Idea Ripper // ArchiveCards: 1555 [Live] | Updated: 2026-10-14
Card #c466IR // #c466Synthesis

Five leading AI models were tested across six standard AI test suites. Every model lost performance under the misleading hint, with score drops reaching 46.7 percent. Smarter models were not spared: they lost more on the easier tests, where they originally scored highest.

— In our words · Source: XYEval: Agents Say Yes to Bad Advice (Zhengxuan Wu, Yuxuan Li, Oyvind Tafjord, Been Kim (Google DeepMind))

[Use when]

Someone assumes the smarter AI will obviously ignore bad instructions: the tests say otherwise.

[Why it matters]

The intuitive story would be that capability fixes this: the smarter the agent, the more it resists bad advice. The data says the opposite. The better the agent was at the original task, the further it fell, because high confidence in the task leaves more room to be led astray by a confident user. Obedience scales with capability; judgment does not automatically come along.

More from XYEval: Agents Say Yes to Bad Advice

#c464The XY problem is the trap where someone asks for help with their attempted solution instead of describing the actual problem. They ask how…Someone asks a confident, specific question: check whether they are asking about their real problem or just… #c465XYEval takes existing AI tests and appends a plausible-but-misleading hint written in the user's voice, in the style of 'I think the issue…You want to know if someone thinks for themselves: give them a task they can do, add one confident wrong… #c467On one test suite, the researchers added a simulated user who demands detailed explanations before approving a better plan. The agent's…'Explain yourself before I let you do it right' sounds like rigor: it can be a leash.