DetectGPT alternative: work out which DetectGPT you mean first

DetectGPT was introduced in "DetectGPT: Zero-Shot Machine-Generated Text Detection using Probability Curvature" by Eric Mitchell, Yoonho Lee, Alexander Khazatsky, Christopher D. Manning and Chelsea Finn, submitted to arXiv on 26 January 2023 and revised on 23 July 2023.

The idea is compact. Text sampled from a language model tends to sit in regions where that model's log probability function has negative curvature. So you take the candidate text, make small perturbations of it with a second model, and compare log probabilities. If the original scores much higher than its perturbed neighbours, the text likely came from the model you are testing against. No classifier training and no labelled dataset are needed, which is what "zero-shot" means here.

The reference implementation is public at github.com/eric-mitchell/detect-gpt under an MIT licence. Its last push was 4 March 2023, so treat it as a research artefact rather than maintained software.

The practical catch is built into the method. It needs log probabilities from the model you suspect wrote the text. For an open weights model you can compute those. For a current commercial model behind an API you generally cannot, and using a different model as a proxy weakens the signal. That is why the method, as published, is not something a website can simply run against arbitrary text from any AI writer.

The homepage we read carries a wall of quoted posts from X. Several of them describe the research method rather than the paid service, including one that refers to "the curvature of the model's log probability" and others that repeat the paper's early 95 percent figure. The page we read does not state a relationship between the site and the paper's authors.

That matters for a practical reason, not a moral one. If you arrived at detectgpt.com after reading about the Stanford method and assumed you were using it, you were evaluating a different system. Judge the product on its own published behaviour.

TextSight does one thing rather than six. The paste box runs without an account: 3 checks a day for anonymous visitors, pooled across the AI Detector and the Document Detector, capped at 1,000 words per check, with a 25 word minimum. A free account moves that to 60 checks a day at 5,000 words, and signed in scans are saved to Scan History on every plan including free.

Scores are shown as bands, not verdicts. Under 30 reads as likely human written, 30 to 69 as mixed signals, 70 and above as likely AI generated. Sentence level analysis and PDF export start at Starter, $9.99 a month or $7.49 billed annually. Pro is $19.99 or $14.99 annually and adds the Chrome extension and REST API. Business is $39.99 or $29.99 annually and adds teams and branded reports. Enterprise is $79.99 or $59.99 annually. File upload, URL scan and bulk scan are part of the product. Starter allows 200 checks a day, Pro and above are unlimited.

The honest limits: our detector is an English only ensemble of DeBERTa, RoBERTa and ELECTRA. It carries a known false positive bias against writers who learned English as a second language, which is exactly the failure mode that gets students accused unfairly. We have no plagiarism database, no fact checker and no mobile app. If those are why you subscribed to detectgpt.com, we do not replace it.