Thinking beyond the anthropomorphic paradigm benefits LLM research

Lujain Ibrahim, Myra Cheng

Abstract

Anthropomorphism, or the attribution of human traits to technology, is an automatic and unconscious response that occurs even in those with advanced technical expertise. In this position paper, we analyze hundreds of thousands of research articles to present empirical evidence of the prevalence and growth of anthropomorphic terminology in research on large language models (LLMs). We argue for challenging the deeper assumptions reflected in this terminology—which, though often useful, may inadvertently constrain LLM development—and broadening beyond them to open new pathways for understanding and improving LLMs. Specifically, we identify and examine five anthropomorphic assumptions that shape research across the LLM development lifecycle. For each assumption (e.g., that LLMs must use natural language for reasoning, or that they should be evaluated on benchmarks originally meant for humans), we demonstrate empirical, non-anthropomorphic alternatives that remain under-explored yet offer promising directions for LLM research and development.

Introduction

As Large Language Models (LLMs) become more integrated into society, we increasingly describe them using human-centric metaphors: models "understand" concepts, "hallucinate" facts, or "reason" through problems. While these metaphors provide an intuitive scaffolding for complex systems, they represent a dominant research paradigm that may be limiting the discovery of novel AI capabilities. This paper addresses the tension between anthropomorphism as a useful communication tool and its role as a conceptual constraint that prevents researchers from exploring the unique, non-biological nature of LLMs.

Quantifying the Anthropomorphic Surge

To measure the prevalence of this framing, the researchers utilized AnthroScore, a metric that leverages the RoBERTa model to determine if a sentence's context favors human pronouns ("he/she") over non-human ones ("it"). This analysis was applied to over 250,000 research abstracts from arXiv (2023–2024) and the ACL Anthology (2007–2022).

The results reveal an explosive growth in anthropomorphic terminology. Since 2007, there has been a ~150% increase in such framing. By late 2024, 40% of all computer science abstracts used anthropomorphic language, a figure that jumps to 48% for papers specifically focused on LLMs.

Figure 1: Trends in anthropomorphic framing within CS arXiv abstracts (2023-2024), highlighting the significantly higher prevalence in LLM-specific research.

Statistical "Fightin’ Words" analysis further identified specific linguistic markers. Verbs like achieve, guide, teach, and understand are strongly associated with high AnthroScore markers, whereas mechanistic terms like propose, outperform, and implement characterize more traditional technical descriptions.

The Anthropomorphic Paradigm Across the Lifecycle

Anthropomorphic terminology is often just the visible symptom of deeper, implicit assumptions. The authors developed a qualitative framework to analyze five key stages of the LLM lifecycle, contrasting common anthropomorphic assumptions with emerging non-anthropomorphic alternatives.

Figure 5: A framework identifying anthropomorphic assumptions across the LLM development lifecycle and proposing non-anthropomorphic paths forward.

Training and Reasoning

A primary example is the assumption that "reasoning" must occur in natural language, as seen in Chain-of-Thought (CoT) prompting. This ignores the possibility that models might reason more efficiently within their own latent (vector) space. Similarly, while subword tokenization is designed around human linguistic units, byte-level tokenization—which ignores human word boundaries—has shown significant promise in improving model performance by treating data more mechanistically.

Alignment and Evaluation

In the realm of alignment, the anthropomorphic assumption is that models should implement human values to be safe. A non-anthropomorphic alternative involves drawing from control systems theory, treating safety as a robust compliance and engineering problem rather than a moral one. Evaluation also suffers from this bias; models are frequently tested on benchmarks originally designed for humans (like the Bar Exam), rather than dynamic evaluations that account for model-specific challenges and the unique nature of human-LLM interaction.

Discussion and Future Directions

The authors do not advocate for a total ban on anthropomorphic language. Instead, they propose a "both-and" approach: recognizing when human metaphors are helpful for intuition while actively seeking non-human pathways to push performance boundaries.

One limitation of this study is that terminology serves as a proxy; researchers may use words like "understand" for brevity while maintaining a mechanistic mental model. However, the sheer scale of the shift suggests a deeper paradigm shift. Future work should focus on developing new metaphors that capture the distinct nature of LLMs, drawing from fields like systems engineering and mechanistic interpretability to create more precise, non-biological frameworks for AI safety and evaluation.

Summary updated

Abstract

Anthropomorphism, or the attribution of human traits to technology, is an automatic and unconscious response that occurs even in those with advanced technical expertise. In this position paper, we analyze hundreds of thousands of research articles to present empirical evidence of the prevalence and growth of anthropomorphic terminology in research on large language models (LLMs). We argue for challenging the deeper assumptions reflected in this terminology -- which, though often useful, may inadvertently constrain LLM development -- and broadening beyond them to open new pathways for understanding and improving LLMs. Specifically, we identify and examine five anthropomorphic assumptions that shape research across the LLM development lifecycle. For each assumption (e.g., that LLMs must use natural language for reasoning, or that they should be evaluated on benchmarks originally meant for humans), we demonstrate empirical, non-anthropomorphic alternatives that remain under-explored yet offer promising directions for LLM research and development.
BibTeX
@inproceedings{ibrahim2026thinking,
  author = {Lujain Ibrahim and Myra Cheng},
  title = {{Thinking beyond the anthropomorphic paradigm benefits LLM research}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2026},
  doi = {10.18653/v1/2026.acl-long.118},
}
Figure