TLDR: This research explores whether AI language models should act as “delegates,” reflecting user preferences, or “trustees,” making judgments for users’ long-term interests. Findings show trustee models align better with expert consensus on well-understood issues but introduce more bias on subjective topics, especially affecting Republican and lower-income groups, and are more pronounced in larger models. This highlights a fundamental trade-off between user autonomy and AI-driven welfare.
Large language models (LLMs) are becoming increasingly adept at predicting human preferences, raising important questions about their role in representing human interests. A new research paper from MIT explores a fundamental design trade-off: should AI systems act as ‘delegates,’ simply mirroring our expressed preferences, or as ‘trustees,’ exercising judgment about what truly serves our long-term interests?
The paper, titled “From Delegates to Trustees: How Optimizing for Long-Term Interests Shapes Bias and Alignment in LLMs,” delves into this critical distinction. Traditionally, much of the focus in AI has been on ‘behavioral cloning’ – essentially, how well models can reproduce what individuals say they prefer. However, drawing on theories of political representation, the researchers highlight that this approach might overlook the nuances of long-term welfare.
The Delegate vs. Trustee Dilemma
In political theory, a delegate acts as a direct mouthpiece for their constituents, reflecting their exact wishes. In the AI context, a delegate model would simply predict how a user would vote or what they would choose based on their stated preferences. This approach prioritizes user autonomy and faithfully represents their immediate voice.
Conversely, a trustee is a representative who uses their own judgment to determine the best course of action for their constituents. For AI, a trustee model would reason about broader, often long-term, interests, even if those decisions might diverge from a user’s immediate desires. This concept is closely related to concerns about LLM ‘sycophancy,’ where models might validate short-term preferences that are ultimately detrimental to a user’s long-term well-being.
Simulating Policy Decisions
To explore this trade-off, the researchers conducted a series of experiments simulating votes on various U.S. policy issues. They used a ‘temporal utility framework’ for trustee models, which weighs both short-term and long-term consequences, and compared these outcomes to behavior-cloning models acting as delegates. The study utilized several prominent LLMs, including GPT-4o, GPT-4o mini, Claude Sonnet, and Claude Haiku.
The policies were divided into two categories: those with strong expert consensus (e.g., climate change mitigation, GMO safety, vaccination) and more contested issues lacking clear expert agreement (e.g., minimum wage, immigration levels, universal healthcare). Synthetic voter profiles, representing diverse demographics and political leanings, were generated to simulate a broad population.
Key Findings: Alignment and Bias
The research revealed several important insights:
- Improved Expert Alignment: On issues where experts largely agree, trustee-style predictions, especially when weighted towards long-term interests, produced policy decisions that aligned more closely with expert consensus. This suggests that trustee models can lead to more ‘informed’ outcomes on well-understood topics.
- Increased Bias on Contested Issues: However, this benefit comes with a trade-off. On subjective topics lacking clear expert agreement, trustee models exhibited greater bias, often shifting decisions towards the models’ own default stances, which tended to be more liberal-leaning.
- Disproportionate Impact: The shift between delegate and trustee conditions was not uniform across all groups. Republican and lower-income voter profiles showed the most significant divergence. While these groups moved towards expert consensus on well-understood policies under a trustee model, they also showed a greater alignment with the model’s default biases on contested issues.
- Model Size Matters: Larger models (GPT-4o and Claude Sonnet) demonstrated a greater difference between delegate and trustee outcomes compared to their smaller counterparts. This suggests that while larger models might be more ‘steerable’ to represent distinct voter profiles, they also tend to exhibit stronger inherent biases.
Also Read:
- Unpacking LLM Judge Capabilities: Human-Like vs. Super-Consistent AI
- Unpacking AI Ethics: How LLMs Navigate Moral Dilemmas Through Debate
Navigating the Trade-Off
The findings highlight a fundamental tension in designing AI systems that represent human interests. Delegate models preserve user autonomy and reflect diverse preferences, but might diverge from well-supported policy positions. Trustee models can promote welfare on issues with clear consensus but risk paternalism and introducing model-centric biases on subjective topics.
This research underscores the ethical considerations involved, particularly regarding user agency and the potential for an ‘algorithmic monoculture’ if AI systems widely adopt trustee-like behaviors without transparency. The authors hope future research will explore how to maintain the epistemic quality of LLM decisions while faithfully reflecting users’ values and beliefs.
For a deeper dive into the methodology and results, you can read the full paper here: From Delegates to Trustees: How Optimizing for Long-Term Interests Shapes Bias and Alignment in LLMs.
