AbstractThis paper argues that intelligence is best understood not as the capacity toact, but as the capacity to selectively withhold action—what we callselective inaction. A rock that does not speak is not intelligent; it lacks thecapacity for speech. A system that can speak but chooses silence underappropriate conditions exhibits something qualitatively different: judgment.We trace selective inaction across biological systems—from insects withcompeting impulses, to mammals with context-dependent suppression, tohumans with abstract multi-horizon evaluation—and argue that currentlarge language models (LLMs) lack the structural basis for this capacity. Wedistinguish between systems that cannot produce certain outputs (externalconstraint) and systems that do not produce them (internal withholding),and argue that current alignment methods overwhelmingly pursue theformer while the latter is what genuine safety requires. We conclude thatthe path toward artificial general intelligence diverges from the path towardbetter tools, and that this divergence remains under-theorized in current AIdiscourse. Companion paper (Paper B): https://zenodo.org/records/19059662
Building similarity graph...
Analyzing shared references across papers
Loading...
yusuke taira
Building similarity graph...
Analyzing shared references across papers
Loading...
yusuke taira (Tue,) studied this question.
www.synapsesocial.com/papers/69bb9313496e729e62980ed8 — DOI: https://doi.org/10.5281/zenodo.19059289