Can AI detect AI-generated images?

The useful part

This question is worth asking, but a confident AI answer can be misleading. AI is useful for organising provenance checks and identifying claims to investigate. It is not enough for authenticating a photo from appearance alone.

NIST review of synthetic-content detection provides the basis for the key limitation in this guide. This percentage is our editorial estimate of what AI can reliably handle on its own for this question, not measured accuracy or a probability that an answer is true.

Before you start

Define the decision you need to make and gather the best source material you can access. Remove unnecessary private information before pasting anything into an AI chat. If a decision affects another person, keep the process fair and preserve the original evidence.

How to do it

1. Save the best available original file and record the claim.
2. Check Content Credentials if present and note exactly what they cover.
3. Trace the publication source and compare independent accounts.
4. State what remains unknown instead of declaring the image real or fake.

Copy this prompt

Help me investigate whether AI can detect AI-generated images in my specific situation. Do not give a final verdict yet.

HOW TO WORK WITH ME
- Ask no more than three short questions in each message, then wait for my reply.
- Keep observed facts, assumptions and open questions separate.
- Say what evidence you can actually inspect. Never claim to have opened a file, searched the web or used a product unless you did.
- Give a cautious conclusion with its limitations. Never invent confidence percentages.
- If another person's rights, reputation or safety may be affected, avoid accusations and recommend a fair human review.

START WITH ONLY THESE THREE QUESTIONS
1. What image are you checking, and what claim is being made about it?
2. Do you have the original file or only a screenshot or repost?
3. What reliable source could independently confirm the context?
Then stop and wait.

IN LATER ROUNDS
Ask about the first known publication, original file, embedded provenance and independent reporting. Distinguish 'generated', 'edited' and 'misleading context'.
Ask for the minimum material needed. Let me remove private or confidential information. If you lack the original file, links or browsing access, explain how I can run the check myself.

BEFORE ANALYSIS
Summarise the task and the evidence I provided as numbered items. Put assumptions and missing evidence in separate lists. Ask me to correct the summary. Wait for my confirmation.

AFTER CONFIRMATION
Produce a provenance checklist, source comparison and a cautious conclusion with confidence limits. For every substantive finding, say which supplied item or opened source supports it. Explain what a specialist tool could help me check, and what it cannot establish. Give me an ordered, practical next-step checklist.
First explain how to inspect any available Content Credentials and what that evidence does and does not establish.


USE THIS CHECK SEQUENCE
1. Save the best available original file and record the claim.
2. Check Content Credentials if present and note exactly what they cover.
3. Trace the publication source and compare independent accounts.
4. State what remains unknown instead of declaring the image real or fake.
If a step is impossible because evidence or access is missing, mark it 'not checked' and tell me what I would need to complete it. Separate a tool's signal from your own interpretation. End by asking which unresolved point I want to investigate next.

IMPORTANT LIMIT
No credentials does not mean an image is AI-generated; credentials do not prove that the scene itself is truthful.

Begin with the three starting questions only.

What you should get

You should get a provenance checklist, source comparison and a cautious conclusion with confidence limits. Each finding should point to a real source or to something you supplied, and unknowns should stay visible.

Where AI falls short

Credentials can be absent, stripped or incomplete; visual and detector clues can be wrong. No credentials does not mean an image is AI-generated; credentials do not prove that the scene itself is truthful.

When another tool helps

Start with Content Credentials Verify for a free, separate check. Record what it actually shows and what remains unknown.

TruthScan can support one part of the check. Review its current features, access and results before relying on it. Its output does not settle the whole question.