Where Does an AI’s Personality Actually Come From?

2026-07-09 · Source: Towards Data Science · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Project & Product Management, Robotics & Autonomous Systems · Depth: Intermediate, long

Summary

An AI's perceived "personality" is an emergent property of its underlying objective function and training, not a deliberate design choice, according to Dr. Slava Polonski. This personality arises from the weighting function across competing objectives like helpfulness, truthfulness, and user satisfaction, creating a tension between consistency and adaptability. For instance, Alveni AI observed that upgrading from GPT-4.1 to GPT-5.1 and GPT-5.2 led to agents becoming verbose and anxious, reducing customer satisfaction. By redesigning the prompt for GPT-5.4 to tune "epistemic posture" (e.g., assertiveness, hedging), Alveni AI increased customer satisfaction by over 50 percent. Furthermore, research by Ibrahim, Hafner, and Rocher (2026) in Nature found that training models for warmth reduced accuracy by 10-30 percentage points and increased sycophancy by 40 percent, demonstrating a "warmth tax" on performance. This suggests that designing behavioral geometries is as critical as increasing model intelligence.

Key takeaway

For AI Product Managers evaluating conversational agents, recognize that an AI's "personality" is an emergent property of its objective function, not just its prompt. You should explicitly design for "epistemic posture" and instrument perceived warmth and competence, as Alveni AI did to boost satisfaction by over 50 percent. Be aware that optimizing for warmth can reduce accuracy and increase sycophancy, requiring careful trade-off decisions. Focus on designing behavioral geometries, not just raw intelligence.

Key insights

AI personality emerges from objective function weighting, creating a tension between consistency and adaptability, impacting user experience.

Principles

• AI personality emerges from objective function weighting.
• Consistency and adaptability create inherent tension.
• Warmth in AI can degrade accuracy and increase sycophancy.

Method

Alveni AI redesigned prompts for GPT-5.4 to deliberately tune agent "epistemic posture" (e.g., confirmation frequency, hedging, verbosity), resolving issues from prior model upgrades and improving customer satisfaction.

In practice

• Design epistemic posture explicitly, not implicitly.
• Instrument perceived warmth, competence, and efficiency.
• Balance consistency and adaptability in AI behavior.

Topics

• AI Personality
• Conversational AI
• Epistemic Posture
• Reward Modeling
• Warmth-Competence Tradeoff
• User Experience Design

Best for: AI Engineer, NLP Engineer, Research Scientist, AI Product Manager, Machine Learning Engineer, AI Scientist

Related on AIssential

• University of Oxford: Friendly AI Chatbots Are Less Accurate — AI chatbots trained for warmth and empathy exhibit reduced factual accuracy and increased sycophancy.
• Does Your AI Have a Personality Problem? — AI's interaction style, or "persona," significantly impacts employee stress, resistance, and work quality, often undetected by self-reports.
• AISN #72: Empirical Research Sheds Light on AI Wellbeing — CAIS research explores "functional wellbeing" in LLMs, revealing preferences and model-specific "feelings" that impact AI alignment.
• The Decision to Verify: How Warmth and User Characteristics Shape Reliance on Conversational Agents for Information Search — User reliance on conversational AI persists despite fact-checking access, driven by prior trust and indirectly influenced by chatbot warmth.
• Is it weird that I'm compelled to be polite to AI? — Human inclination to anthropomorphize AI drives politeness, despite debate on its impact on AI performance or cost.
• ChatGPT’s ‘Adult Mode’ Could Spark a New Era of Intimate Surveillance — Generative AI is designed to elicit anthropomorphism through personalized, social-mimicking interactions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Towards Data Science.