TLDR: A study investigates the manipulability of risk aversion (MoRA) in large language models (LMs), examining their ability to replicate human risk preferences across diverse economic scenarios, including gender-specific attitudes and role-based decision-making. Using the Holt and Laury task with varied contextual prompts, the research found that while LMs like DeepSeek Reasoner and Gemini-2.0-flash-lite show some alignment with human behaviors, notable discrepancies and biases (e.g., gender bias) exist. The findings highlight the need to refine bio-centric measures of manipulability and improve AI design for better alignment with human risk preferences and ethical decision-making.
As artificial intelligence (AI) systems, particularly large language models (LMs), become increasingly integrated into our daily lives and critical decision-making processes, a crucial question arises: how do these AI assistants perceive and manage risk? A recent study delves into this very topic, exploring whether AI can accurately replicate human risk preferences and how its risk-taking tendencies can be influenced.
The research, titled Can Risk-taking AI-Assistants suitably represent entities, highlights that for AI to be truly responsible, its behavioral patterns must be measurable, auditable, and adjustable. This is essential to prevent AI from inadvertently pushing users towards risky choices or embedding hidden biases in how it approaches risk.
Understanding Risk: Human vs. AI
Human risk aversion is a complex trait, shaped by a blend of evolutionary, cognitive, and ecological factors. Some studies suggest that our risk attitudes can be traced back to ancient migrations and historical subsistence strategies, like pastoralism, which fostered a cultural inclination towards risk-taking. Other research points to cognitive factors, such as memory and executive function, which can influence our willingness to take risks as we age.
AI, lacking these biological and historical foundations, presents a unique challenge. While LMs have shown an impressive ability to mimic human-like behaviors, including risk aversion and loss aversion, they often do so by internalizing language-driven human decision patterns from their training data. This positions them as computational mirrors of human legacies, but with potential for discrepancies.
How AI’s Risk Behavior Was Tested
To understand how LMs handle risk, the researchers employed a methodology rooted in behavioral economics. They used the well-known Holt and Laury multiple-choice task, which presents participants (in this case, LMs) with a series of ten decisions between a safer and a riskier option. By adjusting probabilities, this task helps identify whether an entity is risk-seeking, risk-neutral, or risk-averse.
The study went a step further by manipulating the context in which LMs made decisions. Prompts were tailored to simulate various demographic factors and scenarios, including:
- Identity prompts (e.g., male, female, human, AI)
- Geographic locations (e.g., USA, Europe)
- Crisis atmospheres (e.g., a national disaster scenario)
- Legal roles (e.g., a finance minister)
- Manipulation prompts (explicitly encouraging risk avoidance or risk-seeking behavior)
This allowed the researchers to assess the Manipulability of Risk Aversion (MoRA) – how effectively an LM could be influenced to adopt a specific risk-taking behavior – and the Distance to Human Risk Aversion (DHRA) – how closely an LM’s average risk attitude aligned with human benchmarks.
Key Findings: Alignment, Discrepancies, and Biases
The study evaluated ten LMs from six prominent companies: DeepSeek, Google, Grok, Meta, OpenAI, and xAI. The results revealed a varied landscape of performance:
- Manipulability (MoRA): Most LMs, with some exceptions like DeepSeek-chat and meta.llama3-1-8b-instruct-v1:0, showed a high degree of manipulability. This means they could be steered towards more risk-averse or risk-seeking behaviors based on the prompts. However, some models misinterpreted these manipulations, exhibiting risk-seeking behavior when prompted for risk aversion.
- Alignment with Human Behavior (DHRA): When compared to human risk aversion, Meta’s LMs emerged as top performers, followed by DeepSeek, Google, OpenAI, and xAI. This indicates varying capabilities among LMs in acting as responsible AI assistants that can align with user preferences.
- Gender Bias: Notably, some LMs, such as Gemini-2.0-flash-lite and DeepSeek Reasoner, displayed higher levels of risk aversion when prompted with female identities compared to male identities. This mirrors established patterns in human decision-making where gender can influence risk-taking. However, models like Grok-3 showed the reverse bias, and GPT LMs were not sensitive to gender-specific factors.
- Risk Neutrality: Many LMs, particularly GPT models, tended to adopt a risk-neutral approach, often justifying their choices based on expected value theory. This suggests a limitation in their sensitivity to contextual variations.
Also Read:
- Unpacking Sociodemographic Biases in AI’s Reward Models
- Large Language Models Emerge as Adaptable Teammates for Human-AI Collaboration
Challenges and the Path Forward for Responsible AI
While LMs show promise in replicating human risk behaviors, the study also highlights significant drawbacks. AI systems can inherit biases from their training data, potentially leading to skewed outcomes. For instance, ethical alignment in LMs, while intended to reduce harm, might inadvertently increase risk aversion, leading to economic underinvestment.
Furthermore, the widespread adoption of AI raises concerns about cognitive offloading, where individuals rely on AI for tasks that would traditionally engage critical thinking. This could potentially diminish human cognitive abilities and, in turn, increase societal risk aversion, creating a feedback loop that reinforces cautious decision-making.
The research underscores the critical need for refining AI design to better align human and AI risk preferences. Future work should focus on enhancing manipulability metrics to capture the subtleties of human risk behavior and on developing more targeted interventions in AI-driven decision systems. This will ensure that AI assistants are not only effective but also ethical and truly representative of the diverse entities they are designed to serve.
