How to Tell if an Image Is AI-Generated: A 7-Step Verification Guide

When an image looks convincing but somehow feels wrong, the usual response is to count fingers, inspect the text in the background, or upload it to an AI image detector.

These methods may reveal useful clues, but no single visual mistake or probability score can resolve every case. High-quality generated images may contain no obvious defects, while real photographs can look unnatural after compression, filters, stitching, or conventional editing.

A stronger approach considers the file, its source, Content Credentials, platform signals, visual details, and model analysis together. The seven steps below can be used for social-media posts, news images, product photos, message attachments, and other images with an uncertain origin.

The question “Is this image real?” can refer to at least three different issues:

1. Was the image generated or edited with AI?
2. Where did the image originate, and who published it?
3. Did the event shown actually happen, and is the accompanying claim accurate?

AI image detection mainly addresses the first question. Provenance research addresses the second. Fact-checking addresses the third. A camera-captured photograph does not automatically make its caption true. Likewise, evidence of AI editing does not establish that every claim associated with the image is false.

Write down the specific claim you need to verify before examining the image. This keeps the investigation focused.

Whenever possible, obtain the file downloaded directly from its source, exported by its creator, or supplied by the photographer. Record the filename, page URL, publication date, and surrounding text.

Do not begin by taking a screenshot. A screenshot creates a new image and may leave behind camera information, editing records, Content Credentials, and other file-level evidence contained in the original. Reposting, downloading through a messaging app, and repeated compression can cause similar losses.

You can still investigate a screenshot, but note that the original file was unavailable and place less weight on conclusions that depend on metadata or embedded provenance. For more detail, see Why Are Cropped AI Images Harder to Detect?

Use a reverse image search service to look for the same or visually similar images. Check:

• which page appears to have published an earlier version;
• how different sites describe the image;
• whether a higher-resolution or less-cropped version exists;
• whether the image predates the event it supposedly shows; and
• whether a creator, news organization, or fact-checking source explains its origin.

Where available, Google’s “About this image” feature can show when Google may first have found the image or a visually similar version, how other sites use it, and pages that appeared earlier than others.

“First seen by a search engine” is not the same as “first created” or “first published.” A reverse search with no results is not evidence that an image was recently generated. Cropping, mirroring, overlays, and extensive editing may also prevent a close match.

An original photograph may contain EXIF metadata such as camera model, capture time, dimensions, and information about editing software. These fields can provide context, but ordinary metadata can be removed or changed. It should not be treated as conclusive evidence by itself.

Content Credentials serve a different purpose. Based on the C2PA standard, they can record information about a media file’s origin and editing history and use digital signatures to make changes detectable.

Do not stop after seeing a “C2PA present” label. Check:

• whether the signature validates;
• who issued the credential;
• whether it describes capture, generation, or editing;
• whether the current file still matches the credential; and
• whether referenced ingredients or earlier assets are available for verification.

A valid Content Credential can provide strong provenance evidence. Its absence, however, does not make an image suspicious. Many cameras and applications do not yet create Content Credentials, and credentials can be lost through screenshots, transcoding, or platform processing.

Some generation platforms attach verifiable signals to their output. These may include file metadata or an invisible watermark embedded in the media. For example, OpenAI’s public image verification tool checks supported C2PA metadata and SynthID signals to help determine whether an image may have originated from ChatGPT, the OpenAI API, or Codex. Google’s verification features check supported content for SynthID watermarks and Content Credentials.

The scope of these tools matters. A platform verifier is generally designed to answer “Did this come from a supported tool on this platform?” It is not a universal detector for every AI image on the internet.

A verified signal is meaningful provenance evidence. No detected signal means only that the verifier did not find a signal it supports. The image may have been created by another system, generated by an older model, or altered enough to weaken the available signal.

Visual review still has value, particularly for locating areas that deserve closer examination. Look for:

• garbled, repeated, or incomplete text inside the image;
• inconsistent fingers, teeth, glasses, jewelry, or small accessories;
• shadows and reflections that do not agree with the apparent light source;
• unnatural repetition in fences, windows, textures, crowds, or vegetation;
• object boundaries that merge unexpectedly;
• people, clothing, landmarks, weather, or timing that conflicts with the claimed context.

However, normal-looking hands do not prove that an image is not AI-generated. Strange details do not prove that AI was involved. Motion blur, shallow depth of field, low-resolution compression, panoramic stitching, and ordinary photo editing can all create visual anomalies.

When direct provenance evidence is unavailable, an AI image detector can analyze pixel-level indicators associated with generation, synthesis, or local editing. Upload the original file whenever possible rather than a page thumbnail or messaging-app screenshot. If you have both an original and a processed version, analyze them separately and compare which evidence survived.

Review the result in this order:

1. Validated provenance credentials or platform signals: Do they identify an issuer, generation tool, or editing history?
2. Traceable publication history: Can you find a credible earlier version, creator statement, or capture record?
3. Model analysis: How strongly does the current file exhibit AI-generation or editing indicators?
4. Visual observations: Do the visible inconsistencies agree with the other evidence?

A conclusion becomes more useful when independent signals point in the same direction. When the evidence conflicts—or when only a low-quality screenshot remains— the most accurate answer may be “There is not enough information to determine the origin.”

After the review, preserve the submitted file, detection report, request ID, and SHA-256 fingerprint. The fingerprint can help establish whether a later file is byte-for-byte identical to the analyzed file. It does not independently prove that the image is truthful or identify its creator.

An image review does not always need to end with “real” or “fake.” More precise conclusions include:

• A verifiable AI provenance signal was found: A credential or watermark points to a supported generation platform.
• Multiple indicators suggest AI generation or editing: Model analysis and supporting evidence align, but no direct provenance record is available.
• No strong AI indicators were found: The current file does not provide substantial generation signals, but this does not prove camera capture.
• Insufficient information: File quality, source history, or available evidence does not support a stable conclusion.

This language communicates what was actually found without turning a probabilistic result into an absolute claim.

The best way to check whether an image is AI-generated is not to search for one perfect detector. Preserve the original file and context, search for earlier versions, inspect Content Credentials and platform signals, and then use visual review and model analysis as additional evidence.

To analyze an image with ShanHaiYin, use the AI image detection tool and upload a supported JPG, PNG, WebP, GIF, or BMP file. Review the primary finding together with provenance information, model evidence, request reference, and content fingerprint.

For important or disputed images, retain the original source, publication trail, and related records, and continue with human review where necessary. See how to read detection results and content fingerprints, the limitations of AI content detection, and the AI content detection insights.