Can AI Headshots Be Detected? SynthID, C2PA and EXIF

Three different things people mean by "detected", only one of which anybody in your hiring process can actually see.

Every few weeks someone emails us a version of the same question. They have generated a headshot, they like it, and now they are lying awake wondering whether a recruiter is going to run it through some tool and catch them out.

The question deserves a real answer rather than reassurance, because the honest version is more interesting than either of the stories you usually get. One story says AI images are invisibly tagged forever and you will be found out. The other says nobody can tell anything. Both are wrong, and they are wrong in different directions.

Here is what is actually true, including the parts that are not flattering to us.

The short version

Layer	What it is	Survives an upload?	Can a recruiter check it?
Looking fake	Waxy skin, dead eyes, a collar that melts	Yes	Yes, instantly, by eye
SynthID	Invisible pattern inside the pixels	Usually	No
Content Credentials (C2PA)	Signed record in the file header	Usually not	Only if it survives
EXIF	Camera, GPS, timestamp	Usually not	No

The layer that will actually affect how people treat you is the first one, and it has nothing to do with technology. It is about whether the photo looks like a human being.

Three different questions wearing the same coat

Before anything else it is worth separating what people mean, because the same sentence covers three unrelated problems.

"Will someone look at this and think it is AI?" That is a craft question. It is answered by your eyes and by other people's eyes, and it is the only one that changes outcomes for almost everybody reading this.

"Is there something hidden in the file that identifies it?" That is a provenance question. There genuinely is, and we will get into exactly what it is and who can read it.

"Could someone prove it in a dispute?" That is a forensics question. It matters for journalism, court evidence, and insurance claims. It does not meaningfully apply to a LinkedIn profile picture, and treating it as though it does is where most of the anxiety comes from.

Mixing these up produces articles that scare people about a capability nobody is using on a capability nobody has access to, while ignoring the thing that is actually costing them interviews.

What SynthID actually is

Google states that every image generated or edited with Gemini 2.5 Flash Image carries an invisible digital watermark called SynthID. We run that model. That means every headshot this site produces carries one, and we would rather tell you that directly than have you learn it from somewhere with an axe to grind.

The part that catches people out is where it lives.

SynthID is not metadata. It is not a line in the file header that a stripping tool deletes. It is a pattern embedded in the pixel values themselves, calibrated to sit below the point where your eye would notice any difference but above the point where software can still recover it. If you opened the file in an editor and compared it pixel by pixel against a hypothetical unwatermarked version, the differences would be there, and they would be too small and too well distributed for you to see.

That has one consequence worth internalising, because it inverts the advice most people have absorbed.

The standard suggestion for making an image untraceable is to strip its metadata. Remove the EXIF, remove any provenance record, and the file looks anonymous. Stripping metadata does absolutely nothing to SynthID, because SynthID was never in the metadata. It is in the picture. You can delete every byte of header information and the watermark is still sitting there in the pixels, entirely undisturbed.

This is why the "just remove the metadata" advice you find on forums is not so much wrong as aimed at the wrong target. It removes the information about where a file came from. It does not touch the information inside the image.

What survives an upload and what does not

When you upload a photo to a platform, several things happen to it, and they affect these layers very differently.

What happens on upload	Effect on SynthID	Effect on C2PA and EXIF
Re-encoded to JPEG	Usually survives	Usually stripped
Resized to platform dimensions	Usually survives	Usually stripped
Cropped to a square or circle	Usually survives	Usually stripped
Screenshotted by someone else	Usually survives	Gone
Heavy filtering or repainting	Degrades	Gone

Google designs SynthID to remain recoverable through the operations that normally happen to an image on the internet: compression, resizing, cropping, screenshotting, and moderate colour adjustment. It does not claim the watermark is indestructible, and heavy editing does degrade it. But the everyday handling that people assume will scrub an image mostly does not.

Metadata is the opposite. It is fragile by design and by accident. Most large platforms remove it on upload, and they have a good reason to, which we will come to.

Can a recruiter actually check?

No. And this is the part that gets cut from every alarming article on the subject.

Reading a SynthID watermark requires Google's own detection tooling. It is not a public web page you paste a JPEG into. Access has been deliberately limited, initially to a controlled set of testers and partners, because a freely available detector is also a freely available tool for figuring out how to defeat the watermark.

So the situation is genuinely this: the watermark exists, it is durable, and effectively nobody in your hiring process can read it. There is no button on LinkedIn labelled "check this photo". A hiring manager at a 200 person company has no route to that capability, no budget line for it, and no reason to go looking.

SynthID is infrastructure for platforms, researchers and regulators. It was not built as a tool for the person reviewing your application, and it is not being used as one.

Do AI detector websites work?

Somewhat, unreliably, and not in the way their marketing implies.

The free "is this image AI" sites are not reading SynthID or any other official watermark. They have no access to one. They are running their own classifier, which looks at statistical properties of the image and produces a probability. It is a guess made by a model that was trained on some set of examples and is now being shown yours.

