A step-by-step guide to telling whether an image was made with AI, from checking where it came from and its Content Credentials to the anatomical, text, lighting and physics tells that still give AI images away.
No single clue proves an image is AI-generated. Stack several signals and give the most weight to where the image came from.
Start with the source: find the earliest place the image appeared with a reverse image search, and check who posted it and whether they say it's AI.
Check for Content Credentials, the C2PA standard many AI tools and some cameras attach, but don't read anything into missing data, since screenshots and most social platforms strip it.
Hands and short text are no longer reliable tells. Lighting, shadows, reflections and long or dense text still trip up AI images more often.
People are close to chance at spotting AI faces, and automated detectors misfire on real photos and break on compressed ones, so treat both as a hint, not a verdict.
The obvious mistakes are disappearing. In a 2022 study published in PNAS, Sophie Nightingale and Hany Farid found people identified AI-generated faces correctly only 48% of the time, about what you'd get by guessing, and rated the synthetic faces as more trustworthy than real ones. A larger 2025 study by Microsoft's AI for Good lab, based on about 287,000 judgments from more than 12,500 people, put overall accuracy at 62%, and its authors said that probably overstates how well people do against today's generators.
That doesn't mean you can't tell. It means a quick glance isn't enough, and the best checks are about the image's origin as much as its pixels.
Before you study the pixels, find the source. Upload the image to a reverse image search engine to find its earliest appearance and original caption. An image that first appeared on an AI art account, or that has no history before a viral post, deserves more suspicion than one traced to a news agency or the photographer who took it.
Then look at who posted it. Check whether the account discloses AI use, whether the caption makes claims the image can't support, and whether reputable outlets have reported the same scene. If an image is built to make you angry or amazed, slow down before you share it.
Content Credentials are a signed record of how a file was made, based on an open standard called C2PA. Many AI image generators now attach them, and some cameras and phones sign photos at the moment they're taken. You can inspect a file's credentials with a free C2PA verification tool, and some sites show a small "CR" icon on images that carry them.
Image metadata can also carry a "digital source type" field from the IPTC photo metadata standard. A value of "trainedAlgorithmicMedia" means the image was created with generative AI, and "compositeWithTrainedAlgorithmicMedia" means AI was used to edit it.
Two cautions. A valid credential shows who signed a file and how it was made, not that the scene is true. And missing data proves nothing: screenshots remove it, and most social platforms strip it on upload, so always check the original file if you can.
Some AI companies also hide an invisible watermark in the pixels that survives cropping and screenshots better than metadata does. Each company's checker only reads its own watermark, though, so a negative result means only that the image didn't come from that company's tools.
Researchers at Northwestern University sorted the mistakes AI images make into five groups, starting with anatomy. Look for:
Hands gripping objects, where fingers merge with the object or each other, even though a simple open hand is now usually right
Teeth that overlap, run together or look too uniform
Eyes with glassy surfaces or pupils that aren't round
Necks, limbs or bodies that blend into other people, clothing or furniture
Real bodies vary, so one odd detail proves little. Several together are a stronger signal.
AI portraits often have a polished, cinematic look: waxy or overly smooth skin, oversaturated color, perfectly windswept hair and lighting on the face that doesn't match the background. Backgrounds can look smudged or stitched together. Heavily edited real photos share some of these traits, so treat style as a weak clue. The Northwestern research also found that images without the stock "AI look" fooled people more.
Short words are often correct now, but long or dense text still breaks down. Read multi-word signs, labels, book spines, menus and logos closely for garbled letters or words that don't exist. Then check objects that have to work a certain way: watch faces, buttons, buckles, zippers, eyeglasses, bag straps and ties often merge into clothing or don't make mechanical sense.
This is where AI images still slip most often. Generators produce lighting that looks plausible without being consistent.
Shadows: outdoors in sunlight, shadows from different objects should point the same way. Shadows at different angles, or missing shadows, are a warning sign.
Reflections: a mirror, window or puddle should show what's in front of it, matching the scene. Check that a reflection has the same clothing, pose and objects.
Perspective: parallel lines, such as building edges, rails and tiles, should meet at consistent points. Warped edges and stairs that lead nowhere are common in AI images.
Gravity and contact: feet should touch the ground, and objects should rest on surfaces rather than float.
Group photos and busy scenes give AI more chances to slip, so they're often easier to catch than a single close-up face.
Add up what you've found rather than relying on one clue. A traced source, intact camera credentials and consistent lighting point to a real photo. No history, an anonymous source and several physical errors point to AI. When the signals conflict, say you can't tell, since a wrong confident call spreads as fast as a fake.
Use them as a hint at most. When The New York Times tested more than a dozen detectors and chatbots in 2026, most missed altered images, and the Times concluded they're useful for backing up a suspicion but can't give a definitive ruling. Academic tests have found that ordinary JPEG compression, resizing and the processing social platforms apply can push some detectors' accuracy down toward chance. Detectors also flag some real photos as AI, which can cause harm of its own.
Sometimes. Major social platforms read C2PA and IPTC signals to label AI content, and some ask creators to disclose realistic AI images and video, but labels are often missing, especially after an image has been downloaded and reposted. In the EU, the AI Act's transparency rules require deployers to disclose deepfakes from August 2, 2026, and require AI providers to mark generated content in a machine-readable way, with a grace period to December 2, 2026 for systems already on the market under the EU's Digital Omnibus changes.
Start with the source. Run a reverse image search to find where the image first appeared, check whether the account that posted it discloses AI use, and look for Content Credentials on the original file.
Less often than before. Most generators now draw a simple hand correctly, though hands gripping objects still go wrong. Lighting, shadows, reflections and long text are more reliable clues in 2026.
A signed record of how an image was made and edited, based on the open C2PA standard. Many AI tools attach them, and some cameras and phones sign photos when they're taken, but screenshots and most social platforms remove them.
Not reliably. Tests have found they miss altered images, flag some real photos as AI and lose accuracy when an image is compressed or resized.
Generally no, but it depends on the image. In the US, publishing nonconsensual intimate images, including AI-generated ones, is a federal crime under the TAKE IT DOWN Act, and in the EU deployers must disclose deepfakes.
