Real or fake checks
Is This Photo Real? Check a Picture for AI Generation or AI Edits
Is this photo real or fake? Upload it and you get one of eight verdicts, from Real to Certainly AI Slop, and the evidence that produced it. We do not start with a guess. Slop or Not reads the signed Content Credential first, then looks for a SynthID watermark in the pixels, and only when neither settles the file does it score the picture with a classifier trained on millions of human and AI images. Online checks stay away from outside AI companies.
The only AI image detector that reads Google DeepMind's SynthID watermark from both Google and OpenAI images.
Google's own detector checks only Google's images. OpenAI's checks only OpenAI's. Slop or Not reads both.
What it checks
Real, AI-Generated, AI-Edited, or Too Close to Call
A signed C2PA credential, or an unsigned IPTC declaration, that names the generator or declares an AI edit of an existing picture, as ChatGPT and Gemini edits do
An invisible SynthID watermark in the pixels of a Google or OpenAI image, which survives the screenshot that strips a credential
A pixel score from the image model when nothing readable remains, reported on a seven-step scale from Real to AI Slop
Good Uses
- Settling a real-or-fake argument about a viral photo before you share it
- Checking a profile or listing photo that looks a little too clean
- Finding out whether a real photo was edited by ChatGPT or Gemini before it reached you
Limits to Know
- A screenshot or a re-save strips the credential layer, so the check falls back to the watermark and the pixels
- Manual retouching in Photoshop or Lightroom leaves no AI record, and Slop or Not does not run error-level or forgery forensics
- Real means no AI evidence was found, not proof of who took the picture or that the scene happened
FAQ
Real or Fake Photo Checker Questions
Is this photo AI generated, AI edited, or real? How does the verdict tell them apart?
It depends on which layer answered, and the verdict tells you. A trusted signed credential that declares AI media settles the file as Certainly AI Slop, and the credential line under the verdict says which kind: Trained Algorithmic Media means the picture was generated, Composite With Trained Algorithmic Media means someone took a real picture and edited it with AI. A SynthID watermark says the picture came out of a Google or OpenAI model, nothing more. When neither is readable, the image model scores the pixels and you get a probability band instead: Real, Likely Real, Leaning Real, Inconclusive, Leaning AI, Likely AI, or AI Slop. Inconclusive is the honest one. The score sits at the decision line, so treat it as no result, not as a soft yes.
Can you tell if a picture has been edited by ChatGPT?
Yes, as long as the edited file still carries OpenAI's proof. ChatGPT writes a C2PA Content Credential into the pictures it edits, and that credential declares a composite of real and AI media rather than a generation from nothing. So Slop or Not flags the picture as AI and shows the edit declaration beside the verdict. Gemini edits leave a record that names Google, and it reads the same way. Here is the catch. Send that picture through a screenshot or a chat app and the credential is usually gone. The SynthID watermark in the pixels is then the only provider signal left, and it cannot tell you whether the picture was generated or edited.
Does screenshotting remove AI metadata?
Yes, and this is the part most people get wrong. A screenshot is a new file, so the C2PA credential and the IPTC fields of the original never make it across. SynthID is different. Google and OpenAI write it into the pixels themselves, so a screenshot of one of their images still carries it and Slop or Not still reads it. With no credential and no watermark, the image model is all that is left, and the answer turns into a probability.
Is this photo AI if the verdict says Likely Real?
Probably not, but we chose the word Likely for a reason. Likely Real means the pixel score landed well on the human side of the model's line and no AI evidence turned up. That is evidence, not proof. An AI picture with its watermark stripped and heavy post-processing on top can drift toward the human side, and per internal tests the image model reads about 90% of non-watermarked AI images correctly, not all of them. Results vary with new AI models and methods designed to trick detectors.
Does the check send my photo to an AI company?
No. In the browser the picture travels to Numen's private Mac server, gets checked there, and is never passed to OpenAI, Google, or any other AI company. The iPhone, iPad, and Mac apps run the same check on-device, so with the app the photo never leaves the device at all.
Run the Next Check on Your iPhone or Mac
The Slop or Not app checks images and text on-device, so nothing you check ever leaves your device.
More from Numen
Two more apps that run on your device. Neither needs an account.
Private LLM
A Local AI chatbot for uncensored open-source models that runs offline on iPhone, iPad, and Mac. Private LLM uses OmniQuant and GPTQ quantization tuned per model, so its 3-bit models match the 4-bit models other apps ship.
Clean Links
Most tracking now hides behind link redirects. Clean Links unwraps short links and strips every tracking parameter from a URL before you share it, on iPhone, iPad, Mac, and the web.
Related checks
Compare Related Detector Workflows
AI Watermark Detector
Use this AI watermark detector and checker to read image SynthID, verified C2PA Content Credentials, and IPTC metadata before any pixel check.
C2PA Detector
Use this C2PA detector to check Content Credentials. Slop or Not reads signed provenance, then checks pixels when no provenance or watermark resolves the file.
Deepfake Photo Detector
Use this free deepfake detector to check photos, portraits, and profile pictures for AI-generated faces. Online checks run on Numen's private Mac server.