How to tell if a photo was edited with ChatGPT or another AI tool

AI tools can now change a photo in seconds. You can ask ChatGPT to remove a person, shorten a skirt, add damage to a car, or clean up a messy kitchen, and the result often looks like an ordinary photo. So it is reasonable to ask: was this picture changed by AI?

The honest answer is that sometimes a file carries strong evidence, sometimes it carries weak hints, and often it carries nothing that settles the question. This page explains what you can check, what each kind of evidence is worth, and why "unknown" is often the correct answer.

Start with what is inside the file, not how it looks

Odd hands or shadows can raise questions, but impressions are not evidence you can show someone else. Start with the technical record inside the file. There are three kinds of things to look for, from strongest to weakest:

1. Signed Content Credentials (C2PA). Some AI tools, including ChatGPT image generation, attach a cryptographically signed record to the images they produce. That record can say an AI system was involved, and the signature lets you check that the record belongs to this exact file.
2. Unsigned metadata. Some tools write labels such as an IPTC "digital source type", a software name, or a prompt into the file's metadata. Anyone can write or delete these fields, so they are hints, not proof.
3. Processing clues. Signs that a file was re-saved, recompressed, converted or resized. These show that something happened to the file after it was made. They do not show what, or whether the picture's content changed.

What signed Content Credentials can establish

Content Credentials are a standard (C2PA) for attaching a signed history to a media file. When a file has them, you can check three separate things:

• Is the signature valid? If it is, the record has not been altered since it was signed.
• Is the record bound to this file? The record includes a cryptographic hash of the image. If the file you have still matches, the record describes this file, not some other one.
• Who signed it? The signer's certificate can be checked against the official C2PA trust lists. A recognized signer, such as an AI company's media service, means the claim comes from the party it says it does.

When all three hold and the record declares an AI source type, you have strong evidence that an AI system produced or processed this file.

A real example: the "shorten the skirt" edit

We tested this with a real photo. The original was an iPhone picture. We asked ChatGPT to shorten the skirt, downloaded the result, and ran both files through PIT separately, without telling PIT they were related.

For the ChatGPT file, PIT found Content Credentials signed by OpenAI, with a valid signature, a valid binding to the file, and a signer on the official trust lists. The record declared AI-generated media. PIT's headline was AI PROCESSING ESTABLISHED.

Just as important is what PIT did not claim:

• It did not say the skirt was edited. Nothing in the file identifies which region changed.
• It did not say a real photo was the starting point. OpenAI writes the same signed record for an edit of a real photo and for a brand-new image, so the file alone cannot tell those apart. PIT states this as a limitation.

That is the right level of confidence: the file proves an AI system was involved, and stops there.

What a missing label does not prove

Content Credentials are easy to lose. A screenshot, a re-save in some editors, or an upload to many social apps can strip them. Many AI tools do not add them at all.

So if a file has no Content Credentials:

• It does not mean the photo is genuine.
• It does not mean the photo is AI-made.
• It means this particular kind of evidence is not available.

The same goes for camera metadata. A file that lists an iPhone model and exposure settings is consistent with camera capture, but metadata can be copied or written by software. It supports a conclusion; it does not prove one.

Why "unknown" is often the honest answer

Many tools give a single "AI probability" score. A number like 87% feels precise, but it hides what it is based on. If you rely on the answer in a claim, a dispute, or a story, you need to know why.

PIT does not give a real-or-fake score. It answers four questions separately:

• Creation: how does the evidence say the image was made?
• Alteration / Processing: was it processed or changed after it was made?
• Origin: where did it come from?
• Integrity: how strongly is that evidence tied to this exact file?

Each answer comes with the evidence behind it and its limits. When the evidence can't support a conclusion, the answer is "unknown". In our accuracy benchmark of 235 photos (Galaxy S26 FE and iPhone 17 Pro Max originals plus scripted copies), PIT made no overclaims: it never stated a conclusion stronger than the evidence supported. That is a small, controlled test, and we describe it that way.

A simple checklist

1. Keep the file you were sent. Don't screenshot it; that throws away the evidence.
2. Check for Content Credentials and whether the signer is recognized.
3. Treat metadata and signs of processing as hints, not proof of editing.
4. If you need the photo to hold up, ask the person for the original, or send them a secure link to upload or take it live.

FAQ

Can PIT tell me exactly what the AI changed?

Not from the file alone. Signed records usually say that AI was used, not which part of the picture changed.

If PIT says "unknown", does that mean the photo is fake?

No. It means the available evidence does not support a conclusion either way.

Does PIT store my photo?

No. Photos are analyzed in an encrypted, isolated environment and are not stored.

Can someone remove Content Credentials?

Yes, easily. That is why their absence proves nothing, and why their presence, when valid and bound, is strong.