How to tell if an image is AI: the free checks that actually work

The fastest check costs nothing and needs no tool: ask where the image came from. Before you upload anything anywhere, run through these questions.

• Who posted it first? A photo shared by an anonymous account with no original source deserves more suspicion than one from a named photographer or an official account.
• Does anyone else have it? Real news photos of real events usually exist from several angles and several people. A single dramatic shot that no one else captured is a warning sign.
• Is it too perfect for the moment? Famous people in improbable situations, flawless lighting at a chaotic scene, or an image that perfectly confirms what a post claims should all slow you down.
• Can you find older copies? A reverse image search in Google Lens or a similar tool can show whether a picture appeared earlier with a different story attached.

None of these prove anything alone, but together they catch a lot of fakes before you need a detector.

SynthID is an invisible watermark that Google first introduced in 2023. Google’s Gemini models stamp it into everything they make, and partners including OpenAI, Nvidia and South Korea’s Kakao now use it too. You can’t see it in an image or hear it in audio, and Google says it survives common edits such as cropping and compression.

Until recently, only researchers and journalists could check for it directly. As of October 7, the SynthID Detector website is open to everyone, in English, worldwide. Here’s how to use it:

1. Save the file. Download the image, video or audio clip you want to check to your phone or computer.
2. Go to synthid.com. You’ll be asked to agree to the terms of service the first time.
3. Sign in. You need a Google, Apple or ChatGPT account.
4. Upload the file. Results arrive within a few seconds.
5. Read the answer carefully. The tool tells you whether it found a SynthID watermark. It doesn’t tell you which parts of the media were made or changed by AI.

A few practical details: you’re reportedly limited to around 10 checks a day, a cap meant to stop people using the tool to learn how to strip the watermark. On privacy, Google says uploaded files are processed in real time and deleted once results are shown, while a digital signature of the file is linked to your account and deleted after 24 hours.

This is the mistake most people will make. SynthID Detector isn’t a general AI detector. It only looks for one specific watermark. If an image was made with a tool that doesn’t use SynthID, the detector has nothing to find.

When the tool doesn’t find a mark, it says it’s “unlikely” the media came from a SynthID partner, and warns that it may still have been made by another company’s AI, or before a partner adopted SynthID. That wording matters. Plenty of popular image generators sit outside the system, including tools from Meta and xAI’s Grok.

There are gaps even among partners. Apple has said it will support SynthID, and Google says Apple support is coming soon with no firm date. In one test, a photo from an iPhone 18 Pro edited with Apple’s Reframe tool was correctly flagged, but an image made in Apple’s Image Playground app was not.

A positive result has limits too. It tells you AI was involved, not whether the picture’s claims are true or false.

No single checker covers everything, so it helps to know a few others:

• OpenAI’s own checker looks for OpenAI’s watermarks and AI-confirming metadata, but it only covers OpenAI’s tools.
• Meta’s detector looks for its own watermark, called Content Seal. A 2026 investigation found it sometimes failed to flag content made with Meta’s own AI.
• Content credentials (C2PA) are tamper-evident labels some cameras and phones attach to photos to record how they were made. Google’s Pixel phones add C2PA data to every photo and add SynthID when AI editing is used. If a photo carries credentials showing it came straight from a camera, that’s a strong sign it’s genuine.
• Apple’s Reference mode keeps a cryptographically verified original of a photo, but only for shots taken in that mode in the iPhone Camera app.

Anthropic and Microsoft have also started watermarking AI media, but neither offers a public tool to check for those marks yet.

Why you shouldn’t just ask a chatbot

It’s tempting to paste a picture into ChatGPT, Gemini or Grok and ask, “Is this real?” Be careful. In one test, Gemini correctly said an edited image was AI, then invented a detailed backstory about an “original” photo that doesn’t appear to exist. Grok, asked about a fake image it had generated itself, guessed it came from a different AI tool. In another case, chatbots asked about a fake picture of a famous tech boss gave conflicting answers, with one calling it authentic.

Chatbots are good at describing what’s in an image. They are not reliable judges of where it came from. The same goes for many free “AI detector” websites that guess based on patterns rather than watermarks: they often flag human-made work as AI.

Visual clues that still help

Your eyes aren’t useless, just less reliable than before. When you look closely, check for:

• A plastic sheen on skin and surfaces, especially from cheaper image generators.
• Garbled details in text, signs, logos, jewelry, glasses frames and the edges of hair.
• Odd faces in the background, where generators often put less effort.
• Lighting and shadows that don’t agree with each other.
• A gut feeling that something is off. It’s not proof, but it’s a good reason to run the other checks.

For everyday social media use, a sensible routine is: check the source, look closely, run the file through SynthID Detector, and if it matters, look for other coverage of the same event. If the detector finds a watermark, you know AI was involved. If it doesn’t, you’ve ruled out only some AI tools, not all of them. When in doubt, don’t share it. A fake that you don’t repost is a fake that stops with you.