How to tell if an image is AI-generated

There is no single test that tells you whether a picture is real. What works is a short routine that starts with the strongest evidence, the records an AI tool leaves in the file and the watermark it hides in the pixels, and only falls back to eyeballing the image and tracing its source when those are missing. Here it is in five steps, with the free tools for each.

Step 1: Get the original file

Labels live inside the image file as it was made or downloaded. Screenshots never have them, and most social apps and messengers strip them when you post or forward an image. If someone sent you the picture, ask for the original. If you found it online, save the largest version you can find rather than taking a screenshot.

Step 2: Read what the file records

Most big AI image tools now write something into their files:

• Content Credentials, a signed record defined by the C2PA standard. ChatGPT, Adobe Firefly, Microsoft’s image tools and Google’s newer image models add them, and they name the tool and the company that signed them. See Content Credentials explained.
• An AI label in the metadata, the IPTC “digital source type”, which says an image was generated or edited with AI.
• Generator settings. Stable Diffusion apps such as AUTOMATIC1111 write the prompt and settings into a text field in PNG files, or into the EXIF comment of JPG and WebP files (AUTOMATIC1111). ComfyUI saves the prompt and its workflow inside PNG files (ComfyUI).

The free uncrop.si AI image checker reads all three on your device, verifies Content Credentials against the public C2PA trust lists, and shows any prompt it finds. A trusted record that says “created with generative AI” is close to proof. A label or settings are strong evidence, though they can be copied into another file.

Step 3: Check for an invisible watermark

Some AI companies also hide a watermark in the pixels, which survives far more editing than a label does. The most widely used is Google DeepMind’s SynthID, which Google adds to images from its own AI and which OpenAI has added to ChatGPT images since May 2026 (OpenAI Help Center). Only the companies that add it can read it:

• For Google images, use Google’s free SynthID Detector, or ask the Gemini app whether the image was created or edited by Google AI (Gemini Apps Help).
• For OpenAI images, use OpenAI’s verify page, or run OpenAI’s check with your own API key in the uncrop.si checker.

A positive result is strong evidence. A negative one only rules out recent images from those companies whose watermark survived. Read what SynthID is and how to check it for the details.

Step 4: Look closely at the image

When the file and the watermark checks come up empty, the picture itself can still give clues. Fact-checkers suggest looking at:

• Hands, ears, toes and teeth, where errors often appear (Full Fact).
• Text, letters and numbers, which are often garbled in AI images (AAP FactCheck).
• Objects that cannot work, such as Full Fact’s example of a rifle with barrels at both ends.
• Shadows, reflections and textures that do not match the scene.
• Skin and surfaces that look too polished or have an unnatural sheen for the setting (Euronews with AP).

Researchers at Northwestern group these clues into anatomical, stylistic, functional, physics and sociocultural errors, and stress that such artifacts are not always present (Kamali et al.). People are not very good at this: in a 2025 Microsoft study of about 287,000 judgements, participants identified real and AI images correctly only 62% of the time (Microsoft AI for Good Lab). Fact-checkers agree the visual signs are getting rarer as models improve.

Step 5: Check the source

Where an image comes from often settles the question faster than its pixels do.

1. Run a reverse image search with Google Lens, Bing Visual Search or TinEye and look for the earliest copy.
2. Check who posted it first, when, and whether that account has a record of reliable posts.
3. Look for the same scene from other angles or in reporting from trusted outlets.
4. Search Google Fact Check Explorer to see whether fact-checkers have already covered it.

Why “AI detectors” are not enough

Many websites offer an AI image detector that guesses from the pixels. Independent tests do not support relying on them. On a benchmark of real-world images collected in 2024, the best open-source detectors scored about 45% lower than on older test sets (measured by AUC, a standard accuracy score), and commercial tools, while better, did not reach the level of trained forensic analysts (Deepfake-Eval-2024). Bellingcat found that one popular detector labelled prize-winning real photographs as AI-generated, while compressed AI images often passed as real (Bellingcat). Chatbots asked in an ordinary conversation do no better: in tests reported by AFP, Google’s AI Mode called an image made with Google’s own Nano Banana authentic (AFP via Gulf News).

Use those tools as one more hint at most, and lean on the checks that read real evidence: the file’s records and the companies’ own watermark checks.

Which AI tools leave which marks

AI tool	Content Credentials	Invisible watermark	Visible mark or label
ChatGPT and the OpenAI API	Yes, signed by OpenAI	SynthID, since May 2026	None
Google Gemini and Nano Banana	Yes, on Nano Banana Pro and newer models	SynthID	Gemini sparkle on free and AI Pro plans (Google)
Adobe Firefly	Yes, when every pixel comes from Firefly (Adobe)	Not documented	None
Bing Image Creator	Yes	Not documented	Mark in the bottom left corner (Bing)
Meta AI	Not documented	Meta’s own watermark, with no public checker	“Imagined with AI” on photorealistic images (Meta)
Stable Diffusion apps and ComfyUI	No	None you can check	None, but the prompt and settings often stay in the file

The uncrop.si AI image checker reads every mark in this table that can be read outside the company that made it.