How to Spot a Fake Photo Before You Send It to the Group Chat

In the days after Hurricane Helene tore through western North Carolina in the fall of 2024, an image went everywhere: a little girl in a life vest, sitting in a boat in the rain, crying, holding a puppy. It was devastating. It was also generated by a computer. Plenty of the people who shared it were good-hearted folks who wanted the country to pay attention to a real disaster with real victims. They handed the other side an easy win instead, because once one image in a story turns out to be fake, every image in the story gets an asterisk.

That's the actual cost of sharing a fake. Not embarrassment. Credibility. You spend it every time, and you don't get it back cheap.

So build the habit. Most of what follows takes under a minute.

The tell is in your chest, not on the screen

Before any technical check, notice what the image did to you. Fakes are tuned for one of two reactions: a surge of outrage, or a surge of tenderness. Both shut down the part of your brain that asks questions.

If a photo makes you want to share it right now, that urgency is the product. Sit on it for sixty seconds. Almost everything true will still be true in sixty seconds.

The most-shared fake of the AI era so far was Pope Francis in a white designer puffer coat, which went around the world in March 2023. Nobody was angry about it. Everybody just thought it was great. That's exactly why it worked.

Reverse image search, and what defeats it

Save the image. Drop it into Google Lens, TinEye, or Bing Visual Search. What you're looking for is the oldest appearance of that image on the internet.

A photo captioned as being from a protest last Tuesday that shows up on a Brazilian news site in 2017 isn't from a protest last Tuesday. That's the single highest-yield check there is, and it catches most recycled-photo hoaxes on the first try.

What defeats it: cropping, mirroring, and heavy filters. If your first search comes up empty, crop the image down to the most distinctive object in it and search again. Then flip it horizontally and search a third time. TinEye is better than Google at finding exact copies; Google Lens is better at finding similar scenes. Use both.

Empty results don't mean the image is real. They mean it's new, which for a purported news photo is either a scoop or a fabrication.

The visual tells are getting worse, not better

For a couple of years, hands were the giveaway. Six fingers, a thumb bending the wrong way. Don't lean on that anymore. The current generation of image tools mostly gets hands right.

What still breaks:

• Text. Signs, jerseys, license plates, storefront lettering. Generated text tends to be almost-words that dissolve when you zoom in.
• Repeats. Look at crowds and foliage. The same face three times in a row, the same parked car twice, a fence post rhythm that hiccups.
• Physics. Shadows pointing in two directions. Reflections that don't match what's in front of the glass. Rain that falls on some people and not others.
• Jewelry and teeth. Earrings that don't match, glasses arms that vanish into hair, too many teeth.

Zoom to 300 percent and look at edges where one object meets another. That's where the seams are.

Video takes two extra steps

Screenshot three frames from the video, ideally frames with distinct backgrounds, and reverse-search each one. Old footage relabeled as new is far more common online than any deepfake, and frame-searching catches it.

Then check the speed. The 2019 video of Nancy Pelosi that made her appear slurred hadn't been synthesized at all. It had been slowed down. That category, often called a cheapfake, remains the workhorse of political video manipulation because it requires no skill and no software budget. Slowed audio has a distinctive muddy bottom end. Sped-up video makes people's small movements look twitchy.

For actual synthetic video, watch the mouth on hard consonants, the B and P and M sounds, where the lips have to fully close. Watch the neck and the jawline for a faint shimmer. Watch whether the person ever blinks awkwardly or turns far enough to show an ear.

And listen with headphones. Cloned voices tend to be too clean. No room tone, no breath before a long sentence, no scuffed chair.

Check provenance, but don't trust its absence

Cameras and editing tools from a growing number of manufacturers now attach Content Credentials, the tamper-evident provenance standard developed through the C2PA coalition that Adobe, Microsoft and the BBC helped start. When it's present, it tells you what device captured the file and what was done to it afterward. Google Search offers an "About this image" panel that shows where else an image has appeared and when it was first indexed. Google also watermarks its own generated images with SynthID.

The catch, and it's a big one: any screenshot strips all of it. Someone photographs a real event, posts it, another account screenshots the post, and the credential is gone. No credential means no information. It doesn't mean fake, and anyone telling you otherwise is selling something.

The same goes for the browser extensions and websites that claim to score an image's probability of being AI-generated. Treat them as one weak signal among several. They produce false positives on heavily compressed real photos all the time.

Check the account, then check the event

Click the profile that posted it. How old is the account? What did it post before this? An account created six weeks ago that posts nothing but content on one geopolitical conflict isn't a citizen journalist.

Then do the cheapest check of all, the one people skip: search the claim, not the image. In May 2023, a fabricated photo of black smoke near the Pentagon circulated widely enough that the S&P 500 briefly dipped before the Arlington County Fire Department confirmed no such explosion had occurred. Anyone who took ten seconds to search "Pentagon explosion" would have found local fire officials saying the opposite.

Two independent outlets with names you recognize and a correction policy, or wait. That's the bar. A screenshot of a headline isn't a headline; open the outlet's own site and confirm the story exists there.

When you get it wrong

You will, eventually. Everybody does.

Delete it, then post a separate note saying what you shared, that it was false, and what the actual facts are. Don't quietly remove it and hope nobody noticed. The correction is what buys the credibility back, and it buys it back almost in full, which is a better deal than most mistakes offer.

Teach the kids the reverse-image trick on something harmless. Show them a photo of a "rescued" animal from a viral post, run it through Lens together, and watch them find it on a 2014 blog from another continent. Ten-year-olds love catching a liar. Give them the tools while the stakes are still a puppy.

The people manufacturing this stuff are counting on you being generous, angry, or in a hurry. Be the first two. Never the third.