That produces two failure modes, and the second is the one that should make you distrust the whole category.

The first is straightforward: they miss things. Newer models produce output that older classifiers were never trained on.

The second is that they flag genuine photographs. A real photo of a real person, retouched by a studio, exported a few times, and compressed by a platform, can come back scored as probably AI. Modern phone cameras make this worse rather than better, because computational photography already blends multiple exposures, smooths skin, sharpens eyes and reconstructs detail. The output of a recent flagship phone has been through a great deal of machine learning before it ever reaches you.

Which means these tools regularly accuse real photographs of being synthetic. If a tool cannot reliably distinguish a genuine iPhone portrait from a generated one, it is not evidence about anybody, and it should not be treated as evidence about you.

What Content Credentials add

Content Credentials, built on the C2PA standard, is the other half of the provenance story and it works on a completely different principle.

Rather than hiding a signal in the pixels, it attaches a cryptographically signed record to the file. That record can describe what produced the image, what edits were applied, and by which software. It is auditable in a way a watermark is not: you can read the manifest and see the chain. LinkedIn surfaces a small "CR" badge on images that carry one, which opens the provenance information.

Because it lives in the file's metadata, it is fragile in precisely the way SynthID is not. Anything that strips metadata strips the credential, and stripping metadata is extremely common.

This creates a tension the industry has not resolved. Platforms strip EXIF on upload largely for privacy, because EXIF can contain GPS coordinates, device serial numbers, and timestamps. Removing that is genuinely protective. It also removes the provenance record, which is the thing meant to establish trust.

So you end up in a position where a missing Content Credential tells you nothing at all. Ordinary photographs lose theirs too. Absence is not evidence of anything, which substantially limits how useful the badge is as a signal until platforms find a way to preserve credentials while discarding location data.

What the EU AI Act requires now

Article 50 of the EU AI Act, the transparency provisions, became applicable on 2 August 2026. Most coverage blurs the two obligations inside it, and the distinction matters a great deal for whether any of it lands on you.

The provider of the AI system has to mark synthetic output in a machine-readable format. That is an obligation on the company operating the model, which in this case is us and, upstream, Google. SynthID is one mechanism for satisfying it. Nothing about that obligation asks anything of you.

The deployer, meaning the person or organisation publishing the content, carries disclosure obligations in specific circumstances. Those provisions are aimed principally at deepfakes and at synthetic content presented as authentic on matters of public interest. A professional headshot of your own face, used to represent you on your own profile, is not what that provision was written to catch.

This area is new and the guidance is still settling, so here is where we stop being confident. If your use is commercial in a regulated sector, or public-facing in a way that gives you pause, that is a question for a lawyer in your jurisdiction and not for a headshot company's blog. We would rather say that plainly than manufacture certainty we do not have about somebody else's legal exposure.

The thing that actually gets noticed

After all of the above, here is what we have concluded from several years of watching how this plays out.

Nobody is running forensics on your LinkedIn photo. There is no queue of recruiters feeding candidate pictures into detection tools. Hiring teams are busy, the tools do not work reliably, and the question of whether your photo was generated has almost no bearing on whether you can do the job.

The realistic risk was never that somebody proves your headshot was generated. It is that they look at it for half a second, register that something is off about it, and quietly think slightly less of everything else on the page. They will not articulate why. They will not run a test. They will just move on with a marginally worse impression, and you will never know it happened.

That is a craft problem, not a technology problem. It is caused by specific, visible, fixable things:

• Skin with no texture. Real skin has pores and slight unevenness. Over-smoothing is the single most common tell.
• Eyes that do not agree. Catchlights pointing different directions, or two eyes that appear focused at different depths.
• A collar that does not sit right. Where fabric meets neck is genuinely hard, and it is where generated images most often fall apart.
• Lighting that does not match the background. A face lit from the left in front of a background lit from the right reads as composited even when every individual element is fine.
• Hair that becomes a helmet. Fine strands averaged into a solid mass, especially with curly or textured hair.

Every one of those is visible to a person who has never heard of SynthID, and every one of them is worth more of your attention than the invisible layer.

The risk people actually underrate

There is a second failure mode that gets almost no coverage and matters more than detection.

You show up to the interview and you do not look like your photo.

This has nothing to do with AI. It happens with professional photography too, and with photos from four years and one haircut ago. But generated headshots make it easier to drift, because it costs nothing to keep pushing until the result flatters you more than you flatter yourself.

The moment somebody meets you and quietly recalibrates, you have spent trust you needed. Nobody says anything. The conversation just gets very slightly harder for reasons neither of you names.

The test we would apply: would a colleague who sees you weekly say "that is a good photo of you", or would they say "that does not really look like you"? The first is what a headshot is supposed to be. The second is a different person with your haircut, and no amount of technical invisibility fixes it.

