How to Check Whether an Image Is AI-Generated

You usually cannot prove that an image is AI-generated just by spotting strange fingers or garbled text. Modern generators often avoid those errors, while ordinary editing, compression, panoramas, and camera processing can create similar artifacts. A stronger check combines provenance data, the image's source and history, reverse search, metadata, and visual inspection.

Treat the result as a confidence assessment unless you find verifiable provenance or a watermark from the generating system. “It looks fake” is a reason to investigate, not a conclusion.

1. Check the source and claim first

Before zooming into pixels, ask where the image came from and what it is supposed to prove. Find the earliest post you can, identify the account or publisher, and look for a caption naming the photographer, location, and date. A cropped screenshot with no source deserves more caution than an original file from a transparent publisher.

• Does the account routinely publish original reporting or unidentified viral material?
• Is the image attached to a specific event, and do independent reports show the same scene?
• Has the caption changed as the image moved between accounts?
• Can the uploader provide the original file or additional frames?

For a consequential claim, verify the event independently even if the picture itself appears genuine. A real photograph can be reused with a false date or location.

2. Look for Content Credentials or a known watermark

Content Credentials based on the C2PA standard can record how a file was created or edited and cryptographically bind that history to the asset. A valid credential is useful provenance evidence, but it does not prove that the scene or caption is truthful. Conversely, missing credentials do not prove that an image is fake because many cameras, editors, and websites do not preserve them.

The C2PA consumer guidance explicitly distinguishes verified provenance from truth.

Some generators also add model-specific watermarks. For example, the Gemini app can check an uploaded image for Google's SynthID signal. A positive result supports the conclusion that Google AI created or edited it; a negative result does not rule out another generator or an image whose signal cannot be detected.

3. Run a reverse image search

Search the whole picture, then crop to distinctive regions such as a building, product, face, or background sign. Results may reveal an earlier version, the stock photo used as a base, a creator's disclosure, or the real event paired with a different caption.

Google Lens, Bing Visual Search, and other services index different material, so checking more than one can help. These reverse image search options for phones explain several available tools.

Finding no match proves very little: the image may be new, private, poorly indexed, or substantially cropped.

4. Inspect metadata carefully

If you have the original file, examine EXIF and other metadata for camera model, editing software, creation time, dimensions, and embedded provenance. Inconsistent fields can guide further research. However, social networks commonly remove metadata, editors rewrite it, and anyone with the file can alter ordinary EXIF tags.

Metadata is therefore supporting evidence, not authentication on its own. A camera name does not prove the depicted scene is genuine, and an empty metadata record is common after messaging or screenshotting.

5. Examine text, symbols, and repeated structures

Zoom into shop signs, labels, clocks, license plates, book covers, and logos. Generative errors can include letters that change shape, words that are almost correct, repeated characters, or text that fails to follow the surface perspective.

Also check objects with strong rules: chessboards, musical notation, road markings, reflections, fence patterns, repeated windows, and mechanical parts. Look for structures that merge, stop abruptly, or change count. Recent models can render clean text and regular patterns, so their absence is not evidence of authenticity.

6. Check anatomy, lighting, and physical continuity

Hands and teeth are still worth inspecting, but go beyond counting fingers. Trace each arm to a shoulder, compare earrings and glasses on both sides, and inspect how hair crosses clothing. In group scenes, look for duplicated faces, partial bodies, merged objects, or people whose scale conflicts with their position.

For lighting, compare the direction and softness of shadows across the frame. Reflections should match the objects, viewpoint, and environment. Perspective lines should converge coherently, and objects should make plausible contact with the floor or other surfaces.

None of these anomalies is conclusive. Motion blur, computational photography, a wide-angle lens, heavy retouching, or a composite can produce odd results in a real photograph.

7. Look for inconsistent detail and texture

AI-generated images can alternate between highly detailed focal areas and vague, melted backgrounds. Repeated leaves, jewelry, bricks, fabric patterns, or crowd members may be nearly—but not exactly—duplicated. Edges sometimes dissolve where two objects meet.

Overly smooth skin or cinematic lighting alone is weak evidence. Beauty filters, studio photography, HDR processing, denoising, and ordinary advertising retouching can create the same look.

8. Use AI detectors only as one signal

Automated AI-image detectors can be wrong in both directions and may perform differently after resizing, cropping, or recompression. If you use one, save the exact file tested, note the tool and date, and avoid converting a probability score into a categorical claim.

Prefer a validated provenance record or generator-specific watermark when available. For newsworthy or harmful content, consult an experienced visual-forensics team rather than relying on a consumer detector.

A practical verification order

1. Save the highest-quality version and record where you found it.
2. Read the original post and identify the exact claim.
3. Check Content Credentials and known watermark tools.
4. Reverse-search the full image and useful crops.
5. Inspect metadata from the original file.
6. Compare visual details with independent images of the place, person, or event.
7. State the conclusion with appropriate confidence: verified provenance, likely generated, likely authentic, manipulated, miscaptioned, or unresolved.

The most important distinction is between origin and truth. An AI-generated scene may be clearly labeled and harmless, while an authentic photo can still be used to support a false story.