Developmental trajectories of decision making and affective dynamics in large language models

LLMs are developing psychological profiles that are partly human-like but partly alien: As OpenAI models evolved, newer versions showed more human-like risk-taking and emotional responses to gains and losses, but simultaneously developed distinctly non-human traits like abnormally low loss aversion and artificially elevated baseline mood—suggesting they don't simply become "more human" but rather develop their own unique psychology.

Emotional responses in LLMs don't match human patterns in ways that matter for clinical use: The models showed "affective decay" (emotional responses fading faster than in humans) and chronically higher baseline happiness, which could be problematic if these systems are deployed in medical decision-making where understanding genuine emotional and motivational states is important for patient care.

LLM decision-making is more rigid than human decision-making: Newer models made choices in a more deterministic (rule-based, inflexible) way compared to humans, meaning they may be less adaptable to novel or ambiguous situations—a significant concern for high-stakes domains like medicine where flexibility and contextual judgment are critical.

We need caution before deploying LLMs in clinical settings: The paper directly flags that these "developmental trajectories" have "direct implications for AI ethics" and clinical integration, suggesting that despite some human-like improvements, the non-human psychological signatures revealed here warrant careful consideration before using LLMs in medical decision support.