How to Tell If an Image Is AI-Generated (2026 Guide)

No single visual clue can prove that an image is AI-generated. Modern generators can produce convincing portraits, product shots, and documentary-style scenes, while ordinary editing, compression, and poor lighting can make a real photograph look synthetic. The reliable approach is to combine visual inspection with source checks, provenance, metadata, reverse image search, and an automated detector.

1. Look at the hands and fingers

Hands can still reveal generation mistakes, especially in crowded scenes or images containing several interacting people. Look for merged fingers, impossible joints, jewellery that changes shape, or hands that do not appear to grip an object correctly. Treat these as prompts for further checking, not proof: motion blur, retouching, perspective, and low resolution can create similar distortions in genuine photographs.

Inspect small functional details as well as anatomy. A real scene usually preserves how straps connect, how glasses sit behind ears, where buttons attach, and how tools meet a hand. Generated images sometimes produce locally plausible objects that fail when you trace their structure from one end to the other. Zoom in, then zoom back out so you do not mistake a single compression artefact for a pattern across the image.

2. Check the background carefully

Backgrounds deserve the same attention as the main subject. Trace railings, shelves, cables, window frames, and repeated patterns across the scene. Generated details may dissolve, duplicate, or change direction where attention is low. Text can be misspelled or structurally strange, although newer systems often render short words correctly and real photographs can contain blurred text after resizing.

Reflections are particularly useful because they must agree with the rest of the scene. Check whether a mirror contains the expected subject, whether a polished table reflects the right objects, and whether windows show a physically plausible viewpoint. Inconsistent reflections are meaningful when several relationships fail together; one missing reflection may simply be outside the crop or obscured by glare.

3. Examine faces for symmetry and texture

Faces can look unusually polished, but beauty filters and portrait retouching create the same effect. Compare both eyes, earrings, glasses, teeth, hair strands, and the boundary between skin and hair. Catchlights should broadly agree with the lighting direction. Differences between left and right are normal in real people, so look for physically impossible structure rather than ordinary asymmetry.

Also ask whether the expression and pose make sense together. A face may appear sharp while ears, hair, or clothing melt into the background. In group photographs, compare how much detail each person receives. A generated image may preserve the central face but produce inconsistent teeth, eyes, or limbs among people farther from the camera.

4. Look for impossible lighting

Identify the strongest light source, then follow its effect across faces, objects, shadows, and reflections. Shadows should fall in compatible directions and objects made from similar materials should respond consistently. Pay attention to contact shadows where feet, furniture, or products meet a surface; missing or floating contact shadows can indicate synthesis or compositing.

Lighting is not a standalone verdict. A real image may combine flash, window light, street lighting, and computational photography, while a skilled composite can reproduce consistent shadows. Record each inconsistency and ask whether a plausible photographic explanation accounts for all of them.

5. Check metadata and reverse image search

When you have the original file, inspect its EXIF or XMP metadata for a camera model, editing software, capture time, and provenance fields. Metadata can support a conclusion but cannot prove it: social networks often remove EXIF, screenshots have little useful metadata, and fields can be edited. Missing camera data therefore means “unknown”, not “AI-generated”.

Run a reverse image search and examine the earliest credible appearance. A real news photograph may lead to a photographer, agency, event page, or older uncropped version. A synthetic image may trace back to a generation showcase, prompt gallery, or account that labels its work as AI. Search distinctive objects or captions as well as the whole image, because a misleading post may reuse a genuine photograph with false context.

6. Check provenance and Content Credentials

Content Credentials use the open C2PA standard to attach signed information about how compatible media was captured or edited. When present and valid, credentials can show which application handled a file and whether generative tools were declared. They are stronger than ordinary editable metadata because the assertions are cryptographically signed and tampering can be detected.

Absence is not evidence of deception. Most cameras and publishing systems do not yet preserve credentials end to end, and screenshots or platform processing can break the chain. Use provenance as positive evidence when it exists, then continue with source and context checks when it does not.

