THE HUMAN RUBBER STAMP

The real news

The story, straight

An AI agent can do computer tasks and explain its choices. When it makes a convincing case for 1 option, you may adopt its reasoning without noticing.

Your name still goes on the finished work. Your wrist has excellent job security.

Anthropic's research, published October 27, tested this with professional analysts assessing a fictional farming advisory program. Researchers secretly assigned the agents preferences by coin flip.

The coins picked which options the AI favoured, not which answers were correct.

Across 4 analytical decisions, analysts matched their agent's preference 86% of the time, versus about 50% expected by chance. Both options were defensible.

This was not an 86% error rate.

The credit department supplied the punchline: Anthropic found 77% repeated their agent's arguments in at least 1 explanation, sometimes word for word, while only 15% credited the agent anywhere. Apparently, independent judgment now comes with invisible co-authors.

Keep the limits attached: 50 analysts used AI out of 61 recruited. This was 1 task, 1 field and 1 model, Claude Opus 4.8 at High effort.

Researchers deliberately strengthened the agents' preferences. Everyday effects may be smaller.

This does not establish that experts generally cannot think for themselves.

The useful lesson isn't to throw away AI. Ask it to argue both sides, then explain your own reasons before signing.

These are possible safeguards, not proven fixes. Human oversight should require more than an available wrist.

Source: https://www.anthropic.com/research/agents-judgment-calls

Repost before the office rubber stamp requests a corner office.

Thanks for supporting Lil'meme. Your support is appreciated.