Practical verification guide
How to Tell If an Image Is AI-Generated
Short answer: one strange detail or detector score cannot prove an image was made with AI. Start with the original file, check any record of how it was made, and compare more than one clue. If the clues disagree or you only have a screenshot, the honest answer may be “not enough information.”
A five-step checking process
- Preserve the earliest original file.A screenshot, chat compression, or social media download often removes metadata, compresses pixel blocks, and detaches Content Credentials. Always request or search for the earliest uncompressed file.
- Look for a cryptographic origin record.Check for valid Content Credentials (C2PA). A certified signature from camera hardware or an editing suite can verify how the file was produced.
- Inspect file details and EXIF metadata.Use our free Image Metadata Checker to inspect camera make, lens model, shutter speed, and creation timestamps.
- Review visual anatomy and lighting patterns.Examine fine anatomical details, lighting direction, eye reflections, and background coherence where generative diffusion models frequently introduce subtle artifacts.
- Compare all clues together.Do visual patterns, metadata, and the image context align? If signals contradict each other or evidence is missing, leave the result inconclusive rather than jumping to false certainty.
Key visual clues: where AI models leave artifacts
Modern AI generators (Midjourney v6, DALL-E 3, Stable Diffusion XL, Flux.1) produce photorealistic textures, but they still struggle with underlying physical logic and complex 3D structures:
- Hands, fingers, and contact points: Look at where fingers touch surfaces or grip objects. AI frequently generates extra digits, unnaturally merged knuckles, or fingers that fade into held items.
- Eyes and corneal reflections: In authentic camera photography, both eyes reflect the same ambient light source (catchlights). AI models frequently generate mismatched reflections, oval pupils, or irises with irregular circumferences.
- Hair and fine edges: Real hair consists of distinct strands with natural depth of field. AI hair often blends into skin, clothing collars, or background elements like painted brushstrokes.
- Text, logos, and signage: Background street signs, book covers, and clothing labels often display characters that resemble letters or typography but cannot be read as real language.
- Symmetry and accessories: Check earrings, eyeglasses, shirt collars, and buttons. AI frequently renders one earring with a different design than the other, or melts eyeglass frames into temples.
- Lighting and shadow consistency: Verify whether cast shadows on the ground align with specular highlights on faces and ambient light directions in the scene.
What each signal can and cannot establish
| Signal | Can support | Cannot establish alone |
|---|---|---|
| Valid, trusted Content Credential | A signed provenance claim is bound to the inspected asset. | That every real image has credentials, or that the depicted event is true. |
| EXIF or XMP metadata | Camera, software or workflow claims recorded in the file. | Authenticity; metadata may be edited, removed or copied. |
| Visual anomaly | A specific region worth closer inspection. | The exact generator, author, intent or complete editing sequence. |
| AI detector score | How a trained neural network evaluated pixel distributions. | A universal mathematical probability that the image was made by AI. |
| No obvious anomaly | Nothing strong was found by the available checks. | Authentic camera capture; newer models may leave minimal visual clues. |
Frequently Asked Questions
How can I tell if an image is AI-generated?
Check the image across three independent evidence channels: inspect fine visual details (hands, pupils, hair, text, and lighting consistency), inspect file metadata (EXIF/XMP headers and camera details), and look for cryptographic Content Credentials (C2PA) from certified capture hardware.
What are the most common visual flaws in AI-generated photos?
Common clues include unnatural skin smoothness ('airbrushed plastic sheen'), distorted hands and fingernails, mismatched pupil catchlights, text rendered as gibberish glyphs, asymmetrical jewelry or glasses, and contradictory cast shadows.
Can an AI image detector prove whether an image is AI or real?
No single detector score can provide absolute proof. AI models evaluate pattern probabilities based on their training sets. Because generators evolve and compression or filters degrade image signals, responsible detection presents inspectable evidence rather than an unverified certainty score.
Why do AI generators struggle with hands, text, and eyes?
Generative diffusion models synthesize images by predicting pixel co-occurrences rather than understanding physical anatomy or 3D geometry. While general body shapes are learned easily, complex articulations (fingers gripping objects, reciprocal iris reflections, linguistic typography) remain frequent failure points.
Do AI images have camera EXIF data?
Original outputs from generators like Midjourney, DALL-E, and Flux do not contain optical camera EXIF data. However, anyone can easily paste or fake EXIF tags using editing tools, and social platforms strip metadata automatically, so missing EXIF is a clue rather than definitive proof.
When is “not enough information” the right answer?
Keep the result inconclusive if you only have a heavily compressed screenshot, the clues disagree, a signed provenance record cannot be validated, or you do not know the context of the image. When an assessment could impact someone’s career, reputation, or safety, scientific modesty and evidence-first transparency must come before false certainty.
Learn more in our guide on how to interpret AI detector scores or what Content Credentials can show.