Booth 21-25 | AI Data Management Zone | Tokyo Big Sight
There’s a quiet result from Cheng et al. (Renmin University and Microsoft Research, 2026) that I keep coming back to. They gave large language models access to a minimal sandbox
Most AI text detectors look good in testing. They’re trained and evaluated on specific models, specific prompt styles, specific domains — and they perform well within that distribution. Then they get deployed into the real world, which has different models, different prompting patterns, different domains, and the performance drops. This
