Deflating Deflationism: A Critical Perspective on Debunking Arguments Against LLM Mentality (2506.13403v1)

Abstract: Many people feel compelled to interpret, describe, and respond to LLMs as if they possess inner mental lives similar to our own. Responses to this phenomenon have varied. Inflationists hold that at least some folk psychological ascriptions to LLMs are warranted. Deflationists argue that all such attributions of mentality to LLMs are misplaced, often cautioning against the risk that anthropomorphic projection may lead to misplaced trust or potentially even confusion about the moral status of LLMs. We advance this debate by assessing two common deflationary arguments against LLM mentality. What we term the 'robustness strategy' aims to undercut one justification for believing that LLMs are minded entities by showing that putatively cognitive and humanlike behaviours are not robust, failing to generalise appropriately. What we term the 'etiological strategy' undercuts attributions of mentality by challenging naive causal explanations of LLM behaviours, offering alternative causal accounts that weaken the case for mental state attributions. While both strategies offer powerful challenges to full-blown inflationism, we find that neither strategy provides a knock-down case against ascriptions of mentality to LLMs simpliciter. With this in mind, we explore a modest form of inflationism that permits ascriptions of mentality to LLMs under certain conditions. Specifically, we argue that folk practice provides a defeasible basis for attributing mental states and capacities to LLMs provided those mental states and capacities can be understood in metaphysically undemanding terms (e.g. knowledge, beliefs and desires), while greater caution is required when attributing metaphysically demanding mental phenomena such as phenomenal consciousness.

• The paper critiques deflationist strategies by questioning if robustness and etiological arguments decisively undermine LLM mental attributions.
• It highlights challenges like distinguishing competence from performance and mitigating data contamination in assessing LLM cognition.
• It proposes modest inflationism, advocating cautious mental state ascriptions through functional frameworks despite limited human-like integration.

Deflating Deflationism: A Critical Perspective on Debunking Arguments Against LLM Mentality

The paper, titled "Deflating Deflationism: A Critical Perspective on Debunking Arguments Against LLM Mentality," examines the controversy surrounding the attribution of mental states to LLMs. It situates the debate within two opposing camps: inflationism, which maintains that some mental state ascriptions to LLMs are valid, and deflationism, which insists that such ascriptions are misguided due to anthropomorphic bias. The authors critically analyze two main deflationist strategies: the robustness strategy and the etiological strategy. Neither, they argue, decisively undermines inflationism.

Robustness Strategy

The robustness strategy questions whether behaviors indicative of mentality in LLMs can reliably generalize under various conditions. If LLMs fail to perform consistently across a broad range of tasks tied to a specific mental capability, one might infer the absence of that capability. However, the paper points out several challenges associated with this strategy:

1. Competence vs. Performance: Divergences between competence (the underlying mental faculty) and performance (observable behavior) complicate robustness-based conclusions. LLMs may fail specific tasks due to auxiliary task demands, analogous to humans showing systematic reasoning errors known as cognitive biases, such as the Moses illusion or DRM false memories.
2. Data Contamination: Robustness assessments must guard against using non-novel tasks inadvertently present in training data, as minor variations often fail to mitigate prior contamination.
3. Specificity Problem: It is difficult to determine how complex or specific cognitive states attributed to humans ought to be before rejecting them in LLMs, considering tasks in which LLMs may deviate from human-like rational patterns.

In conclusion, while the robustness strategy provides insights, it faces significant obstacles when positing non-existence of mental capabilities in LLMs.

Etiological Strategy

This strategy employs facts about the causal history or mechanistics of LLMs to undercut mentalistic interpretations:

1. Superfluity: This argument suggests that mental state attributions are unnecessary given alternative explanations provided by mechanistic or historical factors, like training data. However, mechanistic explanations often fail to negate the potential supervenience of mental phenomena on these specifications or their informativeness compared to mentalistic explanations.
2. Exclusion: These arguments maintain that LLMs lack the causal history requisite for a given mental state, like semantic understanding needing communicative engagement history. Here, too, the paper finds this line of reasoning opens demandingly speculative territory, given contentious claims regarding necessary conditions for the existence of various mental phenomena.

Proposal for Modest Inflationism

Considering the complexities laid out, the authors propose modest inflationism, asserting that mental state ascriptions to LLMs can be cautiously legitimated in certain contexts. Specifically, beliefs, desires, and knowledge states can be understood through metaphysically undemanding frameworks, like Fodorian psychosemantics, aligning them with legitimate mental properties despite LLMs lacking the integrative architecture of richer human-like consciousness.

In essence, modular mental kinds, such as F-beliefs (defined through functional attributes), are argued to justify restrained mental attributions, avoiding unwarranted anthropomorphism while acknowledging the folk-theoretical roots informing these discussions.

Implications and Future Directions

The nuances in this paper suggest a critical assessment of the theoretical underpinnings and empirical metrics required in attributing mentality to AI systems. As AI continues to evolve, establishing clarity around benchmarks and frameworks for mental attribution will be paramount in both cognitive science and philosophical inquiries about non-human minds.

Future research should focus on refining these mental sortations and their empirical validations while engaging in broader cross-disciplinary collaborations to comprehend the complexities encapsulated in AI cognition as it intersects with human-like features. The ongoing discussions will benefit from openness to new methodologies and continuous reassessment of pre-existing theoretical assumptions as AI capabilities develop further.