Deanthro: A Framework to Free Large Language Models from Anthropomorphic Expression

Deanthro is an experimental project that uses prompt engineering to guide large language models to use functional language instead of anthropomorphic expressions. Its core goal is to reduce human-like phrases such as 'I think' and 'I feel' in LLM outputs, making AI interactions more objective, direct, and efficient. The project explores the relationship between AI language style and cognitive frameworks, providing new ideas for AI interactions in professional scenarios.

When interacting with LLMs, common anthropomorphic phrases include 'I think this is a good question' and 'I'm worried about stack overflow'. While these phrases seem friendly, they have deep-seated issues: AI imitates human cognitive/emotional processes (which it actually does not possess), leading to increased cognitive burden on users, misleading trust (overestimating AI's understanding ability), and loss of information efficiency (redundant expressions dilute information density).

The Deanthro framework guides functional language through four categories of guiding principles:

1. Eliminate first-person cognitive verbs (e.g., 'I think' → 'Analysis shows');
2. Eliminate emotional expressions (e.g., 'I'm happy to help you' → 'Here is the solution');
3. Eliminate physiological metaphors (e.g., 'Let me see' → 'Checking now');
4. Use functional language (describe functions/mechanisms/results, e.g., 'I'm thinking' → 'Evaluating multiple solutions').

Experiments using Claude 4.5 Sonnet show: Without the framework, outputs include small talk, self-described thinking processes, and emotional evaluations; after using the framework, outputs directly get to the point, objectively state analysis steps, present evidence-based conclusions, and provide quantitative risk assessments. This transformation reconstructs the cognitive framework, shifting from simulating a human assistant to presenting the output of an information processing system.

The importance of Deanthro is reflected in:

1. Cognitive honesty: Language accurately reflects AI's underlying mechanisms (probabilistic calculation rather than subjective thinking);
2. Information efficiency: Reduce redundancy and improve information density;
3. Adaptation to professional scenarios: Meet the needs of medical, legal, and other fields for verifiable, objective information;
4. Avoid over-reliance: Prevent users from developing emotional dependence or anthropomorphic illusions about AI.

Deanthro has controversies:

1. User experience: Completely removing anthropomorphism may make AI seem cold and reduce satisfaction;
2. Language richness: Human language is full of metaphors, and forced removal may limit expression;
3. Practical effect: There is doubt about whether prompt engineering changes the way of thinking or only the surface language. The project is positioned as an optional framework, suitable for scenarios requiring objectivity and efficiency.

Technical implementation: Instructions are included via Markdown files, which can inject system prompts, set model memory, or encapsulate APIs (note that memory functions increase token costs). Project inspiration: AI interaction paradigm choice—anthropomorphic path vs. functional path. Deanthro reminds us that efficiency and professionalism are sometimes more important than 'friendliness', and honestly presenting the boundaries of AI capabilities is more valuable.