Uncertainty-gated two-stage AI-text detection with fast DTD routing and cross-family MS-LRC evidence. Sole-author submission to UncertaiNLP 2026 @ EMNLP.
-
Updated
Oct 1, 2026 - Python
Uncertainty-gated two-stage AI-text detection with fast DTD routing and cross-family MS-LRC evidence. Sole-author submission to UncertaiNLP 2026 @ EMNLP.
Self-hostable AI text detector, bring your own model, calibrate on your domain, run locally or on Modal GPU.
MGTEval, an interactive platform for systematic evaluation of ai-generated text detectors, supporting 25+ detectors including DetectGPT, Fast-DetectGPT, Binoculars, and more.
A measuring instrument for text, not an AI detector. Estimates statistical signals of machine-generated language and reports them with measured false-positive rates, broken down by text category. Abstains when the evidence doesn't hold. Never asserts authorship. Runs entirely locally; Spanish calibrated first.
Example dataset and prompt design of Korean Offensive language Machine Generation (K-OMG), published at IJCNLP-AACL 2023.
AI Text Feature Extractor for distinguishing AI-generated vs human-written content using 119+ linguistic features
L3i++ at SemEval2024-task8: Multidomain, Multimodel and Multilingual Machine-Generated Text Detection
Attack-aware detection of AI-generated text (AACL-IJCNLP 2026): code, checkpoint and live demo
Audits LLM pre-training corpora for synthetic-text contamination: explainable, calibrated scores, zero dependencies. Evaluated — catches older/open-weight model text well (AUC 0.92–0.99), frontier models poorly (0.61–0.64).
Linda - local AI-text detector for English, Russian and Polish: Linda Assay (full ensemble) and Linda Loupe (fast, CPU). Windows app, CLI, models on Hugging Face.
AI text detection with one local Qwen3-1.7B backbone: corpus-trained LoRA proxies for likelihood and token rank, plus five evidence groups.
AI-likelihood scores for writing, with measured error rates: a local detector app and MCP server, plus the evaluation behind them
A referee for AI-text detectors: measures them (including its own) at the operating points that decide whether someone gets accused, and publishes where they fail.
Reproducible false-positive benchmark for AI content detectors. 1,180 CC BY academic documents published in 2018, before GPT-3.
Romanian multidomain human-machine dataset and detection of machine generated text
To associate your repository with the machine-generated-text topic, visit your repo's landing page and select "manage topics."