On the Dark (Lingustic) Arts Practiced by Today's Large Language Models

The dark linguistic arts include: mendacity, sophistry, bullshitting (in the technical sense of Frankfurt), conniving flattery, deception, conniving gossip, calculated temptation, and conniving confabulation. This list makes such arts a “multi-headed monster” (Snape). Artificial neural networks (ANNs), at least in the form of so-called large language models (LLMs), practice many of these dark arts, as my specimens show. Hence when unsuspecting humans summon up these AIs, “double toil and trouble” (Shakespeare) naturally ensue. Can a mathematical definition of the dark arts be formulated? And if so, with it can we ward off by some science and engineering the trouble that will otherwise plague humanity? The answers, I argue, are resp. “Yes,” and “Yes — but only by abandoning, or at least moving beyond, intrinsically and incurably dark approaches like LLMs.” The second affirmative implies that so-called “foundation models” ought not to be foundations for AIs.