AI Detector Said Your Photo Is AI. Here Is What Actually Overturns It

You cannot argue with the score, and trying is the most common way people lose a week. A detector's percentage is that model's confidence in its own label — it is not a finding, no vendor retracts it on request, and a second detector's score is not an independent witness. What moves a false positive is a different class of evidence: a provenance credential, or the original file with its metadata intact, put in front of whoever made the decision. This page is that order, plus what to collect before you write anything.

First, work out which of three things happened

"False positive" gets used for three different situations. Only the first is one.

1. The file has no AI involvement at all. You took the photo, nothing generative touched it, the label is simply wrong. This is a false positive, and the evidence ladder below is how you beat it.
2. The file contains synthesised pixels you did not consciously add. Your phone merged frames and invented detail; you used generative fill to remove a sign; an app upscaled the image. The detector is describing the file correctly. What is wrong is the step from "some pixels are synthesised" to "this image is fabricated".
3. The file you are arguing about is not the file that was checked. A screenshot, a re-export from a chat app, a compressed copy from a marketplace listing. The metadata is gone and you are defending a derivative nobody can trace.

Sort this out first, because the three need different arguments. Case 1 needs proof of origin. Case 2 needs a fight about what the label means. Case 3 needs you to find the original before you say another word.

The evidence ladder

Ranked by how much weight a reviewer can put on each. This is the part most advice skips, and it is the only part that decides outcomes.

Rung 5 is last because detectors are correlated, not independent. They are trained on overlapping corpora, share architectures, and are tuned against the same public benchmarks. Five detectors agreeing is closer to one detector asked five times than to five witnesses. A folder of screenshots tells a reviewer that you have no evidence and are hoping volume substitutes for it.

Provenance outranks pattern because of what it measures. A watermark or a signed credential records an event. A pattern score estimates pixels. The first can be present or absent; the second can only be more or less confident. Google's own page for its SynthID tool is explicit that it "is not a general AI detector" and only speaks for content from companies that adopted the technology — which is exactly how to read any provenance result: a hit means something happened, a miss means nobody looked.

What to collect before you write a single message

Six steps, in this order. Skipping the first makes the rest impossible.

1. Stop re-saving the file. Every export, screenshot, social upload and messaging pass rewrites or strips metadata. Once the original is gone, rung 2 is gone with it and you are back to arguing about pixels.
2. Pull the original off the device, not out of the conversation. Camera roll, DCIM folder, or "download original" from the cloud album. Not the copy the platform is showing you.
3. Hash it. Run a SHA-256 of the exact file you are submitting and quote it in your message. It is one command, and it ends the "is this even the same file?" loop before it starts.
4. Capture the scene a second way. A photo of the same subject from another angle, or a screenshot of the camera roll with the capture date visible. That is rung 4, and it takes ten seconds at the time and is impossible later.
5. Copy the EXIF out as text. ExifTool or any metadata viewer, pasted into the message body. A reviewer reads text. A reviewer will not install software to open your attachment.
6. Record what was decided, and where. Screenshot the label itself with the URL and the date. Appeals die because nobody wrote down which surface said what, and when.

What to actually say

Four sentences do more than a page of protest: what was flagged, what you have, what you want, and what would change your mind.

A message that works reads like this: "My upload [file name, SHA-256] was labelled AI-generated on [date] at [URL]. I believe the label is wrong. The file is a camera original — [camera body, lens, capture date] — and the EXIF block is pasted below; a second frame from the same burst and a photo of the same scene from another angle are attached. I am asking for a human review of the label and, if it is upheld, for the specific evidence that would change it."

Three things in that message are doing the work. It names the artefact instead of the accusation. It asks for an action a person can perform. And it asks what evidence would be sufficient, which forces a reason out of a system that would otherwise answer with a template.

Do not open with "is this AI?" You will get a link to the same detector. Do not open with "your tool is broken". Both invite a canned reply, and the first reply is often the only one you get.

Who can actually reverse it

Reviewers differ in what they are able to check, and that decides what you send. Note that a human reviewer is not a better detector: when a study put the same question to people rather than models, the overall success rate was 62% across roughly 287,000 image evaluations by more than 12,500 participants (Roca et al., 2025, "How good are humans at detecting AI-generated images? Learnings from an experiment", arXiv:2507.18640). Bring evidence, not the claim that it obviously looks real.

