Why Sampling Is Not Choosing: Intentionality, Agency, and Moral Responsibility in Large Language Models
A philosophical paper challenges the growing narrative that large language models possess genuine agency or moral responsibility. The authors argue that LLMs operate through learned probabilistic mappings with only derived intentionality, not the intrinsic intentionality and commitment-bearing capacity required for moral agency. Stochastic sampling, they contend, produces variation but not authentic choice. This work matters for the AI landscape because it directly counters claims used to justify anthropomorphizing LLMs in policy and product discussions, potentially reshaping how regulators and builders frame accountability for model outputs.
Modelwire context
ExplainerThe paper's sharpest contribution isn't the conclusion that LLMs lack agency (that's familiar territory) but the specific mechanism it targets: stochastic sampling is framed not as a weak form of choice but as categorically different from choice, which forecloses gradualist arguments that LLMs might have 'partial' moral status as they scale.
This connects most directly to the peer review novelty study we covered on June 11, which documented how promotional framing diverges from what evaluators actually find. That paper's concern about capability overclaiming in AI research is precisely the dynamic this philosophical work is trying to interrupt at the conceptual level, before inflated framing reaches policy. The MaxProof coverage is also relevant here: a system achieving gold-medal proof performance on IMO 2025 is exactly the kind of result that feeds anthropomorphizing narratives, and this paper provides the philosophical vocabulary to push back on those readings without dismissing the capability itself.
Watch whether any of the major AI policy frameworks currently in draft (EU AI Act implementing acts, NIST RMF updates) cite philosophical distinctions between intrinsic and derived intentionality when assigning accountability. If they don't engage with this level of conceptual precision by late 2026, the paper will have influenced academic discourse without touching the regulatory layer it's most relevant to.
Coverage we drew on
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsLarge Language Models · LLMs
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.