AI news
AI model reviews
AI tools
How-to guides
Our staffAboutContact

How to tell if a photo is AI generated: 7 steps with a worked example

Check hidden watermarks with Gemini and OpenAI, read Content Credentials, reverse search, then study the pixels. A step-by-step guide.

Magnifying glass on a wooden table
Photo: Hannes Köttner on Unsplash

To tell if a photo is AI generated, check the file for hidden provenance signals first, then look at where it came from, and only then study the pixels for odd hands, garbled text or wrong shadows. Free checkers from Google and OpenAI can now read the invisible watermarks their own image models add, and a Content Credentials viewer reads the "nutrition label" that some apps and cameras attach. None of them can prove a photo is real. Used together, though, they'll settle most cases in about ten minutes.

I'd rank the steps below roughly by how much they're worth. Eyeballing an image for six-fingered hands used to work well; I think it's now the weakest test on the list, because current models rarely make those mistakes.

Before you start

Get the best copy of the image you can. A screenshot of a screenshot strips most of the evidence you're about to look for.

  • The original file if possible (saved from the post, not screenshotted), or a tight crop around the image if a screenshot is all you have.
  • A free Google account, which you'll need for the Gemini check.
  • A browser on a phone or computer.
  • About ten minutes, and a notes app to jot down what each check says.

One more thing: keep the caption and the account name. Context often matters more than any detector result.

Step 1: Look for a label and check who posted it

Start with the cheapest evidence. On Facebook, Instagram and Threads, Meta shows an "AI content" label when its systems find AI signals in a file, and some accounts carry an "AI creator" profile label. Meta's help page on AI labels also admits a weakness: it doesn't require people to label AI images, only realistic video and audio. So a missing label tells you very little.

Next, open the account. Is it a news outlet, a known photographer, or a page that posts a dozen dramatic "photos" a day? That last pattern is a red flag on its own.

Example: a picture of a whale breaching beside the Golden Gate Bridge, posted by a page called "Nature Is Wild" with no photographer credited, deserves more suspicion than the same shot from a local paper's staff photographer.

Step 2: Ask Gemini to check for a SynthID watermark

Google's image models hide an invisible watermark called SynthID inside every picture they make, and Google says it has now marked over 100 billion images and videos this way. The Gemini app can read it.

  1. Go to gemini.google.com or open the Gemini app and sign in.
  2. Tap Add files and upload one image (100 MB or less).
  3. Type a question like the one below.

Is this AI-generated? Check it for a SynthID watermark and for Content Credentials, and tell me what each one shows.

Going by Gemini's verification help page, a detected watermark means all or part of the image was made or edited by Google AI. No watermark means it didn't come from Google's tools, but another AI could still have made it. You'll sometimes get a "not enough details" or "likely too small an edit" result, too. The same check now works in Google Search through Lens, AI Mode and Circle to Search; just ask "Is this AI generated?"

Step 3: Run the file through OpenAI's verify page

This is the "OpenAI detector" people search for. Images made with ChatGPT or the OpenAI API now carry both Content Credentials and a SynthID watermark, and OpenAI's verify page checks for either one in PNG, JPG, WebP and GIF files.

Upload one file and read the result. A hit means an OpenAI tool likely processed the image. It doesn't tell you how much was generated, or who made it. A miss is weaker still, because OpenAI's own provenance help page lists the usual culprits: older images, stripped metadata, and watermarks worn down by cropping, compression or format changes.

Step 4: Read the Content Credentials

Content Credentials (the C2PA standard) are signed metadata that record which app or camera made a file and what edited it. They cover real photos as well as AI ones, which makes them useful both ways.

  1. Open the Content Credentials Verify site.
  2. Drag the image in. It takes JPEG, PNG, WebP, HEIC, TIFF and several other formats.
  3. Read the history: the tool that created it, any edits, and whether AI was used.

Most images you'll check won't have credentials at all (the site itself warns they're "still rolling out"). That's normal. But when they're there, they're the strongest single clue you'll get, and they can point the other way too: Pixel 10 phones add Content Credentials to photos straight from the camera app.

Step 5: Reverse search the image

A reverse search tells you whether the picture existed before the event it supposedly shows. In Google Search, open the image viewer, tap About, then "Learn more about an image". Google's About this image shows when Google first saw similar versions, which other pages use it, and sometimes a "How the image was made or edited" section.

Honestly, this step catches more fakes than people expect, because plenty of "AI images" turn out to be old real photos with a new caption. For face-heavy images, a dedicated face search tool can help; our Lenso AI review covers one, along with its privacy trade-offs.

