Drop in a picture and find out whether it was made by AI, and by which tool.
An AI image detector is a tool that analyzes a picture to judge whether it was generated or altered by AI. This free checker returns a verdict and an AI score. It also runs a source check, a metadata and C2PA check that names the generator, such as Midjourney or DALL-E, when the file still carries that evidence.
Free account: 50 scans a month across image, video and voice, files up to 12 MB
Two independent checks on every upload. One reads the pixels. One reads the file. You see both.
The free checker takes JPG, PNG, WebP, BMP, TIFF and HEIC files up to 4.5 MB.
Our in-house AI image detector returns two probabilities: that the image was generated, and that part of it was inpainted, meaning an AI filled or replaced a region. The AI score is the sum of the two, capped at 100. The four verdicts below are cut from that score.
It reads EXIF Software, XMP CreatorTool, PNG text chunks, the JPEG comment, IPTC keywords and the C2PA manifest, and matches them against a signature table of 29 AI generators, 5 editing tools and a screenshot class. This is forensics on the file, not a guess by the model, and it never changes the verdict.
No meaningful signs of AI generation in the pixels.
More natural signals than AI ones, with some mixed areas. A partial AI edit can land here.
More AI signals than natural ones. Treat as probably AI and look for a second signal, such as a source check match.
Strong signs of AI generation across the whole image.
A C2PA Content Credentials manifest, or generation metadata written by the tool itself, such as a Stable Diffusion parameters block or a ComfyUI workflow. The most reliable attribution available.
A generator name in EXIF, XMP or IPTC fields, or provenance data that does not name the tool.
A weak marker, for example a screenshot or export signature. It tells you how the file was captured, not what made the original picture.
Normal for anything downloaded from social media or messaging. Says nothing about whether the image is real.
Most free checkers return one number. We return the number and, when the file still carries it, the name of the tool.
The source check is a metadata and provenance check that names the generator or editing tool recorded inside an image file. It reads the same fields a camera or an app writes when it saves a picture, plus C2PA Content Credentials, and matches them against a signature table.
The detector scores any image, whichever generator produced it, because it reads pixels. Attribution is different. A generator can only be named when the file still carries its signature. These are the tools in the current signature table.
Adobe Photoshop, Adobe Lightroom, GIMP, Canva and Affinity Photo are classified as editing tools, not generators. If the last tool to save a file was Photoshop, the source check says so and does not call the image AI generated.
Screenshots and screen exports get their own class, because a screenshot of an AI image carries the screenshot tool's signature, not the generator's.
Because the metadata is gone. That is the normal case, and a blank result here says nothing about whether the image is real.
A resized, re-saved or screenshotted copy of an image usually carries no metadata, and without metadata the source check has nothing to read.
Most social platforms and messaging apps strip EXIF and XMP on upload. Resizing in an image editor rewrites the file. A screenshot creates a brand new file that knows nothing about the original. In each case the source check reports none, and the result panel says so plainly.
Three things follow. First, a blank source check result does not mean the image is real. Second, the verdict and AI score do not depend on metadata at all, so they still apply. Third, when you can, check the original file rather than a copy, because that is where the evidence lives.
The reverse is also true. A strong source check match on a Midjourney or Firefly file is evidence from the file itself, independent of the model, which is why we show it next to the score instead of folding it in.
The detector scores inpainting on its own, which is how partial AI edits show up in a mostly real photograph.
A deepfake image is a real photograph in which a region, usually a face, has been replaced or regenerated by AI.
Most free AI checkers answer one question: was this whole picture generated? A face swap or an AI-filled background fails that test, because most of the pixels are a real photo. Our model returns a separate probability that part of the image was inpainted, and that probability counts toward the AI score. A convincing swap on a real photo can therefore land in Likely AI or Likely Human rather than Human, and that is the signal to look harder.
For an explanation of edited versus generated, read edited vs AI generated images: how to tell them apart.
Use these alongside the score, especially for a Likely AI or Likely Human result.
Count fingers and look at how they join the palm. Teeth that merge or repeat are a common tell.
Signs, labels and logos that look like letters but do not spell anything.
A mirror or window that does not show what it should, or shadows that fall in two directions.
Buildings that do not meet the ground, railings that change pattern, objects that merge into each other.
Skin that is uniformly smooth, or fabric with a repeating pattern that drifts.
Open the original file's properties. A C2PA Content Credentials label, or a generator name in the software field, is the same evidence the source check reads.
The full walkthrough is in how to tell if an image is AI generated. For what happens inside a detector, read how AI image detectors work.
Yes. The checker on this page needs no account. An account adds volume, bigger files and history.
If yours is not here, mail hello@deepfakedetector.ai.
Drop the image into the free checker on this page. You get a whole-image verdict, an AI score from 0 to 100, and a source check result that names the generator when the file's metadata or C2PA credentials still carry it. For a borderline score, also check the image by eye: hands, text, reflections and background logic.
Yes, when the evidence is still in the file. The source check reads EXIF, XMP, PNG text chunks and the C2PA manifest and matches them against a signature table covering Midjourney, DALL-E, Stable Diffusion, Firefly, Gemini, Flux and others, with a strong, moderate, light or none evidence level. A resized, screenshotted or re-saved copy usually carries no metadata, and the source check reports none.
We do not publish an accuracy percentage, because a single number hides how much results vary by generator, compression and image size. Treat the score as evidence, not proof. A strong source check match is independent evidence from the file itself. For decisions that matter, combine the verdict, the source result and your own inspection.
The free checker accepts JPG, PNG, WebP, BMP, TIFF and HEIC images up to 4.5 MB, ten checks a day per visitor, with no account. A free account raises the file limit to 12 MB and gives 50 checks a month. Send the original file rather than a screenshot, because a screenshot loses the metadata the source check reads.
Yes for the verdict, no for attribution. The detector scores the pixels of any image, so a screenshot still gets a verdict and an AI score. The source check classifies the file as a screenshot or export and cannot name the original generator, because capturing the screen discards the original file's metadata and any C2PA manifest.
Verdict, AI score, and the generator's name when the file still carries it.