Fake Image Detector: Check a Photo for AI-Generated Signs

People use “fake image” for several different things. Before checking a picture, separate how its pixels were made from what someone claims the picture shows.

• AI-generated images: a model creates the scene, including subjects, textures, and backgrounds. A photorealistic picture can depict a person or event that was never photographed.
• Partial edits and composites: a photograph is altered by adding, removing, or combining objects or backgrounds. These edits can use AI or ordinary editing software. A changed region does not mean the entire image was generated.
• Face swaps: someone's face or identity is substituted into an existing image. This is a specific form of manipulation, distinct from generating a whole portrait.
• Fabricated screenshots: text, a profile, a payment, or a conversation is arranged to look like a genuine capture. A convincing screenshot can be created without an image generator.

This check covers visible signs of AI generation and substantive edits, including face swaps and composites. It gives one assessment of the image as a whole. Routine cropping, resizing, color or exposure adjustments, and minor retouching alone do not count as fake.

An authentic photograph can also mislead when it is shared with the wrong date, place, or caption. Keep the image and its accompanying claim as separate questions when reviewing a result.

Start with the largest available original and look at the whole scene before zooming in. Check whether details agree with one another, rather than judging a picture from one unusual patch.

• Hands and fingers: follow each finger back to the hand. Look at joints, how objects are held, and whether hands merge into sleeves or nearby objects. Consider whether a finger is simply hidden by the camera angle.
• Text and symbols: read signs, labels, and logos. Look for letters that change shape within the same word, inconsistent repeated branding, or text that blends into its surface.
• Light, shadows, and reflections: compare the apparent light direction across subjects. Follow cast shadows and check whether reflections match the objects and viewpoint.
• Symmetry and repeated structures: compare glasses, earrings, windows, wheels, or patterns. Look for structures that change unexpectedly across the picture, while allowing for perspective and natural asymmetry.
• Background blending: inspect boundaries around hair, clothing, fences, and crowds. Check for objects that merge together, edges that disappear, or background features that stop making sense when followed.

Use these observations to guide a closer review. Compare the original with other images from the same source, search for earlier appearances, and read the detector's limitations before drawing a conclusion.

Likely fake means visible evidence suggests AI generation or substantive manipulation, such as a face swap, composite, or local edit that changes the scene. The result groups these possibilities together. Likely real means visible evidence supports an authentic photographic scene without substantive manipulation. Inconclusive means the visible evidence did not support a clear assessment.

The percentages describe the model's classification probabilities across those answers. Model confidence is reported separately on a scale from zero to one. Read both alongside the selected answer: an inconclusive result is a completed assessment with insufficient evidence, rather than a failed request.

Your image stays in the browser until you start the check. The prepared PNG is then sent through the Aijev server to the model provider, TypeSafe, in one request. Preparing the PNG does not copy the original file’s embedded metadata. The application does not persist images or results. Third-party processing follows the provider's own policies. Replacing or removing an image clears the result and cancels an active request.

The detector shares five requests per public IP in any rolling 60 minutes with image comparison. Requests that reach the API count toward that allowance, including failed requests and invalid uploads. People using the same public IP share the limit. Requests rejected after the limit do not extend the wait.

A visual assessment can help you decide what to investigate next when a picture arrives without a clear source. Pair it with a follow-up that fits the situation.

• News and viral posts: review a questionable image before sharing it, then look for the original publication, photographer, and independent reporting. Check the date and location separately from how the picture was made.
• Secondhand listings: check an unusually polished product photo, then request fresh photographs showing the actual item from several angles. Compare distinctive marks and details across the seller's images.
• Recruitment and résumé portraits: a headshot can prompt a question about whether a picture is synthetic. Follow up through normal application verification; a portrait assessment does not establish a candidate's identity, qualifications, or honesty.
• Dating profiles: review a suspicious profile picture alongside the account's other images and conversation. A live conversation or video call can provide more context than a single photo score.

Keep the original image, its source, and the assessment together if you need to revisit the question. Resolve important claims through corroborating evidence.

The model's image input is experimental and can be wrong or miss details. Illustration, CGI, compression, or an isolated artifact alone cannot establish that an image is fake; a lack of obvious artifacts does not establish that it is real. Ambiguous images may be inconclusive.

A likely fake result does not distinguish full AI generation from partial editing or identify which region was changed. The tool returns one overall classification, without an edit map or separate face-swap score. It accepts still images, not video or audio. A likely real result does not certify that a photograph is authentic or completely unedited.

The probabilities and confidence are model outputs. They are not independently measured detection accuracy, calibrated probabilities of true provenance, or proof of authenticity. Resizing and compression can obscure details, and a convincing synthetic image may lack visible clues.

This tool cannot identify which AI model made an image, verify a person's identity, or establish a caption, date, location, event, or deceptive intent from pixels alone. It does not inspect original metadata, verify watermarks or C2PA credentials, or generate a written rationale. Use original sources, creation records, and relevant context where available.