100% Free and private AI detector.

GPTgone is a completely free AI detector for text and images. Paste or upload some text and it estimates whether it was written by a person, by an AI model such as ChatGPT or Claude, or by a mix of both, and it highlights the sentences that look most machine-written and explains why. Give it a picture and it checks whether it was made with AI, from the labels AI tools leave in their files and the traces image generators leave in the pixels.

Yes. Open the Image tab, or just drop, paste or upload a picture anywhere on the scanner. GPTgone first reads the hidden labels many AI tools write into their files (content credentials and similar tags) and names the tool when the file does. Then a vision model looks at the pixels for the traces image generators leave. You get a verdict, a likelihood and a list of what was checked, in seconds and on your own device.

On 5,134 test images it was never trained on, it caught 71.1% of AI images from 12 generators released in 2025 and 2026, from 82.8% for Ideogram 3.0 to 66.2% for Z-Image Turbo among the generators it never saw in training. It called 0.8% of camera and phone photos AI generated and flagged 3.6% of them in any way. Photos posted in photography communities, which are often heavily edited, are harder: it flagged 8.8% of them and called 3.4% AI generated. It flagged 1.9% of human-made art and 1.1% of screenshots and graphics. Pictures that were re-saved for social media are harder: 61.2% of those AI images were caught. When a file still carries an AI label, GPTgone reads it and says so. Every number is in our accuracy report. Treat any result as one piece of evidence, never as proof.

We measured it on images from 12 generators released in 2025 and 2026: GPT Image 1.5 and GPT Image 2 (OpenAI), Nano Banana Pro and Nano Banana 2 (Google), Midjourney v7, FLUX.2, Seedream 4.5 and 5.0, Imagen 4, Qwen-Image, Ideogram 3.0, Z-Image Turbo and ERNIE Image, and on older ones: DALL·E 3, Stable Diffusion 1.5 and XL, Midjourney v6, Recraft v3 and Adobe Firefly Image 3. The visual check is built on a model trained on images from 4,800+ AI generators. It doesn't tell generators apart by their pixels; when a file's label names the tool, GPTgone shows it. The results for each generator are in the accuracy report.

Yes. Every feature is free for everyone, with no account, no trial and no paywall. Scan as many texts as you like, up to 1,000,000 characters each, and check as many images as you like. Because the detectors run in your browser instead of on a server, there's nothing to meter and nothing to sign up for.

GPTgone runs two models in your browser. The main one is a compact neural network (a transformer) that we fine-tuned on more than 110,000 labelled human and AI texts. It reads your text in overlapping windows of about 120 words and scores how machine-like each one is. The second is a statistical model over 39 writing signals, such as how much sentence lengths vary, word choice, contractions, hedging and punctuation. GPTgone combines both into a calibrated score for the whole text and for every sentence, and shows you the signals behind it.

On more than 25,000 held-out test documents from 11 sources, GPTgone caught 89.6% of AI-written texts and flagged 2.2% of human ones; only 0.5% of human texts were called fully AI. Our previous rule-based detector caught 59% while flagging 27% of human writing. Harder cases score lower: brand-new models it never trained on, AI text run through paraphrasing tools, and AI drafts a person has edited. Every number, including the weak spots, is in our accuracy report. It works best on 150+ words of prose. Treat any result as one piece of evidence, never as proof.

Fairer than most, but not perfect. A 2023 Stanford study found that popular detectors flagged 52% to 76% of essays written by non-native English speakers for the TOEFL as AI-generated. On the same essays, GPTgone flags 11%, and none of the 88 essays by US eighth-graders from the study. That is a big improvement, but it is still not zero, which is one more reason never to treat a result as proof on its own.

Yes. Both detectors run entirely in your browser, so the text you paste or upload and the pictures you check are never sent to us or anyone else. Your first text scan downloads the text model from this site (about 20 MB) and your first image check downloads the image model (about 48 MB), once each; after that they are cached on your device. The site's security policy blocks the page from sending data to any other server. There are no accounts and no scan history.

GPTgone learned from writing by more than 40 models: OpenAI's GPT-3.5 to GPT-5, Anthropic's Claude 3 to Claude 5.5, Google's Gemini 3, Meta's Llama, plus DeepSeek, Qwen, Mistral, Cohere, GLM and others. We also test it on models it never saw in training, because new ones appear all the time. It doesn't try to name the exact model, only whether the writing reads as machine-generated.

For text, English. The text models were trained on English, so results for other languages aren't reliable, and GPTgone tells you when a text doesn't look like English. Images work whatever language appears in them.

Sometimes, like every detector. Text that a person heavily rewrites by hand is the hardest case. GPTgone was trained on AI paraphrases and casual-sounding AI writing, and it catches common tricks such as invisible characters and look-alike letters from other alphabets. When it finds them, it says so in the result. With images, screenshots, heavy compression and edits weaken the traces the visual check reads, and the hidden labels can be stripped.

Please don't rely on it alone. Use a high AI score as a reason to look closer: talk to the writer, ask about their process and compare with work you know is theirs. The sentence highlights are there to make that conversation easier.

Images: JPG, PNG, WebP, AVIF, GIF (the first frame) and BMP up to 50 MB, plus HEIC photos in browsers that can open them, such as Safari (elsewhere, save the photo as a JPG first). Drop, paste or upload a picture anywhere on the scanner and it goes to the image detector.

Text: PDFs, Word documents (.docx, or .doc from older versions of Word) and plain text files (.txt and .md). The file is opened right in your browser, so it never leaves your device. GPTgone reads the main text and leaves out headers, footers, page numbers, footnotes and comments. PDFs don't store paragraphs, so GPTgone rebuilds them from the page layout; it's worth a glance at the text before you scan. Scanned PDFs are pictures of pages with no text inside, so they can't be checked. Each scan checks up to 1,000,000 characters (about 150,000 words), so a longer file is trimmed to its first 1,000,000. Very long texts take a while: the full 1,000,000 characters can take a few minutes on a computer and longer on a phone, and you can switch to a quick scan while it runs.