How to check if an image is AI generated

When you want to know whether a picture came from a real camera or an AI model, there are two completely different ways to check it: reading the cryptographic and metadata records inside the file, or looking at the pixels.

Only the first method gives you a verifiable fact. The second method gives you a clue, and if you rely on online “98% AI probability” meters without understanding how they work, you will regularly accuse real photographers of using AI and wave clean AI renders through.

What real AI provenance labels look like inside a file

Major AI generators and camera makers now embed structured provenance data directly into the image container before you download it. There are three places inside a JPEG, PNG or WebP file where generator identity is recorded:

1. C2PA Content Credentials (JUMBF): the Coalition for Content Provenance and Authenticity (C2PA) defines a cryptographically signed manifest stored inside APP11 marker segments in JPEG (per the C2PA Specification), caBX chunks in PNG, or JUMBF chunks in WebP. Tools like OpenAI DALL-⁠E / ChatGPT Images, Adobe Firefly, Microsoft Designer and Leica/Sony/Nikon hardware-signed cameras attach a signed C2PA claim showing the generator name and edit chain.
2. IPTC DigitalSourceType and XMP Credit: the International Press Telecommunications Council standardized the trainedAlgorithmicMedia URI (http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia) and compositeWithTrainedAlgorithmicMedia inside XMP/IPTC metadata so search engines and social platforms can read an explicit AI label even without a C2PA certificate.
3. PNG tEXt / iTXt generator chunks: Midjourney writes your prompt, job ID and DigitalSourceType into PNG text chunks (Description, Author, XML:com.adobe.xmp), while local tools like Automatic1111 Stable Diffusion and ComfyUI store the entire prompt, negative prompt, seed, steps and checkpoint filename under parameters, prompt or workflow chunks.

How to check an image in three steps

You can inspect all three metadata layers inside your browser tab without uploading the image to a third-⁠party detector:

1. Ask for or download the original image file rather than a screenshot or a compressed preview from a chat thread.
2. Open the photo metadata remover and drop the JPEG, PNG or WebP file onto the inspection box.
3. Read the inspection report for Content Credentials (C2PA / JUMBF), Software or Generator tags, XMP Credit or DigitalSourceType, and PNG text chunks, and check whether normal camera EXIF tags (Make, Model, shutter speed, ISO, lens focal length) are present or absent.

How to interpret what the metadata report shows

Why pixel-⁠based “AI image detectors” disagree so often

When a file has no metadata left, websites that claim to detect AI images feed the picture into a neural network trained to spot high-⁠frequency diffusion artifacts. Here is why two detectors often report “96% AI” and “4% AI” on the exact same image:

• Modern smartphone cameras use computational photography: every recent iPhone and Android flagship merges multiple exposures, runs neural denoising and applies semantic sharpening to faces, skies and foliage. Those computational passes smooth out natural sensor grain in ways that look statistically similar to diffusion models.
• Re-⁠compression erases high-⁠frequency diffusion residuals: resizing an AI image to 1080 pixels wide and saving it as an 80%-quality JPEG wipes out the microscopic checkerboard frequency patterns that pixel classifiers look for.
• Invisible pixel watermarks (like Google SynthID) are vendor-specific: Google DeepMind’s SynthID embeds an imperceptible signal into the pixel values of Imagen and Gemini images that survives cropping and filters, but only Google’s own SynthID verifier can read that specific watermark: it does not show up in generic third-⁠party AI detectors.

What still gives AI images away visually when metadata is gone

When you are reviewing a stripped image with no metadata, check the physical geometry that image models still get wrong more reliably than any percentage meter:

• Small background text and signage: secondary street signs, book spines, nutrition labels and serial numbers in the background often dissolve into pseudo-⁠Latin shapes or inconsistent letterforms.
• Architectural and reflection geometry: follow tile grout lines, window mullions, brick courses or reflections in eyeglasses and puddles behind a subject’s head; diffusion models frequently shift alignment where an object interrupts a straight line.
• Hardware and mechanical details: check button holes vs buttons on jackets, watch dial markings, bicycle spokes, guitar frets and shoelace crossings.

For a deeper look at what happens when someone strips provenance headers from an AI image, read our guide on how AI watermarks and metadata work.