Tips & Tells: How to Spot AI

Start here: the 30-second checklist

Before you zoom in on pixels, run these four checks. They catch more fakes than any single visual tell.

1. Source. Who posted it first? A named person or outlet you can check, or an anonymous repost?
2. Receipts. Does the file carry Content Credentials or useful metadata? Check it with our provenance checker.
3. Details. Zoom in on hands, text, shadows and edges. For text, look for vague, padded writing with no specifics.
4. Corroboration. Would something this big be reported by others? Search for it before you share or act.

If a post makes you angry, scared or urgent, that is the moment to slow down. Fakes are built to be shared before anyone checks.

Why no detector is 100% sure

AI detectors look for statistical patterns. Those patterns overlap with how real people write and how real cameras and apps process images, so every detector makes two kinds of mistakes: it misses some AI content, and it flags some human work (a false positive).

• New models change the patterns. A detector trained on older AI output can miss newer tools.
• Editing blurs the signal. Cropping, compressing, screenshotting or lightly rewriting can move a score a lot.
• Labels get stripped. Many sites remove metadata and Content Credentials when you upload, so "no label" is not evidence either way.
• Labels can be faked. Plain metadata can be typed in by anyone. Signed Content Credentials are much harder to forge, but they only say which tool a file passed through.

That's why our own AI text detector shows an estimate and highlights, never a verdict. For the evidence, see Are AI detectors accurate?

Images

Generators have improved a lot, so the old "count the fingers" trick works less often. Look for several small problems, not one.

• Hands, teeth, ears and jewelry. Extra or fused fingers, too many teeth, mismatched earrings, glasses that melt into skin.
• Text in the scene. Signs, labels and logos with garbled letters, or brand marks in the wrong place.
• Light and shadow. Shadows that point different ways, or reflections that don't match what is in front of the mirror or water.
• Too-perfect skin. A smooth, plastic look with no pores, or hair that blends into the background.
• Backgrounds. Objects that merge, repeat or trail off into nonsense, and crowds of faces that look alike.
• Watermarks. A generator badge or logo in a corner, sometimes partly cropped out.
• No history. A reverse image search finds no earlier copy and no named photographer.

The FBI lists many of the same clues, including distorted hands, unrealistic teeth or eyes, odd accessories and inaccurate shadows (FBI Internet Crime Complaint Center). For a deeper walkthrough, read how to tell if an image is AI-generated.

Video

• Physics. Water, hair, fabric and crowds that move in strange or perfectly synced ways, or objects that pass through each other.
• Continuity. Things that appear, vanish or change between frames: a ring on one hand, then the other.
• Faces and mouths. Lip movements that drift from the words, unnatural blinking, or a face that flickers at the edges when the head turns.
• Short clips. Many generated videos are only a few seconds long or are stitched from short shots.
• The news test. A dramatic event that no reliable outlet has covered is a red flag on its own.

To check a clip, grab a few still frames and reverse-search them. The InVID-WeVerify plugin listed below does this for you.

Text

AI writing often reads smooth but empty. Common patterns include:

• An even, polite tone with no personal detail or first-hand experience.
• Stock phrases and inflated words ("plays a pivotal role," "a rich tapestry"), plus tidy lists of exactly three.
• Heavy use of em dashes and "not just X, but Y" sentences.
• Confident facts, quotes or citations that turn out not to exist.

Wikipedia's editors keep a detailed public list of these patterns (Wikipedia editors). Our AI Words Finder highlights many of them, and our guide on why detectors flag human writing covers what to do if you're wrongly accused.

Voice-clone scams

A scammer can clone a voice from a short clip posted online, then call pretending to be a family member, a boss or a bank. The FTC and FBI both warn about this (Federal Trade Commission; FBI Internet Crime Complaint Center).

Red flags

• An urgent emergency: an accident, arrest or kidnapping, and money needed right now.
• Pressure to keep it secret or not to call anyone else.
• Payment by wire, gift cards, cryptocurrency or a payment app.
• A voice that sounds right but oddly flat, with strange pauses or breathing.

What to do

1. Hang up and call back on a number you already have for that person, not the one that called you.
2. Use a family code word that only your family knows, and ask for it.
3. Check with someone else in the family if you can't reach them.
4. Report it at ReportFraud.ftc.gov or the FBI's ic3.gov.

Free tools to check yourself

All free to use. Outside sites have their own privacy terms, so check them before uploading anything personal.

• Our provenance & metadata checkerReads Content Credentials and photo metadata in your browser. Nothing is uploaded.
• Content Credentials VerifyThe Content Authenticity Initiative's official viewer for Content Credentials.
• Google LensReverse image search to find earlier copies of a picture.
• Bing Visual SearchAnother reverse image search, which often finds different matches.
• TinEyeReverse image search that can sort results by oldest first.
• InVID-WeVerify pluginA free browser plugin used by fact-checkers to split videos into keyframes and search them.
• FotoForensicsShows metadata and error level analysis for a photo. Note: this site uploads your image to its server.
• ExifToolA free desktop program that reads almost every kind of metadata.
• SynthID (Google DeepMind)Explains Google's watermark and links to Google's own detector, which has limited access.

Practice: spot the tell

Six quick scenarios. Pick the strongest clue, then see why. Nothing is saved.

Glossary

AI detector

Software that estimates how likely a text, image or clip was made by AI. It gives a probability, not a verdict.

C2PA

The Coalition for Content Provenance and Authenticity, an industry group that publishes the open standard behind Content Credentials.

Content Credentials

A signed record attached to a file that lists which tools created or edited it. Think of it as a label, which can be missing or removed.

Deepfake

A video, image or audio clip that uses AI to make someone appear to say or do something they did not.

Digital source type

A standard label (from IPTC) that says how media was made, for example "digital capture" for a camera photo or "trained algorithmic media" for generative AI.

EXIF / XMP metadata

Hidden fields inside a photo, such as camera model, date and editing software. Easy to edit or strip.

False positive

When a detector says something is AI but a person actually made it.

Provenance

The history of where a piece of content came from and how it was changed.

Reverse image search

Searching with a picture instead of words to find where else it appears online and who posted it first.

SynthID

Google DeepMind's invisible watermark for content made with Google's AI tools. It can only be checked with Google's own tools.

Voice clone

A synthetic copy of someone's voice made from a short recording.

Watermark

A visible logo or an invisible signal added to AI output so it can be identified later. Not every AI tool adds one, and some can be removed.