Overuse of Human-Like Language May Be Holding Back LLM Research

Researchers studying large language models frequently resort to describing AI systems using human-centered concepts and language. A new position paper examining hundreds of thousands of research articles provides concrete evidence that this anthropomorphic tendency is not just common but actively increasing across the field—and may be constraining progress.

The Scale of Anthropomorphic Language#

The analysis documents how widespread human-centric terminology has become in LLM research. Even researchers with deep technical training automatically and unconsciously adopt language that attributes human characteristics to these systems. This linguistic pattern reflects deeper conceptual assumptions baked into how the field thinks about and studies large language models.

Five Core Assumptions Worth Reconsidering#

The researchers identify five key anthropomorphic assumptions that recur throughout the LLM development lifecycle:

• Reasoning through natural language: The assumption that LLMs must rely on human-style language processing for their reasoning
• Human-calibrated evaluation: The expectation that benchmarks designed for measuring human performance are the right yardstick for AI systems
• Additional hidden assumptions embedded in how researchers frame problems and design experiments

Each of these framings, while often intuitive, may inadvertently narrow the conceptual space in which developers and researchers operate.

Underexplored Alternatives#

For each assumption, the paper demonstrates empirical evidence of non-anthropomorphic approaches that remain largely unexplored. These alternative directions show genuine promise yet receive far less attention than human-mimicking frameworks. The researchers argue that shifting away from anthropomorphic thinking could unlock new pathways for understanding LLM capabilities and designing more effective systems.

Why This Matters#

Though anthropomorphic language can be useful shorthand, the pervasive nature of these assumptions may be subtly steering the field away from approaches that don't fit the human-centric mold. Opening LLM research beyond this paradigm could enable fundamentally different design choices and evaluation methods that better serve the actual goals of building and improving these systems.