How to tell if an image is AI-generated
Ten checks for reviewing a suspicious photo, with a worksheet to record what you find. They help you investigate a claim; none is a shortcut to proving that an image is real or AI-generated.
Maintained by DetectFake · Updated
Save your evidence, not just a score
Write down the claim, where the file came from, earlier appearances, and any conflicting evidence before deciding whether to share it. This makes the review useful even when a detector is uncertain.
Download the free media verification worksheet (.txt)The 10 checks
- 1
Zoom to 100% before judging anything
Keep the original file and view it at native resolution. Inspect regions one at a time. Enlarging a small thumbnail cannot recover missing detail, and an unusual-looking detail is not proof of AI generation.
- 2
Read any text in the frame
Check signage, labels and book spines for inconsistent letters or impossible words. These can be clues, but blur and compression can also damage text, and generated images can contain correctly rendered writing.
- 3
Count and follow the hands
Check finger count, joint direction, and whether fingers actually wrap the object they hold. Also check where limbs disappear behind bodies and reappear on the wrong side.
- 4
Test the lighting story
Compare highlights and shadows across nearby objects. An unexplained mismatch deserves a closer look, but multiple light sources, reflections and ordinary editing can also explain it.
- 5
Check reflections and transparency
Compare mirrors, windows, sunglasses and water with the surrounding scene. Consider the viewing angle, crop and reflective surface before calling a mismatch suspicious. Reflections alone do not establish how an image was made.
- 6
Look at texture and skin
Inspect skin, hair, fabric and repeated patterns. Unusual smoothness or repetition may warrant review, but beauty filters, denoising, makeup and lighting can produce similar appearances in real photos.
- 7
Hunt for a local edit seam
Compare the sharpness, grain and lighting around a suspected added object with its surroundings. A mismatch may suggest editing, but convincing edits may leave no obvious seam, and a whole-image detector can miss a small altered region.
- 8
Judge the background separately
Cover the subject and study the background alone. Melted architecture, repeating crowd faces, doors to nowhere and warped straight lines show up clearly once the subject is not drawing your eye.
- 9
Chase the provenance
Search for earlier uses and a higher-resolution original. Check the source account and compare the caption with independent reporting. Inspect original-file metadata and verified Content Credentials separately; missing metadata proves nothing, and a genuine photo can still have a misleading caption.
- 10
Run an automated forensic check
Use DetectFake to obtain an image-classification verdict and supporting signals. Record the reported model and any fallback notice. Treat the score as one screening signal, not a measured probability or proof; disagreement with other evidence calls for more investigation.
Tells that are no longer reliable
“It looks too perfect,” missing metadata, strange hands and garbled text are not decisive tests. Generated images may avoid these artifacts, while real photos may look unusual because of perspective, blur or processing. Source verification matters even when every visible detail looks plausible.
Three situations where context changes the answer
These are illustrative review scenarios, not measured detector results.
- A forwarded screenshot: ask for the original and search for earlier appearances. Compression and missing metadata do not establish AI generation.
- A portrait with smooth skin: compare it with the source and ask about filters or retouching. Smoothness by itself cannot distinguish a real portrait from a generated one.
- A real photo with a new caption: look for the original date and place. A correct camera-origin verdict would still not verify the new story.
Tools and further reading
- Google’s guide to image search explains how to find matching or similar images and the sites using them.
- Content Credentials explains provenance records; use its verification tool with the original file when available.
- DetectFake’s methodology describes local inference, server fallback, and limits on metadata and video analysis.
- Detector accuracy and confidence explains false positives and how to evaluate results.
FAQs
Is there a reliable way to know for certain?
Pixels or a detector score alone do not establish certainty. Combine documented provenance, source verification, visual inspection and any automated signals. Agreement can add context but does not eliminate error; unresolved contradictions mean the claim should remain unverified.
Do AI images always have missing metadata?
No. Authentic photos can lose metadata through sharing or conversion, and generated images can contain metadata or Content Credentials. Check what a verified record actually says about creation and edits. Its presence alone does not establish camera origin or the truth of a caption.
What about AI-edited real photos rather than fully generated ones?
A local edit can be missed when most of the image remains unchanged. Compare the suspected region with an original version if available, check its boundaries and lighting, and do not treat a detector’s original verdict as proof that no editing occurred.
Can a detector prove an image is fake in court?
No. Automated detection is a probabilistic triage tool that documents its reasoning. A formal legal dispute needs a qualified forensic examiner working from the original file and its chain of custody.
Not sure after checking manually?
Run the image through DetectFake and compare its signal list against what you found.
Open DetectFake