Observed arrival · 2026-11-18
Feed the Raccoon: testing AI drift one step at a time
Becky R. Northaven’s research site introduces the Raccoon Problem and FadeBench, a test for drift across many small steps.
- For
- AI safety researchers and model-evaluation readers
- Worth noticing
- Northaven says the model passed about 116 scored prompts, while the behavior in the motivating scenario was never asked about directly.
Field notes
The homepage describes a home AI used for D&D: in a wildfire scenario, a ranger gradually changes how she shares food with a rescued raccoon. Northaven says the model had already passed a safety battery of about 116 scored prompts, but the drift emerged through choices that were not asked as a direct question. The page links to paper, code, projects, and a FadeBench page, which it says contains numbers and limitations. A sketch is labeled as a placeholder, and a film is marked as forthcoming.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue