Is This Image AI? 7 Checks That Work, and 3 That Don’t

In May, NewsGuard took 15 authentic news photos, the kind shot by Reuters and AP photographers, and ran them through five popular AI image detectors. Collectively, the tools called real photos fake 13.33 percent of the time. One of them, ScamAI, flagged six of the 15.

That is the problem with the question in the headline. Most people answer it with a free detector site or a squint at the fingers, and both methods fail in ways that make you more confident and less right. What follows is for anyone who shares, reports on or simply wants to stop being fooled by images online: seven checks that hold up, three that don’t, and the limit of each.

How we chose: every check below is free, works on a phone or a laptop, and is backed by the people who build the tools or by published testing. None of them is a verdict machine. Used together, they get you to an answer you can defend.

1. Content Credentials: read the label, if there is one

Content Credentials are the consumer face of C2PA, a provenance standard backed by more than 500 companies including Adobe, Google, Meta, Microsoft and the BBC. When a camera or app attaches one, it records how the file was made and edited. To check, drop the file into the Verify tool at contentcredentials.org/verify. It accepts JPEG, PNG, WebP and TIFF images as well as video, audio and PDFs, and shows the signer, the app or device used and the edit history, according to its documentation.

The catch is the empty result. Verify will say “No Content Credential” for most images you find, because most cameras and apps still don’t attach one and many platforms strip metadata on upload (see No. 5). A credential that says “made with an AI tool” is strong evidence. No credential is no evidence at all.

2. SynthID: ask Google about Google’s own watermark

SynthID is Google DeepMind’s invisible watermark, embedded at generation time and designed to survive cropping, filters and lossy compression. In May 2026 Google said it had watermarked more than 100 billion images and videos, and that OpenAI, Kakao, ElevenLabs and Nvidia’s Cosmos video models are adopting it. To check, upload the image to the free Gemini app and ask whether it was made with AI; Google says the same check is rolling out in Search through Lens and Circle to Search, with Chrome to follow.

The limit: it can only find watermarks from companies that use SynthID. When The Register tested the Gemini check on ChatGPT-made images in late 2025, before OpenAI signed on, it could not reliably tell, and the same report cites University of Waterloo researchers who built a method for stripping such watermarks. A positive hit means something. A negative one means only that this particular watermark isn’t there. Google’s dedicated SynthID Detector portal launched in 2025 as a waitlist tool for journalists and researchers, so for most people the Gemini app is the way in.

3. Reverse image search, on two engines

Run the image through Google Lens and TinEye. They index different parts of the web, and an image that appears nowhere else before the viral post is a different animal from one that has circulated since 2019 with a different caption. On Google, the About this image panel, reachable from the three-dot menu in Google Images or from a Lens search, shows where else the picture has appeared, including fact-checking sites.

This catches the most common fake of all: a real photo with a false story attached. It won’t help with a freshly generated image that has no history yet, which is itself a signal worth noting.

4. Find the earliest upload

Provenance is mostly a question of when. Google’s About this image reports when the image and similar images were first indexed. TinEye lets you sort results by oldest, though the date is when its crawler found the copy, not when the photo was taken.

Then go to that earliest copy and read around it. Who posted first, with what caption, and did a news agency or a named photographer ever claim it? An image of a breaking event whose first appearance is an anonymous account with no other photos from the scene should worry you more than any pixel-level artifact.

5. Read the metadata, and don’t trust its absence

If you have the original file, an EXIF viewer will show the camera model, lens, timestamp and sometimes GPS coordinates. A phone photo of a protest that carries real camera data is harder to fake than one that carries none. Most of the time, though, you won’t have the original. In one test, Instagram, Facebook and X all stripped GPS, camera and timestamp data from uploaded photos.

Metadata can also be edited with free tools. Treat it as one more witness, never the judge.

6. Look at who posted it

Open the source account and scroll. How old is it? Has it posted original photos before, from the same city, with the same camera look? Did it go from cooking videos to war footage last week? Accounts built to farm engagement tend to post a flood of dramatic images from wildly different places, none of them credited.

This is slow, unglamorous work. It is also what professional verification desks spend most of their time doing, because people leave a longer trail than pixels do.

7. Check the physics: shadows, reflections, lines

UC Berkeley’s Hany Farid, one of the founders of digital image forensics, has a method anyone with a straightedge can try. For shadows, pick a point on a shadow and the matching point on the object casting it, draw a line through both, and repeat. In an authentic sunlit image, the lines meet at one point, the light source. Reflections and the parallel lines of floor tiles or rooftops work the same way. Farid’s latest work argues that generators remain “fundamentally ignorant of how light and geometry work in the real world“, as TechSpot reported in May.

The limit: it needs a scene with clear shadows or straight lines, lens distortion can bend the result, and nearly parallel lines make the intersection unreliable. Farid himself has cautioned that generators may get better at this.

8. Doesn’t work: free “AI detector” sites

They are fast and confident, which is the danger. In NewsGuard’s audit, the five tools disagreed with one another on 35 of 45 images, and two of the five had no false positives while ScamAI flagged 40 percent of authentic photos as AI. NewsGuard also documented the worse outcome: people citing a detector score to dismiss a real image as fake. If you use one, treat its number as a weak hint and never as the reason you believe or disbelieve something.

9. Doesn’t work anymore: counting fingers

Six-fingered hands were the meme of 2022. By March 2023, Hyperallergic was reporting that Midjourney Version 5 had rendered the hand discourse largely obsolete, and current models do far better still. A mangled hand is still a clue when you see one. A normal hand proves nothing, and checking only the hands trains you to look in the wrong place.

10. Doesn’t work: “it looks too perfect”

Intuition is worse at this than people assume. In a 2022 PNAS study by Sophie Nightingale and Hany Farid, 315 participants identified AI-generated faces with 48 percent accuracy, below a coin flip, and rated synthetic faces 7.7 percent more trustworthy than real ones. Training lifted accuracy only to 59 percent. Those faces came from an older generation of models. Gut feel has not improved since; the generators have.

Which brings it back to NewsGuard’s 15 real photos. The checks that would have cleared them in minutes, a reverse search turning up the Reuters original and an earliest upload with a photographer’s name on it, cost nothing and never once guess.

Sources

• NewsGuard: Leading AI image detection tools mislead online users
• Content Credentials: Verify
• Content Authenticity Initiative docs: Using the Verify tool
• Content Credentials (C2PA) home
• Google: Making it easier to understand how content was created and edited (May 2026)
• Google DeepMind: SynthID
• Google: SynthID Detector announcement
• The Register: Gemini tries to sniff out AI slop images
• Google: How to use About this image
• TinEye: Introducing sort by date
• EXIFData.org: Do social media sites strip EXIF data?
• Content Authenticity Initiative (Hany Farid): Photo forensics from lighting, shadows and reflections
• TechSpot: AI images are getting harder to spot, but physics still gives them away
• Hyperallergic: AI image generators finally figured out hands
• EurekAlert: AI-generated faces are more trustworthy than real faces (PNAS, 2022)