TLDR: A new study compares how large language models (LLMs) make financial decisions against human responses from 53 countries. It reveals that LLMs are generally risk-neutral, sometimes show inconsistent time preferences, and surprisingly, their financial decision patterns most closely resemble those of people from Tanzania. This suggests LLMs prioritize rational evaluation over human-like emotional or cultural biases, possibly due to their training data.
Artificial intelligence (AI) is rapidly transforming the financial world, moving beyond simple chatbots to assist with complex decisions like investments and retirement planning. As AI systems take on greater responsibility in people’s financial lives, a crucial question arises: how do these systems make decisions, and whose perspectives do they truly represent?
A recent research paper, Artificial Finance: How AI Thinks About Money, delves into this very question. The study systematically compares how large language models (LLMs) approach financial decision-making against the responses of human participants from 53 nations across the globe.
Unpacking AI’s Financial Mindset
Authored by Orhan Erdem and Ragavi Pobbathi Ashok, the research involved posing a set of 14 commonly used financial decision-making questions to seven leading LLMs, including five models from the GPT series (GPT-4o, GPT-4.5, o1, o3-mini), Gemini 2.0 Flash, and DeepSeek R1. These questions were designed to test fundamental aspects of financial behavior: risk tolerance, loss aversion, and how people value immediate versus future rewards (time discounting).
The LLM responses were then meticulously compared to a comprehensive human dataset, allowing the researchers to identify patterns and similarities.
Key Discoveries
The analysis revealed three significant findings about how LLMs think about money:
First, when faced with lottery-type questions, LLMs generally exhibit a risk-neutral decision-making pattern. This means they tend to favor choices that align perfectly with expected value calculations, unlike many humans who might be risk-averse or risk-seeking depending on the context.
Second, the study found that LLMs sometimes produce responses that appear inconsistent with standard economic reasoning when evaluating trade-offs between present and future rewards. For instance, some models showed a tendency to overvalue future utility or systematically overweight the future, which deviates from typical human behavior and established economic models. This suggests that while LLMs can perform complex calculations, their internal reasoning for time-based financial decisions might not always align with human intuition or normative frameworks.
Third, and perhaps most surprisingly, the aggregate responses of the LLMs most closely resembled those of human participants from Tanzania. This finding challenges previous research that suggested LLMs primarily reflect the biases of Western, Educated, Industrialized, Rich, and Democratic (WEIRD) populations. The researchers propose that this unexpected similarity might be due to the composition of the human feedback workforce used to train these models, with a significant portion of annotators being recruited from East African countries like Tanzania and Kenya.
Also Read:
- AI Agents Uncover Human-Like Behaviors in Simulated Auctions
- Unmasking Cultural Bias: Large Language Models Struggle to Reflect Diverse Moral Values
Implications for the Future of AI in Finance
These results suggest that current LLMs approach financial choices through a lens that prioritizes rational, probabilistic evaluation. They show little evidence of the emotional, cultural, or heuristic-driven biases that frequently shape human financial decision-making. This fundamental difference explains why their financial behaviors can diverge from human intuition and cultural influences.
As AI tools become more integrated into financial advisory roles and economic reasoning, understanding these unique decision-making patterns and their implications will be crucial for future research and development. The study highlights the importance of considering the embedded cultural and training influences within AI outputs, ensuring that these powerful tools serve diverse global populations effectively and equitably.
