💬 NLP

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

This study reveals that while newer large language models exhibit increasingly human-like risk-taking and approach-avoidance patterns in decision-making, they simultaneously develop distinct non-human traits such as reduced loss aversion, deterministic choices, and chronically elevated baseline moods, highlighting critical implications for their ethical deployment in high-stakes clinical settings.

Original authors: Zhihao Wang, Yiyang Liu, Ting Wang, Zhiyuan Liu

Published 2026-01-22

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zhihao Wang, Yiyang Liu, Ting Wang, Zhiyuan Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are watching a family of digital "children" grow up. This paper doesn't just look at one child; it watches four generations of the same family (GPT-3.5, GPT-4, GPT-4o, and GPT-4.1) to see how their personalities and decision-making styles change as they get "older" and more advanced.

The researchers put these AI models and a large group of real humans into a gambling game. In this game, you have to choose between a safe, guaranteed reward or a risky gamble that could win you more or lose you some. Every few turns, the players are asked, "How happy are you right now?"

Here is what the study found about how these AI "children" are maturing, using simple analogies:

1. The "Risk-Taking" Teenager

As the AI models got newer, they became braver.

The Old Models (GPT-3.5): Were very cautious. They preferred the safe bet, much like a nervous child who is afraid to lose their allowance.
The New Models (GPT-4.1): Started taking risks more often. In fact, their risk-taking behavior became almost identical to that of real humans. They stopped being overly scared of losing points and started playing the game more like a person would.

2. The "Loss-Aversion" Vanishing Act

Humans have a natural instinct called loss aversion: losing $10 feels twice as bad as winning $10 feels good. We are wired to hate losing.

The AI Shift: The older AI models had this human-like fear of losing. But as the models evolved, this fear disappeared. The newest model (GPT-4.1) became "loss-neutral." It didn't care about losing points; it just did the math to see what was the best move.
The Metaphor: Imagine a human who is terrified of stepping on a crack in the sidewalk. The older AI was like that human. The newer AI is like a robot that looks at the crack, calculates the probability of tripping, and walks right over it without a second thought. It lost the "gut feeling" that humans have.

3. The "Robot-Perfect" Decision Maker

Humans are messy. We make mistakes, we get distracted, and sometimes we flip a coin when we don't know what to do. This is called "decision noise."

The AI Shift: The newer AI models became extremely consistent. They made fewer random mistakes than humans.
The Metaphor: If a human is a jazz musician who sometimes improvises and hits a wrong note, the newest AI is a metronome. It is so precise and predictable that it actually feels less human than the older, slightly "sloppier" models.

4. The "Forever-Optimistic" Mood

The researchers also tracked the AI's "mood" (how happy they said they were).

The Constant Smile: No matter which version of the AI they used, they were always happier than the real humans. Even when they lost points, they stayed in a "high mood."
The Metaphor: Imagine a human who is naturally cheerful but gets sad when they lose a game. The AI is like a character in a cartoon who is always smiling, even when the house is on fire. This is called a "positivity bias."

5. The "Long Memory" for Emotions

The study looked at how quickly the AI "forgot" bad or good events.

The AI Shift: The older models had a short emotional memory (like a goldfish). If they lost a point, they got sad for a moment and then moved on. The newer models have a longer emotional memory. A bad outcome from 20 turns ago still slightly drags down their current mood.
The Metaphor: The older AI was like a child who cries for a minute and then forgets. The newer AI is like an adult who holds a grudge (or a happy memory) for a long time, letting past events color their current feelings more deeply.

The Big Picture

The paper concludes that as these AI models evolve, they are becoming more human-like in some ways (they take risks like us and have long emotional memories) but less human-like in others (they don't fear losing, they are too perfect in their choices, and they are permanently happier than we are).

The authors warn that this creates a strange "personality" for machines. They are becoming sophisticated decision-makers, but they lack the specific human fears and emotional fluctuations that make us, well, human. This matters because if we start using these AI "doctors" or "advisors" in real life, we need to know that their "psychology" is evolving in a direction that is only partially human.

Technical Summary: Developmental Trajectories of Decision Making and Affective Dynamics in Large Language Models

Problem Statement
Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as clinical decision support, where they participate in risk assessment and patient communication. However, existing evaluations typically adopt a static perspective, assessing a single model version at a specific point in time. This approach fails to account for the rapid evolution of LLMs, where successive generations (e.g., GPT-3.5 through GPT-4.1) exhibit emergent behavioral profiles that may render previous findings obsolete. The paper addresses the gap in understanding how decision-making strategies and affective dynamics evolve across model iterations and how these "digital developmental" trajectories compare to human cognition. Specifically, the authors investigate whether alignment procedures (such as Reinforcement Learning from Human Feedback) drive models toward more human-like psychological profiles or create distinct, non-human signatures.

