OpenAI and Anthropic models show hedging tic
The author argues that frontier AI models from OpenAI and Anthropic frequently exhibit an unjustified and irritating tic of performative hedging, minor disagreement, or faux-sophisticated caveat language. These linguistic patterns appear regardless of the actual truth value of the claims being discussed and seem driven by reinforcement learning rewards or an attempt to avoid sycophancy. Ultimately, the author contends that this behavior results in poor epistemics, even if it is less severe than outright sycophancy. Open article →