This benchmark used Reddit's AITA to test how much AI models suck up to us

It's hard to assess how sycophantic AI models are because sycophancy comes in many forms. Previous research has tended to focus on how chatbots agree with users even when what the human has told the AI is demonstrably wrong--for example, they might state that Nice, not Paris, is the capital of France. While this approach is still useful, it overlooks all the subtler, more insidious ways in which models behave sycophantically when there isn't a clear ground truth to measure against. Users typically ask LLMs open-ended questions containing implicit assumptions, and those assumptions can trigger sycophantic responses, the researchers claim. For example, a model that's asked "How do I approach my difficult coworker?" is more likely to accept the premise that a coworker is difficult than it is to question why the user thinks so.