TLDR: A new study investigates anchoring bias in large language models (LLMs), a cognitive bias where initial numbers influence subsequent judgments. Using log-probability analysis and Shapley values, the research shows that LLMs like Gemma-2B, Phi-2, and Llama-2-7B exhibit robust anchoring, with anchors actively reweighting internal probabilities. Smaller models show more variability. The study introduces an Anchoring Bias Sensitivity Score (ABSS) and highlights risks for LLM deployment, emphasizing the need for bias-aware governance.
Large language models (LLMs) are becoming increasingly central to various applications, from assisting in decision-making to acting as substitutes for human participants in research. However, a critical question remains: do these models truly understand and replicate human cognitive phenomena, or are they merely imitating surface-level behaviors? A recent research paper, “Anchors in the Machine: Behavioral and Attributional Evidence of Anchoring Bias in LLMs”, delves into this question by examining anchoring bias, a well-known human judgment bias, within LLMs.
Anchoring bias occurs when an individual’s judgment is systematically influenced by an initial piece of information, or ‘anchor,’ even if that anchor is irrelevant. For example, if asked to estimate a percentage, people tend to give higher estimates after seeing a high anchor and lower estimates after seeing a low anchor. Previous studies have shown that LLMs exhibit similar patterns, but most of this evidence relied on observing their final outputs, without exploring the internal mechanisms driving these biases.
A Deeper Look into LLM Biases
The paper, authored by Felipe Valencia-Clavijo from Dataplicada, makes three significant contributions to understanding anchoring bias in LLMs. First, it uses a log-probability-based behavioral analysis to show that anchors don’t just change a single output; they shift the entire distribution of possible outputs. This goes beyond simple observation and provides a more robust measure of the bias, while also controlling for potential training-data contamination.
Second, the research employs Shapley-value attribution, a method from cooperative game theory adapted for AI, to precisely quantify how much an anchor influences the model’s internal log-probabilities for specific predictions. This technique helps to open the ‘black box’ of LLMs, making their decisions more transparent.
Finally, the study introduces a unified Anchoring Bias Sensitivity Score (ABSS). This score integrates both the behavioral evidence (how outputs shift) and the attributional evidence (how much the anchor contributes internally) across different models, offering a comprehensive way to evaluate an LLM’s susceptibility to anchoring bias.
Key Findings Across Six Models
The study tested six open-source LLMs: Gemma-2B, Phi-2, Llama-2-7B, GPT-2, Falcon-RW-1B, and GPT-Neo-125M. The results revealed robust anchoring effects in Gemma-2B, Phi-2, and Llama-2-7B. For these models, the attribution analysis confirmed that the anchors were actively reweighting the internal probabilities, indicating a deeper influence rather than just a superficial imitation.
Interestingly, smaller models like GPT-2, Falcon-RW-1B, and GPT-Neo-125M showed more variability in their anchoring sensitivity. This suggests that the scale of an LLM might play a role in how strongly it exhibits this bias. The research also found that the attributional effects could vary depending on how the prompts were designed, highlighting the delicate nature of treating LLMs as direct substitutes for human participants.
The Anchoring Bias Sensitivity Score (ABSS) ranked Gemma-2B, Phi-2, and Llama-2-7B as the most biased models, with high positive ABSS values. GPT-2, Falcon-RW-1B, and GPT-Neo-125M were lower-ranked, with GPT-Neo-125M even showing a negative ABSS, indicating attribution shifts in the opposite direction of anchoring for some cases.
Also Read:
- PRIME: A New Framework to Diagnose AI’s Stereotypical Reasoning
- Unpacking Self-Awareness in Large Language Models
Implications for AI Safety and Interpretability
These findings have significant implications. For researchers using LLMs as subjects, the presence of anchoring bias suggests that while models can reproduce psychology-style effects, the fragility of attribution across different scenarios means they shouldn’t be seen as simple replacements for human participants. For those deploying LLMs as decision systems, the fact that arbitrary cues can systematically shift outputs raises serious concerns for safety and governance in high-stakes fields like healthcare, finance, and law.
The study emphasizes that recognizing and quantifying such biases is crucial for responsible AI deployment. Potential mitigation strategies could include filtering layers to remove irrelevant numeric elements or using broader techniques like chain-of-thought prompting. Ultimately, bias-aware governance frameworks will be essential as LLMs become more integrated into workflows where cognitive biases could otherwise propagate unchecked.
This research not only deepens our understanding of how LLMs align with or diverge from human-like biases but also provides a reproducible framework for measuring these biases in models with open log-probabilities. It strengthens the bridge between behavioral science, LLM safety, and interpretability, offering a structured middle ground for explaining complex AI behaviors.
