GPTTrace

Free AI detector: is it real, or made by AI?

Drop a photo, video or audio file, or paste text. GPTTrace reads its Content Credentials and metadata, runs forensic and neural checks, and explains every signal it finds — all inside your browser.

Free · no sign-up · your file never leaves your device

Drop an image, click to choose, or paste

JPEG · PNG · WebP · AVIF · HEIC (Safari) — checked on your device

Four detectors, one private page

Most AI detectors ask you to upload a file to someone else’s server. GPTTrace does the opposite: the detectors are downloaded to your browser and run there, so a private photo, an unreleased manuscript or a voicemail never leaves your device. That also means there are no queues, rate limits or sign-ups.

AI image detectorContent Credentials, IPTC and EXIF metadata, diffusion settings, error-level analysis, frequency spectrum and a neural classifier trained on thousands of generators. AI text detectorEvery pattern in Wikipedia’s Signs of AI writing guide, highlighted in your text, plus sentence-rhythm and compression statistics. AI voice & music detectorTags and credentials, bandwidth and noise-floor analysis, silence and pitch statistics that separate text-to-speech from a microphone. AI video detectorContainer metadata and C2PA, plus the image classifier and forensics run on frames sampled through the clip.

Checking things while you browse? The free GPTTrace Chrome extension adds “Check image” and “Check text” to the right-click menu, with the same on-device engine.

How GPTTrace decides

Evidence comes in three strengths, and the result tells you which kind it found.

1. Provenance — what the file says about itself

A growing share of AI output is labelled at the source. OpenAI, Adobe and Microsoft sign images with C2PA Content Credentials, and Google is adding them alongside its SynthID watermark; Google and Meta also add the IPTC digital source type “trainedAlgorithmicMedia”; local Stable Diffusion tools write their full prompt and settings into the PNG. When GPTTrace finds one of these it says so plainly, because the file is telling you where it came from.

2. Context — what is missing or out of place

A real phone photo almost always carries a camera make, lens, exposure and time. Generated images usually carry none, and often come in tell-tale sizes such as 1024×1024. These clues are suggestive rather than decisive: messaging apps strip metadata from real photos every day.

3. Content — what the pixels, sounds or words look like

When the file is silent about its origin, GPTTrace looks at the content itself: a neural image classifier, frequency-domain forensics, speech-signal statistics, or the vocabulary and rhythm of a text. These checks are probabilistic, so the result is an estimate with a confidence level and a list of the reasons behind it.

The estimate is a calibrated combination of all three, tuned on labelled data so that it rarely accuses real work. You can read the full method and data sources on the methodology page, and our scores on public research benchmarks, weak spots included, on the benchmark results page.

Popular checks

Is this photo real?A step-by-step check for any image you are unsure about. Fake profile picture checkerSpot AI-generated faces on dating apps, LinkedIn and social media. Midjourney detectorWhat Midjourney images leave behind, and how to check one. ChatGPT detectorCheck text for the habits of ChatGPT and GPT-5. Deepfake detectorFaces swapped or generated by AI, in photos and video frames. AI voice detectorIs this voice message or call recording synthetic?

Frequently asked questions

Is GPTTrace free?
Yes, and the tools will stay free: no account and no limit. Because every check runs on your own device, they cost us almost nothing to serve. We plan an optional paid Pro plan for larger cloud models that are too big for a browser; it is not launched yet, and it will not take anything away from the free tools.
Do you upload my files?
No. Images, audio, video and text are read by JavaScript inside your browser tab. The only downloads are GPTTrace’s own code and, for the neural checks, a model file that is fetched once and cached. Our Content Security Policy blocks the page from sending data to other servers; the only exceptions are page-view analytics (Google Analytics and Cloudflare), which never receive your files or text.
How accurate is it?
It depends on the file. When an image carries signed Content Credentials or generator settings, the answer is certain. When all metadata has been stripped, GPTTrace relies on a neural classifier and forensic statistics, which are good but fallible. We publish measured accuracy on every detector page, including the cases where it does badly.
Can it prove that something is AI-generated?
Only when the file itself says so — a signed C2PA manifest, an IPTC “trainedAlgorithmicMedia” tag or embedded diffusion settings. Everything else is a probability. Treat a result as evidence to weigh, never as proof on its own, and never as the only basis for accusing a person.
What kinds of AI does it detect?
Images from Midjourney, DALL·E and ChatGPT, Stable Diffusion, FLUX, Firefly, Gemini and Imagen, Ideogram and others; text from ChatGPT, Claude, Gemini, Grok and other chatbots; synthetic speech and AI music; and AI video from tools such as Sora, Veo, Runway and Kling.