How to Detect AI Generated Images: What Actually Works

Someone sends you a photo on WhatsApp of a politician at an event that never happened, or a flooded street in your city. You want to know whether it's real, and you want to know in thirty seconds.

Most advice on how to detect AI generated images points at the wrong thing first. Counting fingers worked in 2023; current image models mostly get hands right. The reliable checks come from the file and the context, not from staring at the pixels. Here's the order to work in.

Step 1: Look for Content Credentials

Content Credentials are a provenance standard, backed by Adobe and the C2PA group, that attaches signed information to a file about what made it and what edited it since. A growing number of cameras, phones and AI tools write them.

Drop the image into the Verify tool at contentauthenticity.org. If credentials are present, you'll see which app or device created the file and what touched it afterwards.

The catch, and it's a big one: credentials only exist if the creating tool added them, and they can be lost when a file is screenshotted, re-saved or stripped by a platform. Absence proves nothing. Presence is informative; silence isn't.

Step 2: Check for a SynthID watermark

Google DeepMind's SynthID embeds an invisible watermark into content made with Google's generative AI products. Google describes it as "an invisible digital watermark" added to AI-generated images and video segments, and says it's "designed to stand up to modifications like cropping, adding filters, changing frame rates, or lossy compression." It also covers audio from Google's music model and NotebookLM's podcast feature, and text from the Gemini app.

That durability is what makes it more useful than metadata. The limitation is scope: SynthID marks content made with Google's tools. An image from a different generator won't carry it.

Google has built a SynthID Detector verification portal, but access is limited. Google says it is "collaborating with journalists and media professionals to test the portal," with an early-tester waitlist, so most people can't simply upload a file to it yet.

Some Google surfaces also add a visible watermark to generated images. In Gemini Notebook, for example, Google says a visible watermark is added to generated images and videos, with a toggle for eligible Google AI Pro and Ultra subscribers.

Step 3: Reverse image search

This catches two common cases that no detector will: a real photo being passed off as new, and a synthetic image that's already been debunked.

Run the image through Google Lens, Bing Visual Search or TinEye. Look for the earliest appearance, not the most recent. A "breaking news" photo that shows up in a 2019 article is recycled, whatever made it.

Step 4: Read the context, not the pixels

Ask the questions a picture editor asks:

• Who published it first, and do they say where it came from? A real event has several photographers and several angles.
• Does anything in the frame contradict the claim? Signage language, number plates, weather, the season of the trees, which side of the road traffic drives on.
• Is there video? Synthetic stills are far more common than convincing synthetic video of the same event from multiple angles.
• Does the account that posted it have a history? New accounts posting emotionally charged images are a pattern.

Step 5: Visual tells, with low confidence

Only now look at the image itself. Current tells, in rough order of usefulness:

1. Text in the image. Signs, labels and shopfronts still come out subtly wrong in many generators, though Ideogram and newer models handle short words well.
2. Repeating patterns. Crowds, railings, windows and bricks that repeat too perfectly or dissolve into mush.
3. Physics that doesn't hold. Shadows pointing different ways, reflections that don't match, jewellery merging into skin.
4. Surface smoothness. Skin, fabric and foliage that look airbrushed at full zoom.
5. Edges of the frame. Generators often get lazy at the periphery: a half-formed hand, a bag with no strap.

Treat every one of these as a hint. Real photos have odd lighting and strange compression artefacts too, and AI image editing means a photo can be 90% real with one generated element.

Why AI detector sites cannot detect AI generated images reliably

Upload-an-image detectors give a confident percentage from a model guessing at statistical texture. They produce false positives on real photos and false negatives on current generators, and they rarely tell you why.

The evidence from the text side is instructive. A Stanford study published in the journal Patterns found that seven widely used GPT detectors flagged, on average, 61% of essays written by non-native English speakers as AI-generated, while nearly all essays by US eighth-graders were classified correctly. The researchers warned against using such detectors in evaluative settings. Image detectors are built on the same kind of statistical guessing.

Use them, if at all, as one weak signal after the four steps above.

A 60-second routine

1. Drop the file into the Content Credentials Verify tool.
2. Reverse image search it.
3. Check who posted it first and what they claim.
4. Look for text, repetition and shadows in the image.
5. If it still matters and you still don't know, say you don't know. That's a legitimate answer, and a better one than a wrong confident call.

For context on the tools making these images, see our best AI image generators ranking and our DALL·E vs Midjourney comparison.

FAQ

What is the most reliable way to detect AI generated images?

Checking provenance: Content Credentials attached to the file, and Google's SynthID watermark for images made with Google's AI tools. Both can be absent, so combine them with a reverse image search and the context of who posted it.

Can I check if an image has a SynthID watermark?

Not easily yet. Google's SynthID Detector portal is in testing with journalists and media professionals through a waitlist, so public access is limited.

Do AI image detector websites work?

Not reliably. They give confident percentages from statistical guesswork, and research on the text equivalents found detectors wrongly flagged most essays by non-native English writers. Use them only as a weak extra signal.

Does counting fingers still spot AI images?

Rarely. Hands were a giveaway in early models and most current generators handle them. Text inside the image, repeating patterns and inconsistent shadows are better tells today.

If an image has no Content Credentials, is it fake?

No. Credentials exist only if the creating tool added them, and they can be stripped by screenshotting or re-saving. Absence tells you nothing either way.