How to Detect AI-Generated Images

To detect AI-generated images, combine source verification, provenance checks, visual inspection, and a detector that explains its limitations. No single odd detail or detection score can reliably settle every case. The useful question is not just “Does this look like AI?” but “What evidence supports how this image was made?”

This guide gives you a repeatable way to investigate an image without treating suspicion as proof. It also separates two different questions: whether an image was generated or edited with AI, and whether the claim attached to it is true.

1. Define what you need to verify

Before opening a detector, write down the claim. “This photograph shows yesterday’s flood in this town” contains a location claim, a date claim, and an implied claim that a camera recorded the scene. Each needs different evidence.

A genuine photograph with a false caption can mislead without any AI. An AI-assisted edit might change only the background. A fully synthetic illustration may be clearly labeled and entirely appropriate in its context.

• For a news image, prioritize the original source, location, date, and independent reporting.
• For a product listing, ask for additional original photographs or another way to verify the item.
• For an authorship dispute, request the creation history and original files rather than relying on appearance alone.

Keep these questions separate in your notes. “I could not verify the date” does not mean “AI-generated.”

2. Preserve the best available file

Save the original post URL, account name, caption, and the time you accessed it. If possible, obtain the original downloadable image from its creator rather than a screenshot of a repost.

Work on a copy. Avoid repeatedly exporting or resizing the file before inspection: that changes the material you are trying to assess. Record whether your copy is an original, a platform download, or a screenshot. If that is unknown, say so.

Do not upload private documents, intimate images, or confidential client material to a third-party detector without permission and an appropriate privacy review. You can investigate the public source before submitting the file anywhere.

3. Trace the image’s source and earlier uses

Search the whole image, then try a distinctive crop such as a building, sign, or recognizable object. Read the pages behind matches rather than comparing thumbnails alone. Look for earlier captions, higher-resolution copies, creator disclosures, and a consistent sequence of posts.

Google’s About this image documentation describes context that can include when Google first encountered an image or similar images. Treat this as a discovery aid, not a certified creation date. An older indexed copy does not necessarily identify the photographer or original publisher.

No reverse-search match is not evidence of AI generation. A new photograph, a private image, or a heavily changed crop might also have no accessible match.

For example, suppose a dramatic “new” storm image appears on an older page about another country. That challenges the current caption. You can explain the mismatch without making an unsupported claim about the image’s production method.

4. Check Content Credentials and metadata

When Content Credentials are available, inspect the validation result, signer, and recorded actions. Do they describe a generated image, an AI-assisted edit, or a capture-and-edit history? Does the information actually apply to this file?

The C2PA explainer describes cryptographically bound provenance records. A valid record can support an attributed history; it does not independently certify that a depicted event or accompanying caption is true. Missing credentials are neutral, not proof that a file is authentic or synthetic.

Ordinary metadata deserves a separate, more cautious reading. Camera fields and software names can offer leads, but they can be edited or removed. A software export label tells you something about a processing step, not necessarily how every visible element was created.

Similarly, a visible watermark can be copied or removed. Invisible watermark checks are system-specific. Google DeepMind’s SynthID overview describes its watermarking system; it is not a universal test for every generator, and this guide does not imply that AIFakeScan provides a SynthID verifier.

5. Inspect visual consistency, not a checklist of “AI tells”

Start with the whole scene, then inspect details at a sensible zoom level. Look for relationships that do not hold together: an object joining another without a plausible boundary, a reflection inconsistent with the object it should reflect, or repeated structures that change unexpectedly.

Hands, teeth, lettering, shadows, and fine textures can be useful places to look. They are not permanent fingerprints of AI. A convincing image can contain no obvious errors, while an ordinary photograph can contain motion blur, unusual perspective, processing artifacts, or an optical illusion.

Describe the observation before naming a cause. “The railing merges into the sleeve” is a checkable statement. “The pixels prove AI” usually hides assumptions that still need testing.

6. Use an AI image detector as supporting evidence

If you use the AIFakeScan image detector, read the evidence and limitations alongside the result. Our methodology explains that beta evaluation and calibration are incomplete: passive-only results remain Inconclusive, and the service does not present a measured AI probability or validated accuracy percentage.

Check what the tool examined. Did it report a provenance declaration, a metadata clue, or a passive model signal? Those are different kinds of evidence. A result for a whole image also does not necessarily locate a small AI-edited region.

If tools disagree, preserve the disagreement. Do not average unrelated scores or count a majority vote as proof. Ask whether the tools support the relevant media type and whether their conclusions depend on similar signals.

7. Write a conclusion that matches the evidence

Use a short record with four fields: the claim, the file you examined, the evidence found, and the remaining uncertainty. Link to the original source and distinguish your observations from a tool’s output.

Appropriate conclusions might be “The creator labels this image as AI-generated,” “The available provenance records declare an AI edit,” or “The source could not be verified from this screenshot.” Avoid upgrading the last statement to a definitive accusation.

For high-stakes decisions involving a person’s reputation, employment, or potential wrongdoing, a quick detector check is not enough. Seek original material and qualified review before acting or sharing an allegation.

Keep an image-verification worksheet

Copy this blank worksheet into your review notes. It turns the seven steps above into a record that another person can follow. It is a suggested process, not a completed experiment or an accuracy benchmark.

Comparing an original with a screenshot or compressed copy

Preserve the original first. If a comparison is useful, create each derivative from the same untouched source, change one setting at a time, and label the output. Record the editor, dimensions, export settings, and each separate report. A social-platform download may combine several transformations, so describe the actual route rather than attributing all differences to compression.

Keep metadata, credential validation, watermark verification, and visual-model output in separate columns. Do not average them into an authenticity percentage. Without known source history, a changed result cannot establish which version received the correct classification.

Use the SynthID guide for Google's official check, the C2PA guide for credential states, and the false-positive guide when a flag conflicts with other evidence.

Start with the check that matches the claim

Use the C2PA Checker for embedded provenance and signer status. Use the SynthID verification route when the question is specifically whether Google AI created or edited the file. Use the AI Image Detector for a visual-model signal with stated limits.

The 2026 tool comparison maps these evidence types to AIFakeScan, Gemini, Adobe Inspect and a hosted visual API. The screenshot and compression experiment supplies a completed version-comparison example with downloadable hashes and row-level results.

Frequently asked questions

Can you tell if an image is AI-generated just by looking?

Sometimes obvious inconsistencies suggest where to investigate, but appearance alone cannot reliably settle every case. Use source history and provenance where available, and keep inconclusive cases inconclusive.

Does missing EXIF data mean an image is AI-generated?

No. Missing metadata does not identify a production method. Ask how the file was downloaded, exported, or captured before drawing conclusions.

Can an AI detector prove an image is real?

A negative or inconclusive detection result does not establish authenticity. It also cannot prove the attached date, location, or story. Those claims need their own verification.

What should I check first?

Start with the source and the claim. Preserve the best copy, search for earlier uses, and then inspect provenance and visual details. Use a detector to add evidence, not replace the investigation.

Your next step

Gather the original URL and best available image, then run through the checks above. If you want to inspect supported file signals, open the AI image detector and review how AIFakeScan works before interpreting the result.