How to spot AI-generated content

You scroll past a shocking photo. A grandmother lifting a car. A celebrity saying something outrageous. A "news" story about a local event you cannot verify. AI now makes images, videos, and articles good enough to fool almost anyone in a quick glance. Here is how to tell.

Spotting AI images: visual tells

Hands and fingers

Still the most common giveaway. Count fingers. Six fingers, four fingers, fused fingers, missing thumbs. Newer AI is better but hands remain weird often enough to be the first place to look.

Text in the image

Signs, books, t-shirts, license plates. AI often generates text that looks like letters but does not actually spell anything. "FRUSH" instead of "FRESH." Gibberish in the background.

Background weirdness

Zoom into the background. Buildings that bend impossibly. Crowds where everyone has a weird face. Trees that merge into each other. AI focuses on the foreground; backgrounds often fall apart.

Jewelry, glasses, earrings

Necklaces that disappear into shirts. Earrings that fade halfway. Glasses with one lens different from the other. Watches with no hands or wrong numbers.

Shadows and reflections

Real shadows fall in one consistent direction. AI shadows often come from different directions in the same image. Reflections in mirrors and water sometimes do not match the scene.

Skin and hair

Skin too smooth, almost plastic. No pores, no minor blemishes. Hair that fades unnaturally. Wisps of hair where a hairstyle should end. Symmetry that is too perfect.

Teeth

AI sometimes generates one or two teeth too many. Adult teeth with baby teeth mixed in. Teeth that do not align.

Crowd scenes

Look at the people in the back. Are they wearing the same clothes? Do their faces blur into each other? Are there extra limbs?

Spotting AI videos: more tells

• Blinking patterns: AI faces blink too rarely or too regularly. Real humans blink at slightly varied intervals.
• Lip sync: mouth movement does not perfectly match speech. The bottom lip behaves oddly.
• Skin tone shifts: face color changes slightly as the head moves.
• Edges of the face: hairline, ears, jawline can wobble or flicker frame to frame.
• Background does not move correctly when the person moves their head.
• One-take feel: the video never cuts. Real videos usually have edits.
• The mouth opens too wide or too small for the sounds being made.

Spotting AI text and articles

• Too smooth, too generic. Reads like a Wikipedia summary. No specific details that prove the writer was actually there.
• "In conclusion" patterns. AI loves wrapping up sections with summary statements ("In conclusion," "Overall," "To sum up"). Real writers do this less often.
• Identical sentence structure repeated across paragraphs.
• Hedge words: "may," "could," "potentially" appearing throughout. Real writers commit to claims.
• No links to specific sources that you can verify.
• The author has no traceable identity. No real photo, no LinkedIn, no other articles.
• Strong claims with no specifics. "Studies show..." with no study cited.

Free tools that help

For images

• Google Images reverse search: images.google.com. Click the camera icon. Upload an image or paste a URL. Shows where else the image appears.
• TinEye: tineye.com. Another reverse image search. Good for finding the oldest known version.
• Bing Image Match: Bing's reverse image search is sometimes better than Google's at finding social media reposts.
• "Hive Moderation" AI detector: free at hivemoderation.com. Upload an image, get a percent likelihood of AI. Treat as a hint, not certainty.

For videos

• Reverse search a still frame: pause the video, screenshot, reverse image search.
• Watch in slow motion (YouTube has speed controls). Tells become obvious at half speed.
• InVID (free verification plugin used by journalists): invid-project.eu.

For text

• GPTZero: commercial AI text detector. Useful but flagged for false positives. Treat results with caution.
• Originality.ai: paid, used by publishers. Same caveats.
• The honest answer: there is no reliable AI text detector. Use judgment.

Where AI fakes do the most damage

• Political content: deepfaked photos and videos of politicians shared on social media. Always assume any politically charged image you see on Facebook needs verification.
• Celebrity scams: "Elon Musk endorsed this stock" videos. The endorsement is AI. The stock is a scam.
• Fake news articles: entire blogs of AI-written stories rank in Google, often with affiliate links or scam recommendations.
• Romance scams: AI faces on dating profiles. The person you matched with does not exist.
• Disaster misinformation: fake images of natural disasters shared as if real.

The "too good or too shocking" rule

The biggest tell is emotional. If an image, video, or headline makes you immediately angry, scared, or thrilled, slow down. Scammers and propagandists make AI content that triggers strong emotions because emotional content gets shared fast.

Before you share anything that made you feel strongly:

1. Reverse image search it.
2. Check Snopes or FactCheck.
3. Find a second source (a major news outlet) reporting the same thing.
4. If you cannot verify, do not share.

What about voice and phone calls?

Voice cloning is a different problem. See our full guide on AI voice cloning scams and the family safe word.

Watermarks: helpful but not reliable

Some AI tools embed an invisible watermark in their output (Google's SynthID, Adobe's C2PA metadata). Browsers and platforms are starting to read these. But the watermark gets stripped if the image is screenshotted, cropped, or compressed for social media. Useful as one of several signals; not a final answer.

Quick mental checklist before believing something online

1. Who posted this? Real account or anonymous?
2. Does the image / video appear anywhere except social media?
3. Are major news outlets reporting it?
4. Does it make me feel strong emotion? (If yes, slow down.)
5. Does it benefit someone if I share it?
6. If I cannot answer all of these, I do not share.

If you got fooled, do not feel bad

AI content is now good enough to fool experts. Fact-checking websites have entire teams that do nothing but verify viral content. If you shared something you later learn was fake, delete the post and add a comment saying so. That helps reduce the spread. Then forgive yourself; this is hard now and getting harder.