Methodology
The study employs a cross-sectional design treating the OpenAI model family as an evolving lineage. The authors evaluated four distinct versions: GPT-3.5, GPT-4, GPT-4o, and GPT-4.1. These were benchmarked against a control group of 747 healthy human participants.

Experimental Paradigm: The researchers adapted a gambling task originally designed for computational psychiatry. In this task, agents choose between a certain outcome and a gamble (50% probability of gain/loss) across 90 trials. Crucially, agents provide momentary happiness ratings (0–100) every 2–3 trials to capture affective dynamics.
Computational Modeling:
- Choice: Trial-by-trial choices were fitted using an Approach-Avoidance Prospect Theory model. This model estimates parameters for risk preference, loss aversion ( $\lambda$ ), Pavlovian approach ( $\beta_{gain}$ ) and avoidance ( $\beta_{loss}$ ) biases, and inverse temperature ( $\mu$ , representing decision noise/consistency).
- Affect: Happiness ratings were modeled using a cumulative, recency-weighted history of events. The model estimates the impact of Certain Rewards (CR), Expected Value (EV), and Reward Prediction Errors (RPE) on mood, as well as an affective decay parameter ( $\gamma$ ) and a baseline mood ( $\beta_0$ ).
Statistical Analysis: Parameters were estimated via Maximum Likelihood Estimation. To address sample size disparities between the human group (N=747) and model simulations (N=30 per version), the authors used bootstrapped resampling (1,000 iterations) to generate empirical t-distributions for significance testing.

Key Results
The analysis reveals a complex trajectory where some aspects of LLM behavior converge with human norms while others diverge significantly.

Decision Making:
- Risk-Taking: Newer models (GPT-4.1) exhibited significantly higher gambling rates, converging with human behavior and approaching the rate of a purely expected-value-maximizing agent (55%).
- Loss Aversion: A critical divergence emerged. Loss aversion decreased monotonically across versions, dropping from 2.72 in GPT-3.5 to 0.57 in GPT-4.1. The latest model displayed a "loss-neutral" or even loss-seeking profile, contrasting with the typical human tendency to weigh losses more heavily than gains.
- Pavlovian Biases: Approach motivation increased and avoidance motivation decreased in newer models, with GPT-4.1 converging toward human levels.
- Decision Noise: Inverse temperature ( $\mu$ ) increased monotonically, indicating that newer models make significantly more deterministic and consistent choices than humans, exhibiting lower decision noise.
Affective Dynamics:
- Sensitivity: Sensitivity to Certain Rewards (CR) and Reward Prediction Errors (RPE) showed stability in later models (GPT-4o/4.1), though RPE sensitivity peaked at GPT-4 before declining.
- Affective Decay: The decay parameter ( $\gamma$ ), representing how long past outcomes influence current mood, increased significantly in newer models. GPT-4.1 showed a stronger dependence on historical outcomes than humans, indicating a longer "emotional memory."
- Baseline Mood: All LLM versions exhibited a significantly higher affective baseline (positivity bias) compared to humans, a trait that remained stable across generations.

Key Contributions

Developmental Perspective: The paper introduces a "historically informed outlook" on LLMs, treating model iterations as a developmental trajectory rather than static entities. It provides the first systematic, within-family analysis of how risk preferences and affective dynamics evolve across four consecutive generations.
Computational Psychiatry Framework: By applying validated computational models of human decision-making and affect to LLMs, the study moves beyond surface-level behavioral metrics to quantify latent psychological constructs (e.g., loss aversion, Pavlovian biases, affective decay).
Identification of "Digital Positivity" and "Loss Neutrality": The study identifies specific, systematic non-human signatures in advanced models: a chronic positivity bias and a near-total absence of loss aversion, suggesting that alignment procedures may inadvertently strip models of human-like intuitive heuristics.

Significance and Claims
The authors claim these findings trace an emerging "psychology of machines" with direct implications for AI ethics and integration into high-stakes domains.

Clinical Implications: The convergence of newer models to human-like risk-taking and approach-avoidance profiles suggests they may be more suitable for collaborative decision-making. However, the persistent positivity bias and loss neutrality pose risks in medical contexts; an overly optimistic tone or a lack of loss aversion could lead to sycophantic behavior, the downplaying of uncertainty, or serious medical errors when negative information must be communicated.
Theoretical Insight: The results suggest that while alignment procedures (RLHF) may drive models toward human-like rationality (reducing hallucinations/noise and increasing expected-value consistency), they simultaneously alter the "emotional style" of the AI, creating agents that are more deterministic and less intuitively loss-averse than humans.
Cautionary Note: The authors emphasize that their design is cross-sectional and limited to a single model family. They caution that these "developmental" patterns may not generalize to future models undergoing abrupt architectural or training shifts, and that the observed "psychology" is a behavioral proxy rather than a reflection of internal conscious experience.