# Three frameworks for AI mentality

## TL;DR

This paper analyzes three frameworks for understanding how people attribute mentality to AI systems, focusing on implications for interpretation and cognition.

## Contribution

The paper introduces a novel framework for graded attributions of belief-like states in AI systems.

## Key findings

- Architectural debunking arguments are often too quick but highlight distinctions between deep and shallow folk-psychological concepts.
- Mere roleplay views are psychologically unstable in systems designed to elicit anthropomorphism.
- A minimal cognitive agents framework allows for limited, graded attributions of belief- and desire-like states in LLMs.

## Abstract

Rapid advances in large language models (LLMs) have been accompanied by a striking increase in public and user attribution of mentality to AI systems. This paper offers a structured analysis of these attributions by distinguishing three frameworks for thinking about AI mentality and their implications for interpretation. First, I examine “mindless machines” views, focusing on architectural debunking arguments that claim mechanistic or algorithmic descriptions render folk-psychological explanation redundant. Drawing on Marr’s levels of analysis, I argue that such arguments are often too quick, though they highlight an important distinction between “deep” folk-psychological concepts that are sensitive to implementation and “shallow” concepts such as belief and desire that are more architecture-indifferent. Second, I assess “mere roleplay” views that treat mental-state ascriptions to LLMs as useful heuristics akin to engagement with fiction. I argue that this stance is psychologically unstable in anthropomimetic systems designed to elicit unironic anthropomorphism, and theoretically incomplete insofar as roleplay analogies typically presuppose an underlying agent. Third, I develop a “minimal cognitive agents” framework under which LLMs may warrant limited, graded attributions of belief- and desire-like states. I suggest that moving from binary to multidimensional, continuous conceptions of belief can preserve distinctions between humans, LLMs, and simpler systems while better capturing emerging interpretive practice and its normative stakes.