How to Tell If an Image Is AI-Generated: 6 Checks

Alejandro Rioja
Alejandro Rioja
6 min read
TL;DR

No single test proves an image is real. Run them in order: ask Gemini or Google's SynthID Detector whether Google or partner AI made it, check for Content Credentials, look for a platform's AI label, read the file's metadata, find the earliest copy with a reverse search, then look at the details. A positive watermark is strong evidence; the absence of one proves nothing.

Free newsletter

Every Wednesday. 28,400+ operators. Zero fluff.

There’s no button that tells you for certain whether a photo is real. What you have is a stack of checks. A few can confirm that an image was made by AI. None can prove that it wasn’t. So the job is to run the cheap, high-confidence checks first and treat everything after them as evidence, not a verdict.

This matters for a business more than it used to. Before you repost a customer photo, run a creator’s image in an ad, or reply to a “look what your product did” complaint, you want to know what you’re looking at.

The rule that saves you. A detected watermark or a signed credential that says "AI" is strong evidence. A missing one means nothing, because metadata gets stripped by screenshots and most social uploads. Never call an image real just because a checker came back empty.

The six checks, ranked by how much they tell you

CheckWhat it can tell youWhat it can’tTime
1. SynthID (Gemini, SynthID Detector)That Google AI or a partner’s tool made or altered itAnything about other generators1 min
2. Content Credentials (C2PA)The signed history of how the file was made and editedAnything, once the credential is stripped1 min
3. Platform AI labelsThat the platform detected industry signals or the uploader disclosed AIUnlabelled AI imagesSeconds
4. File metadataCamera, date, editing software, sometimes an AI tool’s tagProof either way; it’s easy to edit or remove1 min
5. Reverse image searchWhere and when the image first appearedWhether the first copy was itself real2–5 min
6. Visual inspectionTells that raise or lower suspicionCertainty; good generators fix the classic tells2 min

1. Check for a SynthID watermark

SynthID is Google’s invisible watermark for AI content. Google says it’s embedded across its generative AI consumer products and is designed to survive cropping, filters and lossy compression (Google DeepMind, SynthID).

Two ways to check:

  • Ask Gemini, Google Search or Chrome. Upload the image and ask whether Google AI created or altered it. Google says the system checks for a SynthID watermark.
  • Use SynthID Detector. Google describes it as a verification portal where anyone can check whether an image, video or audio file was made with Google AI or partner tools. It lists OpenAI, NVIDIA and Kakao as partners, with Apple coming.

A hit here is about as close to proof as you’ll get. A miss only means the image didn’t come from Google or a listed partner, or the watermark didn’t survive.

2. Look for Content Credentials

Content Credentials are the C2PA standard: a signed record that travels with the file and describes “how the content was created, what tools or processes were used, when and where it was made, and how it has changed over time” (C2PA FAQ). Because each credential is cryptographically signed, C2PA says any modification “will break this cryptographic linkage, signalling tampering.”

Drop the file into the Content Credentials Verify site (contentcredentials.org/verify) to read any credential it carries.

The limit is the same as with any metadata: credentials can be separated from the file. C2PA’s answer is “durable” credentials that use watermarking or fingerprinting to find the record again. Until most tools support that, a screenshot will usually arrive with no credential at all.

3. Read the platform’s label

Meta labels AI-generated images, video and audio when it detects “industry-shared signals” or when the uploader discloses AI, and shows photorealistic Meta AI images as “Imagined with AI”. In July 2024 it renamed the “Made with AI” label to “AI info”, and since September 2024 content that was only edited with AI tools has the label in the post’s menu rather than on the image (Meta).

So on Facebook and Instagram, open the post’s menu before deciding there’s no label.

4. Read the file’s metadata

If you have the original file, its EXIF data can show the camera, the date and the software that last saved it. A photo straight off a phone usually has a make, model and lens. An image exported from an AI tool usually has none of that, and some tools write their own name into the software field.

My free EXIF viewer shows everything a photo carries, in your browser. Treat what you find as a clue: metadata is easy to strip and easy to fake, and most social platforms remove it on upload.

5. Find the earliest copy

Most misleading images aren’t brand new. They’re old images under a new caption, or AI images that first appeared somewhere with an honest description. A reverse image search finds the earlier copies.

Use Google Lens first, then a second engine. On a phone, my guide to reverse image search on iPhone and Android has the steps; on a computer, the reverse image search tool sends one image to Google Lens, Bing, Yandex and TinEye at once. If the earliest copy sits on an AI art account or in a prompt-sharing community, you have your answer.

For video, the same logic applies frame by frame. See how to verify where a video really came from.

6. Look closely, and know the limits of looking

Visual tells still catch lazy fakes: text that turns into gibberish, jewellery or glasses that don’t match on each side, reflections and shadows that disagree with the light, hands, and background details that melt into each other. Zoom in on edges where two objects meet.

But don’t let a clean-looking image reassure you. The better generators fix the famous tells, so visual inspection can raise your suspicion and can’t settle it.

What to do with the answer

  1. 01
    Confirmed AI. If you still want to use it, label it. Meta and other platforms ask uploaders to disclose AI content.
  2. 02
    Unclear. Ask the person who sent it for the original file or another photo of the same scene. Real photographers can usually produce both.
  3. 03
    Earliest copy found, and it's real. Credit the original source, not the account that reposted it.

AI image detection FAQ

Is there a free tool that tells you if an image is AI-generated?

Yes, with limits. Google’s SynthID Detector and Gemini check for Google’s SynthID watermark, which covers Google AI and the partners Google lists. The Content Credentials Verify site reads C2PA credentials from any tool that writes them. Neither can clear an image that carries no watermark or credential.

Can AI detectors be wrong?

Yes. Watermark and credential checks are reliable when they find something, because the signal is either there or it isn’t. Classifier-style detectors that guess from the pixels can misfire on heavily edited real photos and miss AI images, so don’t use their score as your only evidence.

Does a screenshot remove the AI watermark?

A screenshot removes the file’s metadata, including Content Credentials. Google says SynthID is designed to survive edits like cropping, filters and lossy compression, so it’s built to outlast that, but the C2PA credential usually won’t.

How do I check if a photo on Instagram is AI?

Look for the “AI info” label on the image or in the post’s menu, which Meta applies when it detects industry signals or the uploader discloses AI. Then save or screenshot the image and run a reverse image search to find the earliest copy.

Advertisement
Keep reading

Related posts

Keep reading

Get the AI playbook in your inbox

Every Wednesday. 28,400+ operators. Zero fluff.

↵ ↵ to see all results esc esc to close