# Resolution control and LLM philosophy A diagnosis of LLM failure modes in philosophy: the problem is often not falsehood but *resolution* — choosing the right level of specificity and commitment. ## The Two Competencies LLM philosophy requires: 1. **Discourse competence** — generating recognisably philosophical moves (distinctions, objections, reconciliations, charitable reconstructions) 2. **Resolution control** — choosing the right level of specificity: committing to definitions, quantifiers, scope, inferential strength, and answer-shape LLMs are strong at (1) because training data is saturated with philosophical and argumentative writing. They're weaker at (2) because nothing in the default objective forces them to pick a resolution and stick to it. ## Why "Plausible but Untrue" Isn't the Main Problem In philosophy chat, the model can avoid being "wrong" by never sticking its neck out: > The problem often isn't "it's making false claims"; it's "it's not being sharp enough to be testable". This explains why the dominant failure modes are: - **Imprecision** — slippery terms, untracked sense shifts, "to some extent" - **Altitude problems** — talks *about* approaches rather than staking claims - **Genre mismatch** — delivers a survey when you wanted an argument These are "philosophy-flavoured" expressions of [[Floridi's critique of LLM abduction|Floridi's point]] that LLMs lack a built-in justification loop — but the symptom is *non-commitment* rather than confident falsehood. ## Resolution as a Measurable Dimension "Philosophical competence" can be partly operationalised as: *how little scaffolding is needed* for the model to stay precise, stay at the right altitude, and match the requested genre. > A model is "more philosophical" to the extent that it can satisfy dialectical and logical constraints with less external scaffolding. This connects to [[The obvious move prompting technique]] — skilled prompting functions as external resolution management, compensating for the model's weak point. ## The Structural Explanation Why should the training objective produce these specific failure modes? **Precision is costly; vagueness is a locally safe equilibrium.** If the model gets specific (explicit quantifiers, sharp claims), it risks being caught out. Staying vague ("one might say…") remains broadly acceptable across many possible user intentions. Philosophy amplifies this because vague meta-level talk is often socially acceptable ("here are some ways to think about it"), whereas in maths you're forced into checkable commitments. Source: [[Web Clippings/chat with ChatGPT about Floridi, LLM reasoning, Move 37|ChatGPT conversation on Floridi et al.]]