SlopTotal launches an open-source, local AI text detector with 23 engines

Summary

SlopTotal is an open-source, self-hosted tool for analyzing whether text was generated by AI. Users can paste text, scan a URL, or upload PDF, DOCX, TXT, and Markdown files; the system runs 23 detectors in parallel, including DeBERTa and RoBERTa classifiers, perplexity and token-rank methods, and linguistic heuristics. A calibrated ensemble combines the results, while the interface exposes each engine's score, agreement, and explanation. The project is designed to run on a user's own CPU without third-party AI APIs or tracking; a 4 GB RAM lite profile is supported, and a first scan downloads about 2 GB of model files. It also provides paragraph-level heat maps, a JSON API, a Chrome extension, and a site checker that looks for deployment fingerprints left by builders such as Lovable, v0, Bolt, Base44, Replit, and Same. In its published evaluation, the project used a 110-text RAID corpus containing human and AI passages from several domains, plus 26 pre-1920 literary passages to test false positives. The reported overall AUC was 0.974, with one of 66 human texts labeled “Likely AI” and none of the 26 literary passages flagged. A September 2026 remeasurement on 180 fresh RAID texts reported an AUC of 0.979, with one of 40 human texts called “Likely AI.” The project says short text below about 80 words, heavily edited AI writing, and source code are unreliable detection targets. It publishes failures, including engines that scored backward or carried weight despite loading random networks, and advises against using any detector as the sole evidence for an accusation. Reports are retained for 30 days by default, and the software is released under the MIT license, while model weights retain their own licenses.