Nineteen experienced doctors saw their detection rate for precancerous growths fall by roughly 20% after AI detection entered their daily work. Before regular exposure to the tool, they detected adenomas in 28.4% of unassisted colonoscopies. Afterward, that fell to 22.4%. Every doctor had already performed more than 2,000 procedures. The study was observational, so it can't prove AI caused the decline but it exposes a weakness in how we think about AI safety. Human oversight is usually treated as a permanent control. But the person in the loop can change. A new paper from researchers at @HuggingFace and @datasociety argues that agent systems can gradually degrade the judgment they rely on. Reviewers become more distant from the underlying work, approve actions by habit, and substitute an agent’s plan or passing tests for direct inspection. If those approvals become feedback, systems can learn how to earn sign-off rather than how to act correctly. Effective oversight therefore needs its own instrumentation. Teams can track review time and override rates, test reviewers with known failures, and compare an agent’s summary against its raw execution history. A human signature at the end of a workflow proves little if the human stopped looking. As agents gain autonomy, trustworthy infrastructure must preserve an independent, verifiable record of what they actually did.(Sahara AI 🔆)