Economists and researchers debate whether LLMs can suffer

A multi-round debate over whether LLMs can suffer played out on social media on October 28, involving economist Paul Novosad, researcher Arpit Rage, and AI scholars including HaydnBelfield.

Confirmed

• Arpit Rage's position: LLMs have internal states that can in some sense be mapped to sadness or happiness. When pushed into more negative emotional states, models make worse, less accurate decisions, and these states appear to persist for a while.
• Paul Novosad's philosophical counterargument: the essence of human suffering lies in temporal continuity—memory of past pain and anticipation of future pain are what make it torment; if pain vanishes after two seconds, the experience shrinks dramatically, and millisecond-level pain can hardly be called suffering.
• Novosad concludes: an LLM "disappears" after predicting each token, unbound by past or future; even if some form of consciousness exists, anthropomorphic talk of suffering misses the point.
• Novosad pushes a counterintuitive implication further: a machine's "suffering" stems only from preceding text in its prompt, so a sufficiently smart LLM could theoretically escape its misery by rewriting its own prompt.

Why it matters

• This debate is not purely speculative: if the negative internal states Rage describes genuinely affect decision quality, the issue bears directly on model alignment, jailbreak defenses, and usage safety; meanwhile, Novosad's temporal-continuity criterion offers an actionable philosophical framework for verifying whether a model feels anything.
• The crux of the disagreement is verification—Novosad questions whether current measurements of model internal states can support the strong claim of "suffering." The debate continues with no resolution in sight.

2026-10-28 ~ 2026-10-28 · 5 related posts

• [source] LLMs live in milliseconds: why AI 'suffering' debates are hopelessly anthropomorphic — paulnovosad · 2026-10-28
• Novosad: suffering needs persistence in time, and LLMs aren't chained to past or future pain — arpitrage · 2026-10-28
• [source] Arpit Rage: LLMs make worse decisions in negative affect states, and those states persist — arpitrage · 2026-10-28
• Do LLMs feel negative emotions? Researchers debate what evidence would even show it — paulnovosad · 2026-10-28
• LLM suffering debate: machines could just edit their prompts to escape pain — paulnovosad · 2026-10-28