I recently read Professor Philip Resnik’s thought-provoking position paper, “Large Language Models Are Biased Because They Are Large Language Models,” published in Computational Linguistics 51(3), which is available via open access. This paper challenges conventional perspectives on bias in artificial intelligence, prompting a deeper examination of the inherent relationship between bias and the foundational design of large language models (LLMs). Resnik’s primary objective is to stimulate critical discussion by arguing that harmful biases are an inevitable outcome of the current architecture of LLMs. The paper posits that addressing these biases effectively requires a fundamental reevaluation of the assumptions underlying the design of AI systems driven by LLMs.
What the paper argues
- Bias is built into the very goal of an LLM. A language model tries to predict the next word by matching the probability patterns of human text. Those patterns come from people. People carry stereotypes, norms, and historical imbalances. If an LLM learns the patterns faithfully, it learns the bad with the good. The result is not a bug that appears once in a while. It is a direct outcome of the objective the model optimizes.
- Models cannot tell “what a word means” apart from “what is common” or “what is acceptable.” Resnik uses a nurse example. Some facts are definitional (A nurse is a kind of healthcare worker). Other facts are contingent but harmless (A nurse is likely to wear blue clothing at work). Some patterns are contingent and harmful if used for inference (A nurse is likely to wear a dress to a formal occasion). Current LLMs do not have an internal line that separates meaning from contingent statistics or that flags the normative status of an inference. They just learn distributions.
- Reinforcement Learning from Human Feedback (RLHF) and other mitigations help on the surface, but they have limits. RLHF tries to steer a pre-trained model toward safer outputs. The process relies on human judgments that vary by culture and time. It also has to keep the model close to its pretraining, or the model loses general ability. That tradeoff means harmful associations can move underground rather than disappear. Some studies even find covert bias remains after mitigation (Gallegos et al. 2024; Hofmann et al. 2024). To illustrate this, consider an analogy: The balloon gets squeezed in one place, then bulges in another.
- The root cause is a hard-core, distribution-only view of language. When meaning is treated as “whatever co-occurs with what,” the model has no principled way to encode norms. The paper suggests rethinking foundations. One direction is to separate stable, conventional meaning (like word sense and category membership) from contextual or conveyed meaning (which is where many biases live). Another idea is to modularize competence, so that using language in socially appropriate ways is not forced to emerge only from next-token prediction. None of this is easy, but it targets the cause rather than only tuning symptoms.
Why this matters
Resnik is not saying we should give up. He is saying that quick fixes will not fully erase harm when the objective rewards learning whatever is frequent in human text. If we want models that reason with norms, we need objectives and representations that include norms, not only distributions.
Conclusion
This paper offers a clear message. Bias is not only a content problem in the data. It is also a design problem in how we define success for our models. If the goal is to build systems that are both capable and fair, then the next steps should focus on objectives, representations, and evaluation methods that make room for norms and constraints. That is harder than prompt tweaks, but it is the kind of challenge that can move the field forward.
Link to the paper: Large Language Models Are Biased Because They Are Large Language Models
— Andrew
5,279 hits
CompLing Insights 