An AI image detector checks whether a picture was made or edited by AI. The most reliable way to tell is provenance: Content Credentials (C2PA), IPTC "digital source type" labels and generator settings that AI tools write into the file. This free tool reads all of them in your browser, adds Error Level Analysis and a visual checklist, and explains the evidence instead of guessing a percentage.
Private by design: no upload, no AI calls, nothing stored. Your image is analysed in this browser tab and never leaves your device.
Drop, choose or paste an image. We read its Content Credentials, IPTC/XMP, EXIF and generator data, run Error Level Analysis, and explain what we found.
Drag and drop an image here, or paste one with Ctrl+V / Cmd+V
JPEG, PNG, WebP, AVIF or HEIC · up to 50.0 MB · use the original file if you can
A URL is fetched by your browser straight from that site (no AI Wins server). Many sites block this; if so, download the image and drop it here.
No upload, no AI calls, nothing stored. Your image stays in this browser tab.
Samples are synthetic images made for this tool, with made-up metadata.
Content Credentials (C2PA) in JPEG APP11, PNG caBX, WebP C2PA and HEIC/AVIF uuid boxes; the IPTC digital source type in XMP; EXIF camera data; and prompts and settings left by Stable Diffusion WebUI, ComfyUI, InvokeAI and NovelAI.
An Error Level Analysis heatmap (large images are analysed at up to 2048px), plus dimensions and whether they match common generator sizes such as 1024x1024, 1792x1024 or 1536x1024. These are clues, not proof.
A guided checklist of tell-tale signs: hands, text, reflections, jewellery symmetry, background faces, lighting, waxy skin, repeated patterns and watermark remnants.
What it can't do: it doesn't verify C2PA signatures (use Content Credentials Verify (opens in a new tab)), it can't read invisible watermarks such as Google SynthID (use SynthID Detector (opens in a new tab) or the Gemini app), and it doesn't use an AI classifier, so it won't give you an "AI probability" score.
The first rule that matches wins. Evidence is marked strong, moderate or weak.
trainedAlgorithmicMedia (strong): Created using generative AI trained on sampled content.algorithmicMedia (moderate): Created purely by an algorithm, without sampled training data.compositeWithTrainedAlgorithmicMedia (strong): Edited using generative AI, e.g. inpainting, outpainting or generative fill.compositeSynthetic (moderate): A mix of elements, at least one of which is synthetic.digitalCapture (moderate): Recorded from real life with a digital camera or recorder.computationalCapture (moderate): Several real-life frames merged automatically by the device (e.g. HDR).negativeFilm (moderate): Scanned from a transparent negative.positiveFilm (moderate): Scanned from a transparent positive (slide).print (moderate): Scanned from a non-transparent medium such as a photo print.humanEdits (weak): Edited by people with non-generative tools.algorithmicallyEnhanced (weak): Altered by an algorithm, main content kept (e.g. sharpening, noise reduction).digitalCreation (weak): Made by a person with non-generative digital tools.dataDrivenMedia (weak): Represents data through human programming, e.g. a chart.screenCapture (weak): A capture of a computer or phone screen.virtualRecording (weak): A live recording of a virtual event.composite (weak): A mix of elements, any of which may or may not be generative AI.compositeCapture (weak): A mix of elements that are all real-life captures.minorHumanEdits (weak): Retired IPTC term; now humanEdits.digitalArt (weak): Retired IPTC term; now digitalCreation.softwareImage (weak): Retired IPTC term.Camera make, model and exposure settings are easy to copy from another photo or edit, so they only count as a weak camera signal. Your checklist ticks are your own judgement: 2 or more can turn an inconclusive result into "Looks suspicious based on your observations", but they never override what the file's metadata says.
Check the file first: Content Credentials (C2PA) or an IPTC digital source type of trainedAlgorithmicMedia mean the creating tool labelled it as AI, and Stable Diffusion or ComfyUI settings in a PNG are a giveaway. Then look closely at hands, text, reflections, earrings and background faces, and run a reverse image search to find the original source.
Yes. This AI Wins tool is free with no sign-up. It reads Content Credentials, IPTC/XMP, EXIF and generator metadata, runs Error Level Analysis and guides a visual check, all in your browser. For Google's invisible SynthID watermark, use Google's SynthID Detector or ask the Gemini app "Is this AI-generated?".
No AI image detector is 100% reliable. Research (Saberi et al., ICLR 2024) shows classifier-based detectors can be fooled and watermarks can be removed or faked. Metadata is reliable when present but is often stripped. That's why this tool shows each piece of evidence and its strength instead of a single AI probability score.
Content Credentials are a signed provenance record defined by the C2PA standard and embedded in the file (an APP11 segment in JPEG, a caBX chunk in PNG, a C2PA chunk in WebP). They can say which app made the image and whether generative AI was used. OpenAI, Google, Microsoft Designer and Adobe Firefly add them to AI images. This tool reads them; verify signatures at contentcredentials.org/verify.
Yes. Screenshots, social media re-uploads, most messaging apps and many editors strip Content Credentials, IPTC and EXIF data, and metadata can also be edited or faked. So missing metadata proves nothing: an image with no AI markers can still be AI generated.
No. The image is read by JavaScript in your browser tab: metadata parsing and Error Level Analysis run on your device, there are no AI or API calls, and nothing is stored or logged. If you check an image by URL, your browser fetches it directly from that website, not through AI Wins.
Facts checked November 2, 2026. Vendors change what they embed; the official pages are the final word.
Write detailed prompts for Midjourney, DALL-E, Stable Diffusion and more.
Open toolCreate a clear statement saying how AI was used in your work.
Open toolCheck citations and sources in AI answers before you publish.
Open toolTest how well you understand what AI can and can't do.
Open toolMake AI-assisted drafts read more naturally.
Open toolFollow the latest AI image models and how to spot them in the daily AI Wins digest. Free, positive AI news, summarized every morning.
Get the free digestOnly positive AI stories, delivered daily or weekly. Free, no spam, unsubscribe anytime.