done this before, and it changed our conclusion rather than confirming it.
The nude outputs from clothed photographsNamed journalists who ran the test, December 2022
Primary
Two publications reproduced it independently, with figures. Stronger than the lawsuit claim our research offered us, and it stands regardless of the litigation.
The photograph and facial-geometry distinctionThe Illinois statute itself
Primary
Read at source. The legal hinge of the entire subject, and it maps closely onto the British definition too.
No platform or union rule existsChecked at each platform today
Primary
Rechecked rather than trusted from our own earlier table. Still true, and a casting-director association statement appeared in May 2026 that our page did not have.
The casting argumentAlpha Tyler, April 2026
Good
A named former casting director, currently teaching, quoted in a dated publication. He is the only such voice we could find anywhere.
The Facetune settlementCourt records and the settlement notice
Good
Around four and a half million dollars, for Illinois users over a seven-year period. Our research named the wrong case; the settlement was reached in a different, later state-court action.
The Lensa case as live proof of harvestingLaw-firm blogs, via our research
No
It was compelled to arbitration in August 2023 and there is no public record since. It proved nothing, and describing it as pending or successful would be wrong.
Every scan is a separate violation worth thousandsThe pre-2024 position, still widely repeated
No
Amended in August 2024 to a single violation and a single recovery, and confirmed retroactive by a federal appeals court in April 2026.
These companies harvest your face to train their modelsThe framing our research was built on
No
Not as a blanket claim. Two of the three vendors we read prohibit it in writing. One authorises it and contradicts its own privacy policy.
A casting source named on our own existing pageCarrie Johnson
No
We could not verify that this person exists in a headshot context at all. The name appears in none of the sources it should. We have removed it pending an explanation.

One gap we are leaving open honestly. We do not know whether commercial, theatrical or background casting differ on this, because no evidence exists either way and our research had none. Rather than reason our way to a plausible answer, we are telling you we do not know.

Where we differ from the standard account

Do these companies train on your face?

Commonly saidYes, that is the whole business model.

What we foundTwo of the three we read prohibit it in writing. The third authorises it and contradicts its own privacy policy.

What happened to the famous lawsuit?

Commonly saidCited as though it proved something.

What we foundCompelled to arbitration in August 2023. No public record since. It established nothing either way.

Can casting directors tell?

Commonly saidYes, the tells are obvious.

What we foundThey will not be for long, and it does not matter. The failure happens when you walk in, not when they look.

Are AI headshots against the rules?

Commonly saidSurely they must be.

What we foundNo platform or union has a rule. That is a gap in the rules, not a verdict on the practice.

Is Illinois law your protection?

Commonly saidBiometric lawsuits are everywhere.

What we foundOnly Illinois lets you sue directly, and the damages rules were narrowed in 2024.

Are British readers covered?

Commonly saidRarely addressed at all.

What we foundBetter rights on paper and a free complaint route. One flagship case has run four years without a final answer.

Is it the cheap option?

Commonly saidThirty dollars against three hundred for a photographer.

What we foundThirty dollars against nothing. A window and a friend cost less than an AI headshot does.

Do this instead of buying anything. Wait for the next overcast day, stand facing your biggest window, and get somebody to talk to you while they take fifty photographs on a phone held six feet away. One of them will be a headshot. It will look like you, which is the only thing the picture has to do.