Is it safe to upload your photos to an AI detector?

Most detectors require an upload, which means handing a copy of your image to somebody else infrastructure. For ordinary pictures that is usually fine; for claim photographs, identity documents and client work it creates obligations people rarely price in.

What an upload actually commits you to

Pasting an image into a web service is a transfer. A copy leaves your device, crosses a network, lands on infrastructure you do not control, and is processed by software you cannot inspect.

For most pictures that is unremarkable and nobody should be alarmed by it. A screenshot from a group chat, a picture off a news site, a listing photograph: none of these carries an obligation, and the risk is close to zero.

The problem is that the interesting cases are almost never those. People reach for a detector precisely when an image matters, and images that matter are disproportionately sensitive: somebody claim photographs, a candidate passport, an unpublished picture, a client deliverable under contract.

At that point the upload stops being a technical detail and becomes a decision with consequences that outlast the check.

The questions worth asking

Before uploading anything that matters, five questions cover most of the exposure. A service that cannot answer them clearly has answered them.

Five questions for any detector

Is the image retained, and for how long? Deleted on completion, kept for thirty days, or kept indefinitely are very different answers.
Is it used for training? Several free services reserve this right. It is a real cost, paid in data rather than money.
Where is it processed? Cross-border transfer matters for regulated data, and often needs its own assessment.
Who can see it? Staff access, subprocessors and anybody with the result link.
Is a contract available? For business use, a processing agreement is usually required rather than optional.

What to establish before an image leaves your machine.

The second row catches people out most often. A free service with no obvious business model frequently has one, and retaining uploads to improve a model is among the more common versions of it.

Where the obligations bite

For an individual, uploading a personal photograph is a personal choice with personal consequences. For an organisation, it is usually a processing decision with rules attached.

Common cases, and what an upload triggers
What you are checking	Upload risk	Why
A picture from a news site	Negligible	Already public, no obligation
A marketplace listing photo	Low	Public, not yours, not sensitive
A claim photograph	High	Somebody personal data, held on trust
An identity document	Very high	Special category data in several regimes
Unpublished client work	High	Contractual confidentiality
A private photo of a person	High	Theirs to control, not yours

The pattern is that risk tracks whose image it is rather than how interesting the answer would be. Checking somebody else photograph by sending it to a third party is a decision made on their behalf, usually without asking them.

For regulated organisations the practical consequence is procedural. A third-party image processor needs a contract, a lawful basis, a retention position and a line in a privacy notice, and the review often costs more than the tool it is reviewing.

Why local analysis changes the question

When the model runs in your browser, the file is read from disk by the page and never transmitted. There is no server to trust, no retention policy to read and no processor to appoint, because no processing happens anywhere but on your machine.

That is a different kind of assurance from a promise. A service that says it deletes uploads is asking you to trust its operations; a tool that never receives the file has removed the thing you would be trusting.

It is verifiable, too, which matters for anyone who has to justify a choice. Open the network panel in a browser, run a check, and observe that the model comes down and nothing goes up. That takes a minute and settles it more convincingly than any policy page.

100%

of the analysis happens on your own device The model downloads once and runs in the browser tab. There is no upload, so there is no retention policy to read, no processor to appoint and no breach that could expose the image.

The number that makes the retention question moot.

The trade is honest and worth stating: the model downloads once, which costs bandwidth, and the analysis uses your processor, which is slower on old hardware than a datacentre would be. That is the whole cost.

Decide whose image it is
If it is not yours, the default should be not to send it anywhere without a reason.
Read the retention and training terms
Two paragraphs, and they are the two that matter. A service reserving training rights on your uploads is charging you in data.
Prefer local analysis for anything sensitive
Claim photographs, documents, private pictures and client work. The check is the same; the exposure is not.
Verify rather than trust where you can
Watch the network activity during a check. It takes a minute and it is the only assurance that does not depend on somebody else honesty.
Write it into your process
A one-line rule about which images may leave the organisation prevents the decision being made ad hoc by whoever is in a hurry.

The threat model most people actually have

It helps to be specific about what could go wrong, because vague unease produces either paralysis or indifference and neither is useful. Four things realistically happen to an uploaded image, in descending order of likelihood.

It is retained and used to improve a model. Common, usually disclosed somewhere in the terms, and mostly harmless for an ordinary photograph. Uncomfortable for a private picture of a person who did not consent to being training data.

It is accessible to staff. Every hosted service has people who can reach stored files for support and debugging, and that is normal engineering rather than misconduct. It is still a set of strangers with access to your image.

It is exposed in a breach. Services holding uploads are targets, and a picture that was fine to share with one company is not necessarily fine to have published. This is the risk that persists long after you have forgotten the check.

It is subpoenaed or requested. Rare, and the most consequential where an image relates to a dispute. A copy on somebody else infrastructure is discoverable in ways a file on your own machine is not.

A proportionate rule

None of this argues for treating every check as a security incident. The proportionate position is the one people already apply to documents: public things can go anywhere, and things belonging to somebody else stay where they are.

Written as a rule it fits in a sentence. If the image is already public, upload it anywhere you like. If it belongs to a customer, a candidate, a claimant or a client, it does not leave the organisation.

Questions people ask

Do AI detectors keep the images I upload?

It varies, and the terms are where the answer is. Some delete on completion, some retain for a period, and several free services reserve the right to keep uploads for model improvement. That last one is a genuine cost, paid in data rather than money.

Is it safe to check a passport or ID?

Not through a service that requires an upload, unless you have a processing agreement in place. Identity documents are among the most sensitive data an organisation handles and are treated as special category data in several regimes. Use a tool that analyses locally.

How is this tool different?

The model and runtime download to your browser and the analysis happens on your device, so the image is never transmitted. That is a structural difference rather than a policy one: there is no server that could retain the file, because no server ever receives it.

Can I verify that nothing is uploaded?

Yes, and it is worth doing. Open the network panel in your browser developer tools, run a check, and watch what happens. The model comes down and nothing goes up. That takes a minute and is more convincing than any privacy policy.

What about checking a client image under NDA?

An upload is a disclosure to a third party, and most confidentiality clauses cover exactly that. Local analysis avoids the question entirely, which is usually simpler than seeking permission or explaining afterwards why the file was sent somewhere.

Is a browser tool less accurate because it runs locally?

No. It is the same class of model doing the same analysis; the difference is where the computation happens. What you trade is a one-off model download and a few seconds of your own processor time, which on modern hardware is barely noticeable.