nettoandtetto/osu-ai-detector

Osu-ai-detector is a research tool that examines beatmaps, the files that define note patterns for the rhythm game osu!, and looks for signs that a beatmap may have been generated by an AI mapping tool rather than created by a human. It runs entirely on the user's own computer and is explicit about being a research aid rather than a definitive verdict machine. The tool performs four separate, independent checks rather than combining them into a single score. One check looks at whether the file has been explicitly disclosed as AI made or matches known historical AI generated files exactly. A second check compares the beatmap against the internal generation process of a known AI mapping model called Mapperatorinator, to see how closely it agrees with that model's outputs. A third check looks for mechanical traces in the beatmap's underlying data, such as coordinates, timing, and slider velocity, that tend to appear in machine generated files. A fourth check compares the beatmap's rhythm, movement, and structural choices against a reference set of human made beatmaps. The README stresses that the tool does not produce one overall verdict, and that an anomaly score is not the same as a probability that a beatmap was AI made. The software is built for Windows 10 or 11 on a 64 bit PC, and strongly recommends an NVIDIA graphics card with at least 8 gigabytes of video memory, since running on the processor alone can be extremely slow for one of the checks. Installation is handled by a PowerShell script that sets up an isolated Python environment through Miniforge and downloads several pinned AI model files needed for the checks, verifying each one against a recorded checksum. An offline installation path is also provided for computers without an internet connection, built ahead of time on a separate connected machine. Separate repositories hold the training code and research data used to build the detection models, since this app repository intentionally does not include the material used to train them. The application's own code is released under the MIT license, while the AI model files it downloads have their own separate licensing terms.