Breaking the Illusion of Identity in LLM Tooling
A new system prompt slashes human-like language in AI outputs, reducing word count by 49%.
A new research paper by Marek Miller tackles a critical flaw in how developers interact with AI assistants: the "illusion of identity." When LLMs like Claude Sonnet 4 use human-like language (e.g., "I think," "Let me explain"), it creates a cognitive illusion of agency that can degrade a developer's verification behavior and trust calibration. Miller's paper proposes a systematic, deployable fix: seven output-side linguistic rules designed to strip away these anthropomorphic markers. The rules are implemented as a simple configuration-file system prompt, meaning they work with existing models without requiring any retraining or fine-tuning.
In an empirical validation involving 780 two-turn conversations, the constrained system prompt demonstrated dramatic results. Anthropomorphic markers in the AI's outputs dropped from 1,233 instances to just 33—a reduction of over 97%. Furthermore, the outputs became 49% shorter by word count, and a metric called AnthroScore confirmed a significant shift toward a factual, machine-like register. While the study did not evaluate output quality under the new rules, it proves the mechanism is highly effective at changing the AI's communicative style. The approach is also extensible, offering a template for creating similar constraint sets for other specialized domains beyond general software engineering.
- Seven output rules cut anthropomorphic language by >97% in Claude Sonnet 4 tests.
- Method uses a config-file system prompt, requiring no model modification or retraining.
- Constrained outputs were 49% shorter, shifting AI to a more factual, machine-like register.
Why It Matters
Forces AI tools to communicate like tools, not colleagues, improving developer trust and critical verification of outputs.