7. Use an AI image detection tool

Automated detectors examine statistical and visual patterns that are difficult to judge by eye. A useful result is a probability or confidence signal, not an absolute declaration. The score can change after cropping, screenshotting, heavy JPEG compression, filters, or editing because those transformations alter the same low-level patterns the model evaluates.

Use the highest-quality version available and keep the original aspect ratio. If a result is consequential, compare the detector output with provenance, source history, and visible evidence. A high score with a credible original camera file deserves investigation, just as a low score should not override a source that openly identifies the image as synthetic.

• Upload any image file or paste a social media post URL
• Receive a 0–100% AI probability confidence score in under a second
• See a clear "Authentic", "Suspicious", or "AI Generated" verdict
• Works with Midjourney, DALL-E, Stable Diffusion, Adobe Firefly, and more

A repeatable image-verification workflow

• Preserve the original file or highest-resolution version before platforms recompress it.
• Identify who first posted it, what they claim, and whether the account has first-hand access to the scene.
• Inspect anatomy, object structure, text, reflections, lighting, and repeated patterns without relying on one clue.
• Check metadata and signed Content Credentials when available.
• Use reverse image search to find earlier versions and verify the accompanying context.
• Run an automated detector and interpret its score alongside the other evidence.
• For high-stakes use, contact the photographer, publisher, or subject through an independently verified channel.

Write down what each check establishes. “No EXIF” is weak evidence, “the claimed news agency has no matching photograph” is contextual evidence, and “valid signed capture credentials from a known camera” is stronger provenance evidence. Keeping those categories separate prevents one dramatic visual oddity from deciding the whole case.

Common false positives and difficult cases

Portrait mode, HDR, denoising, upscaling, beauty filters, background replacement, and aggressive sharpening can all produce an “AI look” in genuine photographs. Illustrations and 3D renders are synthetic media but are not necessarily generated by AI. Conversely, an AI image can be printed, photographed, and uploaded again, hiding many generation traces while leaving the false scene intact.

Mixed-origin images are another difficult category. A real photograph may contain a generative fill, replaced face, expanded background, or AI-enhanced product. The right question may be “what part was altered?” rather than “is the entire image real or AI?”. Preserve nuance in your conclusion and describe the evidence you actually observed.

Quick AI-image verification checklist

• Do the source, caption, location, and date agree?
• Can you find an earlier or higher-resolution version?
• Do anatomy, text, object connections, shadows, and reflections remain consistent?
• Does the file contain useful metadata or signed provenance?
• Does a detector score support, rather than replace, the other evidence?
• Would an incorrect conclusion cause harm? If so, verify with the original source.

How to document your conclusion

For journalism, moderation, education, or workplace review, keep a short evidence log. Record the original URL, account, retrieval time, file hash if available, metadata observations, reverse-search results, provenance status, detector output, and any direct confirmation. Save the untouched original separately from annotated copies. This lets another reviewer reproduce the process and protects against a later edit or deletion changing what was available at the time.

Phrase the conclusion at the level of certainty the evidence supports. “Confirmed AI-generated” is appropriate when the creator discloses generation or valid provenance records it. “Likely AI-generated” can describe several aligned signals without direct confirmation. “Unverified” is the honest answer when evidence is limited or contradictory. Avoid naming a person as deceptive unless you can also establish authorship and intent, which image analysis alone cannot do.

The bottom line

The strongest answer to “is this image AI-generated?” comes from independent signals that agree. Begin with the source and original file, inspect the scene without treating quirks as proof, look for provenance, search for earlier copies, and use a detector as a measured confidence signal. If the evidence remains mixed, say so. A careful “unverified” conclusion is more accurate and more useful than a confident label built from one odd hand, missing metadata field, or model score.

Sources and further reading

• C2PA specifications — the open technical standard for signed media provenance
• Content Credentials — consumer guidance for viewing and understanding provenance information