Step 1 — Read the file’s labels
Many AI tools now write a label into the files they create. We look for:
- Content Credentials (C2PA) — a signed record used by OpenAI (ChatGPT, Sora), Adobe Firefly, Microsoft, Google, Meta, TikTok and camera makers like Leica, Sony and Nikon.
- IPTC digital source type — the standard “trainedAlgorithmicMedia” tag (AI-generated) and “compositeWithTrainedAlgorithmicMedia” (AI-edited).
- Generator settings — Stable Diffusion, ComfyUI, NovelAI, InvokeAI and Midjourney often leave the prompt and settings in the file.
- AIGC labels used by many Chinese AI apps.
- Camera data — make, model, exposure, aperture and ISO from real cameras.
If an AI label is found, the result is marked confirmed. If a real camera signed the photo, it is confirmed real.
Step 2 — Check the pixels (or words) with several AI detectors
Labels are easy to remove (a screenshot does it), so we also run open-source detection models from the Hugging Face community. Our main image and text detectors run directly on our own server; extra detectors join when available, and if they disagree or are unsure, another one breaks the tie. Their opinions are combined using a weighted average in log-odds — a standard way of merging probabilities that stops one over-confident model from dominating.
For videos, your browser extracts six frames: pattern detectors check several of them, and a vision AI reviews all six in order, looking for glitches between frames, impossible physics and staged scenes. One clearly-AI frame counts for more than a blurry one. For text, we split long pieces into sections, score each, and highlight them.
Step 3 — Combine everything into one honest answer
Smaller signals nudge the score up or down: camera exposure data (towards real), filenames from AI tools (towards AI), and exact output sizes that AI generators use. The final AI-likelihood becomes one of six verdicts, with a confidence level that drops when signals conflict, when only one detector answered, or when the file is a video.
| AI likelihood | Verdict |
|---|---|
| 85–100% | Likely AI-generated |
| 60–84% | Possibly AI-generated |
| 41–59% | Hard to tell |
| 16–40% | Probably real |
| 0–15% | Likely real, not AI |
Why it’s free
We use open-source models and cache results for images that have already been checked (viral images get scanned a lot), so the cost per scan stays tiny. No ads in your results, no sign-up, no paywall.
Frequently asked questions
How accurate is it?
When a file still carries Content Credentials or other AI labels (common for ChatGPT, Gemini, Firefly and Stable Diffusion files) the answer is essentially certain. Otherwise we combine several open-source detectors; they are right most of the time on unedited AI images but can be fooled by screenshots, heavy compression, filters and brand-new generators. That is why every result shows a confidence level and the evidence behind it — treat it as strong evidence, not proof.
Do you keep my photos, videos or text?
Only briefly. What you check is kept for at most 24 hours (so we can investigate wrong results and abuse) and then deleted automatically. For videos, your browser pulls out a few frames for the check; the full video is then uploaded in the background so we can review results, and is deleted with everything else. Nothing is published unless you tap Share, and then only a small preview is saved.
Is IsItMadeByAI really free?
Yes. Every scan is free, there is no sign-up, no credit card and no watermark on your results. The site runs on open-source detection models, which is how we keep it free.