# Floridi's critique of LLM abduction [[Luciano Floridi]] et al.'s argument that LLMs have a "stochastic core" with only an "abductive appearance" — they generate text that *looks like* reasoning without performing the epistemic work reasoning is supposed to do. ## The Central Claim > "LLMs generate text based on learned associations rather than performing abductive inferences… they do not… ground [outputs] directly in truth, semantics, verification, or understanding." (Floridi et al. 2025, p. 1) The paper's punchline: LLMs don't have "truth" as an internal objective. Their objective is to output a continuation that is probable given the prompt and learned distributions. ## The Missing Loop In Bayesian terms: > "In statistical terms, LLMs perform prior predictive sampling but lack an external feedback loop for posterior evaluation." (p. 7) They generate plausible candidates (discovery phase) but don't genuinely validate them against reality (justification phase) — "unless they are specifically augmented by other systems, which only proves the point." ## The Smoke/Fire Example Floridi et al. contrast: - **Humans**: infer real fire because they understand causal structure in the world - **LLMs**: output "fire" because "smoke → fire" is a strong linguistic association in training data The output can be identical, but the underlying "aboutness" differs: human inference is world-directed; the model's is text-distribution-directed. ## The "Stochastic Parrots" Framing They accept the "stochastic parrots" metaphor highlights two limitations: 1. LLMs are limited by their training data — they can be creative and remix, but gaps and biases in data produce gaps and biases in output 2. LLMs don't know whether they're right or wrong — they can't lie in the ordinary sense because they have no understanding of truth conditions ## Weak vs Strong Abduction The paper grants LLMs can perform "weak abduction" (generate plausible explanations) and even look like "strong abduction" on multiple-choice tasks. But they redescribe this as "zeroth-order abduction": > Generating plausible continuations that mimic hypothesis generation, without real understanding or verification. ## The Verification Concession Importantly, Floridi et al. acknowledge you *can* make LLM outputs more truth-tracking: > "Any connection to real-world evidence must be deliberately engineered, as in retrieval-augmented systems." But this proves their point: the base token-completion model doesn't have truth as an internal norm. You need to bolt on tools, retrieval, verification modules. ## Response: Philosophy as Special Case See [[Philosophy as internal verification domain]] — the critique has different force depending on the domain. In philosophy, "posterior evaluation" is often internal (validity, consistency, dialectical robustness), so the missing "world verification loop" matters less. Source: [[Web Clippings/chat with ChatGPT about Floridi, LLM reasoning, Move 37|ChatGPT conversation on Floridi et al.]]