How to Tell If a Photo Has Been Edited or Is AI-Generated
By Christopher Elley, Founder, FactHeck · Last reviewed 28 May 2026
Written with AI assistance and reviewed for accuracy by the author.
Telling if a photo has been edited or is AI-generated involves two parallel checks. First, look at the image itself: AI-generated people often have extra fingers, distorted jewellery, garbled text in signs or clothing, and backgrounds that blur inconsistently at the edges. Heavy edits leave noise inconsistencies that show up in error level analysis (ELA). FotoForensics runs this for free and highlights regions saved at a different compression level than the rest. Second, examine the metadata: EXIF data embedded in unmodified photos records the camera model, lens, and GPS coordinates. A claimed press photograph with no EXIF, or EXIF that was stripped and re-added, is a red flag. Free AI-image detectors such as Hive Moderation and Illuminarty can give a probability score, though no tool is 100% accurate on highly compressed social media images.
Why photo fakery is spreading: the two main types
Two distinct techniques underpin most fake images circulating today. Traditional editing tools such as Photoshop and GIMP can remove or add objects, change lighting, or swap faces into existing photographs. AI generation tools (Midjourney, DALL-E, Stable Diffusion, and others) go further by creating entirely synthetic images that have no underlying real photograph at all. A third, hybrid category uses AI in-painting to edit a real photo: removing a person from a crowd, or placing a public figure in a location they never visited. All three types spread as apparent documentary “proof” on social media, often faster than corrections can follow. Researchers at Sensity AI (formerly Deeptrace) have tracked the rapid growth in synthetic image volume, and the World Economic Forum's Global Risks Report lists AI-enabled misinformation among its top near-term threats.
Visual inspection: what to look for in AI-generated images
AI image generators still make characteristic errors. Knowing them lets you catch synthetic images before you reach for any tool.
| Body part / region | What to look for | Reliable on its own? |
|---|---|---|
| Hands and fingers | Wrong finger count, unnatural articulation, fingers fusing or splitting | Fairly; AI still struggles here |
| Text in the scene | Garbled letters on signs, clothing, documents | Fairly |
| Background coherence | Objects blending into each other, walls that “melt” at the edges | Weak; better generators fix this |
| Teeth and eyes | Unnaturally symmetric; identical reflections in both eyes | Weak |
| Ears and jewellery | Asymmetric or misshapen earrings, partially merged with hair | Fairly |
| Skin texture | Unnaturally smooth or plastic-looking in portrait close-ups | Weak |
Visual inspection: signs of traditional Photoshop editing
- Clone stamp marks: repeated texture patterns in grass, crowd scenes, or water.
- Soft or jagged edges around pasted-in objects, especially where they meet complex backgrounds.
- Lighting direction mismatch: the subject is lit from a different angle than the background.
- Missing or misplaced shadows: a shadow absent, pointing the wrong way, or the wrong length for the apparent light source.
- Pixel-level noise: pasted objects often carry different noise and grain from the original photograph.
Tool 1: Error Level Analysis with FotoForensics
Error Level Analysis re-saves a JPEG at a standard quality level and compares the result to the original. Regions that were edited after the original save appear brighter, because they carry a different compression signature. To use it: go to fotoforensics.com, upload the image, then select “ELA” from the analysis menu. A uniformly grey result suggests the image has not been selectively edited; bright patches point to recently altered regions. The FotoForensics ELA tutorial explains how to read the output in detail. One important caveat: ELA is less reliable on images that have been re-compressed multiple times (such as screenshots of screenshots, or images saved from social media), because repeated compression cycles mask the original editing signature.
Tool 2: EXIF metadata inspection
EXIF metadata is embedded automatically by cameras and smartphones. It records the camera make and model, date and time, lens, shutter speed, and often GPS coordinates. ExifTool (command-line, free) reads EXIF from virtually any file format. Warning signs in EXIF include “Software: Adobe Photoshop” in a supposedly candid news photograph, GPS coordinates that do not match the claimed location, or EXIF that appears to have been stripped and re-added. One important nuance: most social media platforms strip EXIF on upload, so the absence of metadata is common and is not itself proof of fakery.
Tool 3: AI-image detectors
Several free tools will score the likelihood that an image was generated by AI rather than captured by a camera.
| Tool | What it detects | Free tier? | Key caveat |
|---|---|---|---|
| Hive Moderation | AI-generated images and video | Yes (web demo) | Degrades on highly compressed images |
| Illuminarty | Highlights AI-generated regions within the image | Yes | May miss newer generator styles |
| Content at Scale AI Image Detector | AI-generation probability score | Yes | Training data cut-off matters |
All AI-image detectors are trained on images from known generators; models released after a detector's training cut-off may evade detection. Social media compression further degrades accuracy. Treat any score as one signal, not a verdict.
Tool 4: Reverse image search to find the original
Google Lens, TinEye, and Yandex can reveal whether an image appeared in a different context or predates the event it is claimed to document. For AI-generated images, you would not expect an earlier appearance; but a context mismatch (for example, the image surfacing as fan art before it appears as a news photograph) is a revealing sign. For a full walkthrough of reverse image searching, see our guide on how to reverse image search a social media photo.
C2PA: the emerging provenance standard
The Coalition for Content Provenance and Authenticity (C2PA) is developing an open standard for embedding cryptographic provenance into images at the point of capture. Adobe, Nikon, Leica, and other manufacturers have begun shipping C2PA-compliant cameras and software under the Adobe Content Authenticity Initiative umbrella. Photographs carrying a valid C2PA manifest can be verified at contentcredentials.org/verify. However, adoption is not yet universal: most smartphone photos and all AI-generated images currently lack C2PA data, so its absence is not meaningful. Only its presence is informative.
The limits of photo-fakery detection
No single method is conclusive. ELA is undermined by re-compression; EXIF is routinely stripped; AI detectors miss newer generator styles; and reverse image search finds nothing when a fabricated image is genuinely new. Combine at least three methods, and consider the source and incentive: who benefits from this image being believed? If an image underpins a high-stakes claim (legal, medical, political), submit it to a professional fact-checker rather than relying solely on automated tools.
Frequently asked questions
Can you spot an AI-generated photo just by looking at it?
Sometimes. Look for wrong finger counts, garbled text on signs or clothing, asymmetric jewellery, and backgrounds that blur or melt unnaturally at the edges. Newer AI generators close many of these gaps, so visual inspection alone is not reliable. Pair it with ELA and an AI-image detector.
What is error level analysis and how accurate is it?
Error level analysis (ELA) re-saves a JPEG at a standard quality and highlights regions that carry a different compression signature, which indicates they were edited after the original save. It is useful for detecting traditional Photoshop edits but becomes less reliable on images that have been re-compressed multiple times, such as screenshots of screenshots or images saved from social media.
Does missing EXIF data mean a photo is fake?
Not on its own. Most social media platforms strip EXIF metadata on upload, so absent EXIF is common and is not itself proof of fakery. It is a red flag only when combined with other evidence, for instance a claimed press photograph with no EXIF and signs of editing in the ELA output.
Which free tools are best for detecting AI-generated images?
FotoForensics for error level analysis, Hive Moderation and Illuminarty for AI-generation probability, and ExifTool for metadata inspection are all free and well-regarded. No single tool is definitive; run at least two and compare results. If both flag the image independently, treat that as a meaningful signal, not proof.
Want a second opinion in seconds? Upload the image to FactHeck. The AI vision analysis examines the image for editing artefacts and AI generation patterns, then fact-checks any claims in the caption.