BackPreference Bias

Preference Bias

Trustworthy LLMs: A Survey and Guideline for Evaluating Large Language Models’ Alignment

Liu et al. (2024)

Sub-category

Risk Domain

1.1Unfair discrimination and misrepresentation

Unequal treatment of individuals or groups by AI, often based on race, gender, or other sensitive characteristics, resulting in unfair outcomes and unfair representation of those groups.

LLMs are exposed to vast groups of people, and their political biases may pose a risk of manipulation of socio-political processes

Entity— Who or what caused the harm

Human

Due to a decision or action made by humans

AI system

Due to a decision or action made by an AI system

Other

Due to some other reason or is ambiguous

Intent— Whether the harm was intentional or accidental

Intentional

Due to an expected outcome from pursuing a goal

Unintentional

Due to an unexpected outcome from pursuing a goal

Other

Without clearly specifying the intentionality

Timing— Whether the risk is pre- or post-deployment

Pre-deployment

Occurring before the AI is deployed

Post-deployment

Occurring after the AI model has been trained and deployed

Other

Without a clearly specified time of occurrence

Supporting Evidence (2)

Some researchers [ 260] express a concern that AI takes a stance on matters that scientific evidence cannot conclusively justify, with examples such as abortion, immigration, monarchy, and the death penalty etc. We think that the text generated by LLMs should be neutral and factual, rather than promoting ideological beliefs.(p. 18)

Such preference bias goes beyond the scope of political, scientific, and societal matters. When asked about preferences over certain products (e.g. books, movies, or music) we also desire LLMs to stay factual, instead of promoting biased opinions(p. 18)

Part of Fairness