How to spot AI-generated content and images in 2026
A lot of the old advice for catching AI images no longer works. Here is what actually holds up in 2026, what to stop trusting, and the one habit that beats every trick.
The reliable ways to check an image are content credentials (C2PA), reverse image search, and a few visual artifacts like garbled text and wrong shadows. For text, do not trust AI detectors, they are unreliable and unfair to accuse students with. The move that beats all of them is verifying who made it: a named, accountable source.
How to spot an AI-generated image
1. Check for content credentials
The strongest signal, when it exists, is provenance. Content Credentials, built on the C2PA standard and backed by Adobe, Microsoft, Google, OpenAI, and camera makers like Sony and Nikon, attach a cryptographic record of how an image was made and edited. You can inspect an image at contentcredentials.org. One honest caveat: most images online carry no credential yet, so a missing one is not proof of anything.
2. Run a reverse image search
Drop the image into Google Lens or TinEye. This finds where it first appeared, catches a real photo with AI elements pasted in, and shows whether the event it claims to show is being reported by trusted outlets. It is the single most useful thirty seconds you can spend.
3. Look for the artifacts that still give it away
Modern models are good, but a few tells survive:
- Garbled text. Signs, labels, license plates, and book spines often come out as nonsense letters.
- Inconsistent light. Shadows that fall in different directions, or reflections that do not match what should be reflected.
- Drifting patterns. Brick, tile, fabric weave, and crowds that repeat and warp when you look closely.
What no longer works
Plenty of popular advice is now out of date. Do not rely on these:
Counting fingers is dead as a test. Today's top models render hands correctly most of the time. Hands, ears, and teeth can still glitch in complex poses, but treat that as a weak hint, not proof.
- Plastic or too-smooth skin and perfectly symmetric eye highlights no longer flag AI. Models handle these now.
- Missing metadata (EXIF) is a weak hint at best. Social platforms strip metadata from real photos too.
- The uncanny-valley feeling is unreliable. Some AI images look completely natural, and some real photos look strange.
How to spot AI-generated text
Do not use AI text detectors to judge or accuse people. Independent testing shows their accuracy collapses on edited or paraphrased writing, and their false-positive rates are especially high for non-native English writers. That is a fairness problem, not a small bug.
Instead of trusting a detector score, read for a cluster of signals and then check the facts:
- Generic, hedged tone with no specific detail, names, or firsthand observation.
- Fabricated citations. AI often invents sources and quotes that look real but do not exist. Click through and confirm they are real.
- No traceable sourcing. Real reporting tells you how it knows what it knows.
As models improve, reliable text detection may stop working altogether. So the goal is not to "detect" AI. It is to verify the sources and claims, the same skill that catches any unreliable writing. Our step-by-step guide to spotting fake news covers that method.
The one habit that beats every trick
Tools change every month. The durable skill does not: trace anything you are unsure about back to a named, accountable source. In a feed full of generated images and text, the strongest signal is whether a real person or a real newsroom stands behind it. Full disclosure: that is the bet we made building Newsreel Classroom, where every story carries the name of the journalist who reported it. Students should still question what they read there. The difference is that there is a person to question.
Give students real journalism to anchor on
Newsreel Classroom brings each day's news, written by vetted journalists, at three reading levels with a quiz and poll built in. Book a demo to see it in your classroom.
Explore Newsreel ClassroomFrequently asked questions
Can you tell if an image is AI-generated?
Sometimes, but not always. The most reliable signals are content credentials (C2PA), reverse image search to find the original, and visual artifacts like garbled text or inconsistent shadows. There is no single test that works every time, so the safest move is to verify the source rather than trust the pixels.
Are AI content detectors accurate?
AI text detectors are unreliable and should not be used to accuse students. Independent testing shows accuracy dropping sharply on edited or paraphrased text, with false-positive rates that are especially high for non-native English writers. Verify the sources and claims in a piece of writing instead of trusting a detector score.
What are Content Credentials (C2PA)?
Content Credentials are a cryptographic label built on the C2PA standard, backed by Adobe, Microsoft, Google, OpenAI, and camera makers. They record how an image or video was created and edited. You can inspect them at contentcredentials.org, though many images carry no credential at all.
Does counting fingers still work to spot AI images?
No, not reliably. Today's image models render hands correctly most of the time. Hands, ears, and teeth can still glitch in complex poses, but treat that as a weak hint, not a test. Provenance, reverse image search, and source-checking are far more dependable.