How to check your own photo before you post it

A short practical routine, worth the four minutes:

1. Look at it at 40 pixels. Shrink it to the size it appears in a feed. If you cannot tell it is you, the framing is too wide. This is also the size most people will ever see it at.
2. Zoom to 100 percent on the skin. If there is no texture at all, it will read as synthetic to anyone who looks closely.
3. Check both eyes for catchlights. The bright reflections should be in roughly the same position in each eye. Mismatched catchlights are a strong tell.
4. Follow the collar line. Trace where the fabric meets the neck and shoulder. Look for edges that dissolve or double back.
5. Ask one person who knows you. Not "does this look AI", which primes them. Ask "does this look like me?" The answer to that is the one that matters.

If the photo passes all five, the invisible layer is a footnote. If it fails any of them, no metadata scrubbing in the world will save it, because the problem is visible.

What we do, specifically

We do not remove SynthID from the images we deliver, and we would not, even though we technically could try. It is doing something reasonable, the obligation to mark synthetic output is a fair one, and a company quietly stripping provenance from its output is not a company we would want to be.

We do strip the metadata from your original upload, which means the GPS coordinates, camera serial and timestamp in the selfie you sent us never travel onward into anything we produce. Your source photo is processed in memory, never written to disk, and deleted within 24 hours.

We also do not claim our post-processing leaves the watermark perfectly intact. We apply tone grading, fine grain and a crop, and any pixel-level operation has some effect on a pixel-level watermark. We neither add to it nor attempt to remove it, and we are not going to pretend to precision about somebody else's system that we have not measured.

That is the whole picture, including the bits that would be easier to leave out.

The bottom line

Can AI headshots be detected? Technically yes, by Google, using tooling nobody in your hiring process has. Practically, the only detection that will ever affect you is somebody's eyes deciding your photo looks wrong.

So spend your attention there. Get a photo that has skin texture, consistent eyes, a collar that behaves and lighting that matches its background. Make sure it looks like the person who will walk into the room. Then stop worrying about the watermark, because the watermark is not the thing that was ever going to cost you the job.

FAQ

Can anyone tell my headshot is AI-generated? By eye, sometimes, if it carries the usual tells: skin with no texture, mismatched catchlights in the eyes, or a collar that melts into the neck. Technically, images from Gemini carry an invisible SynthID watermark, but reading it requires Google's own detection tooling, which is not publicly available. In practice people judge visually and nobody runs forensics on a profile photo.

Does removing the metadata remove the AI watermark? No, and this is the most common misunderstanding about how this works. SynthID lives in the pixel data, not in the file header. Stripping EXIF or a Content Credential removes information about the file, but the watermark is inside the image itself and is completely unaffected by it.

Do AI image detector websites actually work? Not reliably enough to trust. They are not reading any official watermark, because they have no access to one. They run their own classifier that guesses from visual patterns, and they produce false positives on genuine photographs, particularly retouched portraits and images from modern phone cameras that already apply heavy computational processing. A tool that cannot separate a real iPhone portrait from a generated one is not evidence about anyone.

Does LinkedIn detect or flag AI headshots? LinkedIn displays a Content Credentials badge on images that carry one, which is a label rather than a penalty. It does not run SynthID detection on profile photos. Its policy asks that your profile represent a real person, which an accurate headshot of your own face does.

Does the EU AI Act mean I have to disclose an AI headshot? Article 50 became applicable on 2 August 2026. The machine-readable marking obligation falls on the AI provider, which is us. Disclosure obligations for deployers are aimed at deepfakes and at synthetic content presented as authentic on matters of public interest, not at a professional photo of your own face. If your use is commercial in a regulated sector, get advice specific to your jurisdiction.

Will the watermark survive if I crop or resize the photo? Google designs SynthID to remain detectable through cropping, resizing, screenshotting and compression. Heavy editing degrades it and Google does not claim it is permanent, but ordinary handling is not going to remove it.

What is the difference between SynthID and C2PA? SynthID is invisible and lives in the pixels, so it survives most handling but cannot be read without Google's tooling. C2PA Content Credentials are a signed record in the file's metadata, so they are readable and auditable by anyone but are stripped whenever metadata is stripped, which most platforms do on upload. One is durable and opaque; the other is transparent and fragile.

Why do platforms strip the metadata if it is meant to prove where an image came from? Because the same metadata block can carry GPS coordinates, device identifiers and timestamps. Removing it protects users who would otherwise publish their home location with a photo. The side effect is that it also removes the provenance record, which is a real tension the industry has not solved yet.

Should I disclose that my headshot is AI-generated? It depends on context rather than on a rule. For a LinkedIn profile or a resume, a headshot that genuinely looks like you is doing the job a headshot is meant to do, and no disclosure norm has formed around it. For journalism, regulated professions, or any situation where the image itself is evidence, say so plainly.

What actually matters more than any of this? Whether the photo looks like a real person, and whether it looks like you. Those two things decide how people react. The invisible layer decides nothing that will ever reach you.