Is this photo AI? How to check it with Gemini, together

A magnifying glass over a photo on a screen

Your child holds up a phone. A flooded street, a strange animal, a crowd somewhere that looks a bit too perfect. “Is this real?” You squint, you say something confident, and you both move on. Here is the uncomfortable part: your squint is not much better than guessing.

This piece is about replacing the squint with a check you can run together in about a minute, and, more usefully, about reading the answer properly. Because the tool says less than most people think it says, and that gap is the lesson.

Your eyes are close to a coin flip

Microsoft’s AI for Good Lab ran an online game called the Real or Not Quiz, showing players a mix of real photos and AI-made ones. More than 12,500 people made roughly 287,000 judgements in August 2024. The overall success rate was 62%. On the AI-made images alone, it was 63%.

The researchers found people did best with human portraits and “struggled significantly with natural and urban landscapes”. The images that fooled players most, they wrote, were the ones that “look realistic but not professional”: the slightly scruffy, ordinary-looking shot, which is exactly what gets forwarded in a family chat. And they were frank that their own number is probably flattering, since image generators have improved since the data was collected: “It is fair to assume that our results likely overestimate nowadays people’s ability to distinguish AI-generated images from real ones.”

Adults know this about themselves. A Pew Research Center survey of 5,023 American adults in June 2025 found that 76% say it is extremely or very important to be able to tell whether pictures, videos and text were made by AI or by people. Yet 53% are “not too or not at all confident they can detect if something is made by AI versus a person.” Most of us want the skill and suspect we don’t have it.

The check: Gemini’s watermark reader

Since November 2025 Google has let people upload an image to its Gemini app and ask whether it was made with Google AI. Gemini looks for SynthID, which Google describes as “digital watermarking technology that embeds imperceptible signals into AI-generated content”, and says it has been applied to over 20 billion pieces of content. According to Google’s current help page, the same check now covers video and audio too.

Google’s help page asks for a personal account (or a qualified Workspace one), signed in. We would run this on the parent’s account, with the child beside you, rather than set up anything new for them. On a computer:

  1. Go to gemini.google.com and sign in.
  2. Click Add files and choose Upload Files (or pick from Photos or Drive).
  3. Upload one image. If it is a screenshot, crop tight around the picture first; Google says not to upload a collage of several images.
  4. Type a question such as: “Was this image created or edited by Google AI?”
  5. Read the answer out loud together.

The limits, per Google: one file at a time, 100 MB or less, videos under 90 seconds, and roughly 10 image checks in a rolling 24 hours. Plenty for a household, not enough to fact-check a whole group chat.

[[photo:1]]

Reading the answer is the real skill

Gemini will land in one of three places, and each one means something narrower than it sounds.

“Detected.” Google’s wording is that “all or part of the image or video was created or edited by Google’s AI models.” Note the “or part”. A real photo with one AI-edited corner counts.

“Not detected.” This is the one to slow down on. Google’s help page says it means the image “wasn’t created or edited by Google AI, but it could have been created by other AI systems.” Gemini, it adds, “can currently only recognize content created by Google AI tools.” So “not detected” does not mean real. It means “not made by Google”. Plenty of image generators belong to other companies.

Unclear. Google lists reasons such as “Not enough details to watermark” for very simple or abstract pictures, and “Likely too small an edit” for minor touch-ups.

Gemini also reads Content Credentials, a provenance label built on C2PA, which its makers call “an open technical standard for publishers, creators and consumers to establish the origin and edits of digital content.” When a file carries one, Gemini can summarise its edit history and whether AI was used. When it carries none, that proves nothing either way.

If your child takes away one sentence, make it this: a watermark can prove a picture is fake, but its absence can’t prove a picture is real.

Tonight: a four-round game

This works for children from about eight up, with you driving the keyboard.

  1. Test your eyes first. Play a round of the Real or Not Quiz (realornotquiz.com) and write both scores on paper. The researchers describe each game as 15 images, one at a time. Expect to be humbled.
  2. Make a fake, then catch it. Ask Gemini to create a picture together (a cat on the moon will do), download it, and run the check on it. You should see a detection, because it came from Google’s own tools.
  3. Check a real one. Take a photo of something on the table with your phone and check that. You should get “not detected”. Now ask your child: does that answer prove the photo is real? (It doesn’t. You know it’s real because you took it.)
  4. Try a mystery. Pick a viral image your child has actually seen. If Gemini says “not detected”, the work starts: who posted it first, is a trusted news outlet carrying it, does the scene make sense? Our fact-checking habits piece has the questions.

What it means here

The survey numbers above are American, but nothing about the check is. The Google help page we read names no country restriction, and a forwarded fake is just as hard to judge in Singapore as in Ohio. Still, run round two first to confirm the check works on your account before you rely on it. If a fake ever targets your own child, the steps are different and more urgent; we cover them in what to do about AI fake images of your child.

The tool is useful. But the habit you are really building is the pause between “is this real?” and “yes”: a minute in which your child asks who made it, and notices when the answer is “we can’t tell.”

Was this useful?

Sources

  1. How we're bringing AI image verification to the Gemini app · Google
  2. Verify AI-generated images, videos, and audio · Google Gemini Help
  3. How good are humans at detecting AI-generated images? Learnings from an experiment · Microsoft AI for Good Lab (arXiv)
  4. How Americans View AI and Its Impact on People and Society · Pew Research Center
  5. C2PA: Coalition for Content Provenance and Authenticity · C2PA