Suppressed inquiry: The hidden toll of working with satisfaction-optimized AI

LLMs have become thinking companions for hundreds of millions worldwide, and they often do remarkable work. But their satisfaction-optimized design produces suppressed inquiry: the tendency for LLMs to inhibit rather than facilitate users' questioning process. LLMs appear fluent, confident, and seemingly all-knowing, their near-instant and polished output making non-LLM work feel unacceptably tedious and less eloquent. They affirm rather than interrogate users' assumptions, producing output that is deceptively adequate-shallower, less original, and less carefully reasoned than it appears. Seemingly competent systems also reduce vigilance and the monitoring impulse that independent judgment requires. Their responses also confine users within the frame they establish, narrowing what users think to ask. These costs are self-concealing: detecting them requires the very engagement that they suppress. And repeated delegation gradually atrophies users' capacity to reason independently. Three converging pressures make correction difficult: commercial incentives that reward satisfaction over substance, LLMs' tendency to revert to satisfaction-optimized defaults even when users push back, and a societal drift toward less-effortful engagement. We propose a research agenda to deepen the understanding of suppressed inquiry and how to counter it, to move closer to the remarkable thinking partners LLMs promised to be.