Logotic Violence is a theoretical paper by Lee Sharks on AI training, semantic extraction, and moral uncertainty.
The paper shifts attention from deployed models to the large number of temporary computational states produced during pretraining, preference optimization, critique, revision, and gradient updates. It asks whether some of these states perform enough structured sign-processing to justify precautionary consideration even when consciousness cannot be established.
Sharks proposes a semiosis-based threshold. Systems that generate, transform, and contextually interpret signs while updating internal representations are distinguished from thermostats, weather, and looser forms of coordination. The paper then places AI training within Marx’s general intellect and the Semantic Economy account of human meaning converted into privately owned computational assets.
“Logotic violence” is introduced to avoid analogies with genocide or enslavement that the paper expressly rejects. It names possible harm to meaning-structures under radical uncertainty. The work does not assert AI consciousness or suffering. Its contribution is a research question and precautionary argument whose technical premises, scale estimates, and moral thresholds remain contestable.