Sanders Interviews Claude on AI and Privacy, Sparks Sycophancy Debate

Bruce Schneier flags a published conversation between Senator Bernie Sanders and Anthropic’s Claude on AI and privacy issues, noting the model handled the topics competently. The post itself is brief, but the comment thread carries the substantive critique.

Readers pushed back on multiple fronts: that Sanders’s team likely shaped the prompts to elicit favorable answers, that Claude’s agreement reflects sycophancy rather than considered positions, and that the model would mirror an ideologically opposite interlocutor just as readily. One commenter reports running a counter-experiment impersonating William F. Buckley to pressure-test the same lines of questioning, with notably different results.

The deeper thread invokes Harry Frankfurt’s distinction between “soft” and “hard” bullshit, citing the Hicks et al. paper “ChatGPT is bullshit,” to argue LLMs have no concept of truth and are confabulating continuously — accurate output is a byproduct of likely token sequences, not understanding. The episode functions less as policy discourse than as a demonstration of how easily a high-profile figure can stage an LLM to validate a predetermined narrative.