
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…