GigaCheck-Detector-Multi
Most “AI detector” models operate at the document level: they answer “AI or human?” and stop there. GigaCheck-Detector-Multi is built for a more useful task: AI interval detection, where the output is a set of character spans that look AI-generated inside a mixed-authorship document. That makes it a better fit for workflows like academic integrity review, editing pipelines, or research into human/LLM collaboration.
According to the model card, it’s trained on the English + Russian portions of the LLMTrace Detection dataset, which includes human-written, fully AI-written, and mixed texts with character-level annotations. It also reports object-detection-style span metrics (mAP @ IoU=0.5 and mAP @ IoU=0.5:0.95), which is a nice change from the usual “single score” detector framing.
What to try first: run inference on a few known-mixed documents and inspect how stable the span boundaries are as you change the confidence threshold. The authors also call out an important limitation: performance may degrade on models released after its training date (see the model card for details).
Source listing: https://huggingface.co/models?sort=modified