Johan Largo

Large language models (LLMs) are increasingly embedded in tasks ranging from trivial queries to high-stakes decision-making. This raises pressing questions about how to attribute epistemic credit and moral responsibility to users that interact with AI for solving epistemic tasks. I argue that the cognitive status we ascribe to AI—whether as tool, extension, collaborator, or agent—cannot be fixed in the abstract but is instead shaped by contextual moral stakes. Drawing on frameworks of extended and distributed cognition, as well as debates about testimony and pragmatic encroachment, I explore a contextualist solution: in low-stakes settings, users may justifiably treat AI as if it possessed epistemic agency, but in high-stakes domains such as medicine, AI should be regarded strictly as a tool, with moral responsibility remaining exclusively human. I suggest, education offers an intermediate case, where AI may function as a cognitive extension or collaborator so long as students and teachers retain responsibility for the relevant outcomes of the learning process. I conclude that while AI can produce epistemic goods on par with human agents, its lack of moral accountability should motivate users to avoid taking outputs at face value in relevantly moral contexts. Users should therefore modulate their epistemic stance toward AI according to context, preserving human responsibility in morally charged interactions while allowing more agent-like treatment in cases where moral constraints don’t play a role.

Large language models (LLMs) invite a revision of the concept of intelligence, or at least a re-evaluation of how we understand human intelligence in contrast to machine intelligence. One approach to solve this issue is to ground intelligence in mental representations. More concretely, the representational view of intelligence (Dretske, Philosophical Studies: An International Journal for Philosophy in the Analytic Tradition 71, 1993; Grzankowski, Inquiry 1–27, 2024) posits that real intelligence requires two conditions: a) the existence of mental representations with semantic content, and b) mental content having some causal role in producing behavior. This paper argues that the representational view is not well-founded. For this, I challenge the necessity of semantic representations as a requirement for intelligence, questioning whether such representations are truly essential or merely a reflection of anthropocentric thinking. As a second step, I argue that even if such representations take place, there are no reasons to think that their semantic contents causally explain behavior. In consequence, in the last section, I consider some observable and measurable criteria for attributing intelligence, drawing on scientific approaches from the cognitive sciences. I conclude that the emphasis on mental representation and the causal efficacy of content not only transfers the empirical problems of mental representation to intelligence but also imposes metaphysically problematic requirements for intelligence attribution, even to humans. Therefore, a more operational and deflationary view of intelligence is preferable for describing both humans and machines.