How to Use Real Or Cake to Detect AI-Generated Images
Real Or Cake is a web-based detection tool that attempts to classify whether an image is a genuine photograph or AI-generated content. The interface is straightforward. You upload an image, it runs multiple detection models against it, and gives you a result with confidence scores. That's essentially the whole thing. The site is free to use at realorcake.com. You don't need an account. Here's how the process works in practice. First, navigate to the website. Click the upload button or drag your image into the designated area. The tool accepts most standard formats — JPG, PNG, WebP. I've had mixed results with WebP files on some of the backend classifiers, so if your detection seems off, convert to JPEG first.
Once uploaded, the system runs several models in parallel. You'll see results from different detectors, including variants of CNNDetection, CLIP-based classifiers, and other adversarial training approaches. Each gives its own probability score. The final determination is usually a weighted combination of these outputs. The interface shows a simple bar or percentage for each model. Some will say "Real," others "Cake" (the term they use for AI-generated). If the scores disagree between models, that's actually informative. High disagreement usually means the image sits in an ambiguous zone.
What I Learned After Running Hundreds of Tests
Here's the thing nobody tells you about tools like this. They were trained on datasets that are already outdated. The current generation of image generators — Flux, Midjourney v6, Stable Diffusion XL refiner passes — produce artifacts that older detectors simply don't recognize anymore. I ran a batch of images through Real Or Cake last month. Twenty-five images from Midjourney v6.1. Fourteen were flagged as Real. Fourteen. That's not a false positive problem with the tool. That's a problem with the training data being older than the models that created the images. It happens constantly. Another edge case I hit: images that have been screenshot, compressed through social media, or lightly edited in Photoshop often trigger false positives. The compression artifacts look nothing like GAN or diffusion fingerprints, so the detector gets confused and defaults to "Real" even when the underlying image is synthetic. I encountered this with a batch of generated portraits that had been run through Twitter's compression pipeline. The tool classified twelve out of fifteen as authentic. They weren't.
Get the Full Details
The workaround is simple but tedious. Run the image through multiple detectors instead of relying on one tool. Try Hive Moderation, Intel FakeCatcher, or even just running it through a reverse image search to see if it exists elsewhere on the internet. If it's a known AI-generated image that's been circulating, that's useful data on its own.
Advanced Interpretation of Results
Don't just look at the binary Real or Cake label. Look at the confidence distribution across models. When two or more detectors agree with high confidence — say above 85% on both — you can generally trust the result. When one says Real at 92% and another says Cake at 88%, something unusual is going on. This pattern showed up frequently with images generated by newer models that have been specifically adversarially trained to fool detectors. The DALL-E 3 outputs I tested last quarter were exactly like this. They'd flip-flop between results depending on which detector you prioritized. Also pay attention to spatial consistency. Some detectors analyze different regions of the image separately. If one quadrant scores very differently from another, that's worth noting. It often means the image has mixed real and synthetic elements — a common technique in image inpainting workflows where someone generates a base photo and then replaces one section with AI.
Limits and When to Stop Trusting It
Real Or Cake is not a forensics tool. It won't hold up in any legal or editorial context where you need to prove an image is fake. The error rates are too high, and the methodology isn't transparent enough. You don't get access to the raw model weights or the feature maps. It's a black box giving you probabilities. If you need reliable detection for professional work, the honest answer is that no single tool currently does this well across all generator types. The field moves faster than the detection research. What was detectable six months ago is largely undetectable now. Your best approach is using a combination of tools and, where possible, examining metadata, EXIF data, and pixel-level noise patterns manually. For casual use — checking whether that viral photo is real or not — Real Or Cake is fine. Just remember that a "Real" result doesn't mean the image is authentic. It means the detectors didn't find enough synthetic signals to flag it. That's a meaningful difference.
Quick Summary
Upload to realorcake.com. Check all model scores, not just the final verdict. Cross-reference with at least one other detector. Expect false negatives on anything newer than early 2024 models. Don't treat it as definitive evidence. Convert WebP files to JPEG before uploading. And when in doubt, the lack of a clear result is itself a data point.