Step 6: Zoom in on the details

Now look at the pixels. Gemini's own help page lists the tells worth checking:

  • Text on signs, shirts or labels that's garbled or melts into nonsense.
  • Small errors in hands, teeth, jewellery and background objects.
  • Shadows that fall in different directions, or light with no clear source.
  • Repeating patterns, such as identical faces in a crowd or copy-pasted bricks.

Treat these as hints. A blurry phone photo can look "off" for ordinary reasons, and a well-prompted image can look flawless. If you want to see why, our guide to writing AI photo prompts shows how much control people now have over realism.

Step 7: Weigh the evidence and decide

Add up what you found. A watermark or a credential saying "created with AI" is close to conclusive. A clean result from every checker isn't proof of anything, since it only rules out the tools that were checked.

I'd use three verdicts: "confirmed AI", "likely real" (credentials from a camera, or an older original found by reverse search), and "can't tell". That last one is fine. It's more honest than guessing.

Example: a complete worked example

Say a friend sends you a dramatic shot of a flooded city square, captioned "this morning in my town". You'd run it like this.

  1. Label and source: it's a screenshot from a reposting account with no AI label. The original poster isn't credited. Suspicious, but not proof.
  2. Gemini: you crop the screenshot tightly and upload it. Gemini reports no SynthID watermark, so Google's tools probably didn't make it.
  3. OpenAI verify: no signal. Since it's a screenshot, any Content Credentials would already be gone, so this result barely counts.
  4. Content Credentials: nothing to read, for the same reason.
  5. Reverse search: About this image shows near-identical copies on several pages, the earliest from two years ago, credited to a news agency covering a different flood.
  6. Details: the shop signs are legible and consistent, which fits a real photo.

Verdict: a real photo, but an old one with a false caption. The AI checks came back empty, and the reverse search did the actual work. I'd guess that's the most common outcome you'll get.

Common mistakes and how to fix them

  • Trusting one "AI or not" score. Fix: treat any single detector as one vote; accuracy is contested, as our piece on whether AI detectors are accurate explains.
  • Checking a screenshot when the original exists. Fix: save the file itself, because screenshots and conversions strip Content Credentials.
  • Reading "no watermark" as "real". Fix: remember each checker only sees its own company's watermark.
  • Uploading a collage. Fix: Gemini works best with one image, cropped tight.
  • Skipping the caption. Fix: ask whether the place, date and weather match the claim.

Tips to get better results

Run the watermark checks before you crop, compress or edit anything. Each change chips away at the signals, and OpenAI says heavy cropping or compression can make a watermark undetectable.

For a second opinion on images from other generators, you can try one of the third-party tools in our AI detectors roundup, but I wouldn't lean on them alone. They guess from patterns; they don't read a signature. Our explainer on how AI detectors work covers the difference between watermark checks and pattern-guessing.

It also helps to know what current generators produce. Spend five minutes browsing the output of the tools in our AI image generator guide and you'll start spotting the glossy, too-perfect lighting many of them share.

Next steps

Bookmark gemini.google.com, openai.com/verify and contentcredentials.org/verify so they're one tap away. If you post AI images yourself, label them. It seems a small courtesy, and it saves everyone else these ten minutes.

FAQ

How to tell if image is AI generated in 2026?

Upload the original file to Gemini and ask "Is this AI-generated?" to check for Google's SynthID watermark, then run it through openai.com/verify and contentcredentials.org/verify. Next, reverse search it with Google's About this image. Finally, zoom in on text, hands and shadows. No single check is proof, so weigh them together.

Is this AI: is there a quick way to check one picture?

Use Google Lens or Circle to Search: select the image and ask "Is this made with AI?" It checks for Google's SynthID watermark in seconds. It can't see most other generators, though, so a negative result only rules out Google's tools.

Is this AI? Can a free detector give a definite answer?

Only sometimes. A detected watermark or a Content Credential saying AI was used is strong evidence. Everything else is a probability. Pattern-based detectors can misfire on heavily edited or low-quality real photos, and they can miss newer generators entirely, so treat any score as one clue among several, and look for an original source or an earlier copy of the picture before you decide.

How to tell if images are AI generated when they're screenshots?

Screenshots lose Content Credentials, so focus on watermarks and context. Crop tightly around the picture before uploading it to Gemini, since Google says collages and loose crops hurt accuracy. Then reverse search it. If you can, find the original post and save the actual file, which keeps far more evidence intact.

Which AI is this? Can I find out which tool made an image?

Sometimes. Content Credentials often name the app that made the file, openai.com/verify confirms OpenAI tools and Gemini confirms Google AI. Beyond those signals, I'd be wary of any site claiming it can name the generator from pixels alone.

Keep reading