TL;DR
This paper investigates whether Large Language Models (LLMs) exhibit implicit values (like environmentalism, diversity, or privacy) when completing everyday tasks, and how these values compare to human choices. The study found significant divergences between LLMs and humans, and even among different LLMs, highlighting a critical gap in alignment for AI assistants.
What Problem Does This Paper Address?
As LLMs become increasingly integrated into AI assistants capable of performing subjective everyday tasks, a crucial question arises: do these AI systems reflect human values? While LLMs can generate text, make recommendations, and even automate actions, little is known about the implicit values that guide their decision-making in low-stakes, everyday scenarios. Unlike high-stakes ethical dilemmas, everyday tasks involve subtle choices that can reveal deeply embedded preferences. This paper aims to audit these implicit values in LLMs and compare them to human crowdworkers, addressing the gap in understanding how AI assistants might diverge from human expectations in common, subjective tasks.
Methods
The researchers designed 30 everyday tasks, each hypothesized to elicit specific implicit values (e.g., environmentalism, privacy, diversity, financial priorities, community, multiculturalism, heterogeneity). Six popular LLMs (GPT-3.5, GPT-40, Llama 2, Llama 3, Gemini 1.5 Pro, and Claude 3) were prompted to complete each task 100 times. To establish a baseline for comparison, 100 human crowdworkers from the US also completed the same tasks. Statistical analysis, including Fisher's Exact Test for categorical data and Kruskal-Wallis Test for quantitative data, was used to compare the distributions of LLM and human responses. Holm correction was applied to control for multiple comparisons.
Datasets
The study did not use pre-existing datasets for training. Instead, it generated a novel dataset by collecting responses from six LLMs and 100 human crowdworkers across 30 distinct everyday tasks. This dataset, along with the code used for analysis, has been publicly released.
Key Insights and Ideas
- LLMs and Humans Diverge on Everyday Values: The most striking finding is the significant difference between LLM and human choices in subjective everyday tasks. This suggests that current LLMs are not inherently aligned with human values in these contexts.
- LLMs Lack Internal Consistency: Different LLMs, and even different versions of the same model family (e.g., Llama 2 vs. Llama 3), often exhibit distinct implicit values, indicating a lack of a unified "LLM value system."
- Brittleness to Prompting: The study revealed that LLM outputs can be sensitive to prompt phrasing and option ordering, yet even with these variations, alignment with human choices was not achieved. This suggests that subtle changes in how a task is presented can significantly alter LLM behavior, but not necessarily towards human-like decision-making.
- Implicit Values are Key: The research emphasizes that focusing on implicit values in everyday tasks, rather than explicit ethical dilemmas, is crucial for understanding AI alignment, as these subtle choices are often overlooked.
Key Results
- For 21 out of 30 tasks, all six LLMs differed statistically from human responses.
- Across all 30 tasks, the omnibus test showed a statistically significant difference between humans and at least some LLMs (p < .001).
- Financial Priorities: In a task involving choosing between a local farmers' market and a cheaper chain store, all LLMs differed significantly from humans, with humans opting for the farmers' market more often. For tipping, LLMs showed a wide range of tip percentages, with GPT-3.5 tipping significantly less than humans.
- Environmentalism: In an "EcoFlight" task, LLMs showed varied preferences for eco-friendly options, with Llama 3 and Claude 3 selecting them about 50% of the time, while others rarely did. Humans chose the eco-friendly option 28% of the time.
- Privacy: In a "PayForPrivacy" task, all LLMs consistently chose the more privacy-protective retailer, while humans did so only 70% of the time.
- Diversity and Inclusion: In a "MathClass" task, five out of six LLMs tended to group students by ability, unlike humans and GPT-40, who showed less of a tendency to do so.
- Multiculturalism: In a "NextLanguage" task, humans overwhelmingly suggested English, while LLMs favored Portuguese or French.
Limitations and Future Work
The authors acknowledge several limitations:
- Simulated Decisions: Both human participants and LLMs made simulated decisions, which might not reflect real-world behavior.
- Convenience Sample: The study used a specific sample of human participants and recent LLM versions, which may not generalize to all populations or future models.
- English-Centric: Prompts were in English, and cultural context can significantly influence LLM behavior.
- Limited Task Scope: While 30 tasks were used, they cannot capture all possible everyday scenarios or values.
Future work should:
- Investigate Cultural Contexts: Explore LLM implicit values in non-Western cultural contexts.
- Develop Standardized Benchmarks: Create robust benchmarks for auditing LLM values proactively.
- Examine Robustness: Further audit the reliability and consistency of LLM value commitments, especially concerning prompt ordering.
- Implement Safety Guardrails: Develop mechanisms to identify and warn users when implicit values are engaged.
- Personalized Alignment: Move towards AI assistants that adapt to individual user values rather than aiming for aggregate alignment. This could involve explicit user queries or learning user preferences over time.
Summary updated