In the documentation reviewed for this page, no major platform publishes an appeal route aimed specifically at an AI-detection label. That absence is the practical problem: there are forms for a removed post and a suspended account, and the "my photo is not AI" case usually has to be routed through general support. State what you want in the first message, because you may only get one human reply.

The four "false positives" that are not

Some labels are correct about the file and wrong about the meaning. Working out which case you are in changes the argument from "your detector is broken" to "your label is misleading", which is a claim a reviewer can act on.

Computational photography. A modern phone does not take one exposure. It takes a burst, aligns the frames, and synthesises detail no single frame contained — increasingly generating texture where the sensor had none. A detector saying "these pixels were not photographed in one shot" is not necessarily wrong. What is wrong is the step from that to "this is a fabricated image".

Generative fill inside a real photo. You removed a bin, extended the sky, swapped a face in a group shot. The file now contains a synthetic region, and some provenance systems will record the edit honestly. Concede the pixel and dispute the conclusion.

AI upscaling and denoising. Enlargement and noise reduction rewrite exactly the high-frequency detail detectors key on. A two-times upscale can move a file from "real" to "AI" without a single generated object in it.

Merges, stitches and HDR blends. Panorama stitching, focus stacking and HDR merge are algorithmic composites. They have been standard for well over a decade and they leave artefacts that look like generation to a model trained to spot generation.

In all four, the honest sentence is "this is not a camera original", not "this is not AI". Say it before the reviewer says it for you — the second version sounds like a lie, and the first is a claim you can defend.

Five things that make it worse

• Uploading to ten more detectors. Correlated scores plus a privacy cost: you have handed the file to strangers. Rung 5 evidence at its worst.
• Sending the compressed copy. That is the copy with no metadata. You are arguing with your best evidence already deleted.
• Quoting the percentage back at them. "It said 97%" invites the reply that the number is a confidence score, and you have lost the exchange on your own opening line.
• Stripping metadata to "clean" the file. It removes rung 2 and reads as laundering. A reviewer who sees deliberately blank EXIF stops listening.
• Fabricating EXIF to pass a check. This turns a mislabelling dispute into a fraud allegation, and metadata is trivially checkable against the camera's own records.

How to stop this happening to the next photo

Two habits. Turn on Content Credentials wherever the camera or app offers them — a signed credential at capture time is rung 1, and it beats anything you can assemble afterwards. And keep the original: a folder per shoot with the RAW files, not a camera roll that syncs and re-compresses.

On the checking side, start with provenance and only drop to a pattern detector when provenance returns nothing. The SynthID walkthrough covers the first; what a detector's percentage actually means covers why the second should never be your only input.

FAQ

The detector said 99% AI. Does that make my appeal weaker? No. The percentage is the model's confidence, not a measurement of the file, and it carries no weight with a human reviewer. Argue with rung 1 to 4 evidence and do not discuss the number.

Can I just run it through another detector that says "real"? You can, and it usually does not help. Correlated models agree with each other, and a favourable score from a tool nobody has heard of is not something a reviewer can check.

What if I no longer have the original? Then you are arguing with context alone: other frames, witnesses, the camera-roll screenshot, publication history. It is weaker, and it is still worth more than a screenshot of a score.

Is a C2PA credential proof the image is real? No. It is proof that a participating tool handled the file, and a record of what it did. It can show a camera captured it; it can equally show an edit was applied. Read the credential, not just its presence.

How long should this take? The collection steps take minutes if the original still exists and are impossible if it does not. Do them before you appeal, not after the first rejection.

Sources

1. C2PA — technical specifications and Content Credentials overview: c2pa.org
2. Content Credentials Verify: contentauthenticity.org
3. Google DeepMind — SynthID overview: deepmind.google/models/synthid/
4. SynthID Detector portal, supported formats and terms: synthid.com
5. OpenAI — Verify OpenAI-generated content: openai.com/research/verify/
6. Roca et al., "How good are humans at detecting AI-generated images? Learnings from an experiment", arXiv:2507.18640: arxiv.org/abs/2507.18640