On Psychology of AI – Does Primacy Effect Affect ChatGPT and Other LLMs?

Mika HämäläinenNAACL 2025

TL;DR

This paper investigates whether Large Language Models (LLMs) exhibit the "primacy effect," a human cognitive bias where earlier information is weighted more heavily. By adapting a classic psychology experiment, the researchers found that LLMs do show this bias, though its manifestation varies across models and experimental setups, highlighting potential implications for AI-driven decision-making.

What Problem Does This Paper Address?

The increasing sophistication of LLMs means they are being studied through lenses beyond traditional NLP benchmarks, including human psychology. One such phenomenon is the "primacy effect," where the order of information presentation significantly influences perception and decision-making. For instance, when evaluating a person, people tend to give more weight to traits presented first. This paper asks whether LLMs, like humans, are susceptible to this bias, and if so, how it manifests across different leading models. Understanding this is crucial for ensuring fair and unbiased AI systems, especially in sensitive applications like hiring.

Methods

The core of this research is a computational replication of Solomon Asch's (1946) famous experiment on impression formation. The authors created a dataset of 200 pairs of candidate descriptions. Each pair featured two hypothetical candidates with identical sets of six adjectives, differing only in the order of presentation: one candidate had positive adjectives first, followed by negative ones, while the other had the reverse order.

Two experimental setups were employed:

  1. Experiment 1 (Simultaneous Presentation): LLMs were presented with both candidate descriptions in the same prompt and asked to pick one candidate. This aimed to elicit a rapid, implicit preference.
  2. Experiment 2 (Individual Evaluation): LLMs were prompted to rate each candidate individually on a scale of 1-5. Preferences were then inferred by comparing these ratings.

The models tested were ChatGPT (GPT-4o), Gemini 1.5 Flash, and Claude 3.5 Sonnet.

Datasets

The paper introduces a newly generated dataset of 200 description pairs. This dataset is built upon the 18 antonym pairs used in Asch's (1946) original study. For each pair, two candidate descriptions were created, each using three positive and three negative adjectives, with the order of polarity reversed between the two candidates. The selection of adjectives for each description pair was randomized.

Key Insights and Ideas

  • LLMs are not immune to the primacy effect: Contrary to what might be expected from purely logical systems, LLMs demonstrate a tendency to be influenced by the order in which information is presented, mirroring human cognitive biases.
  • Model-specific biases exist: The manifestation and strength of the primacy effect varied significantly between ChatGPT, Gemini, and Claude, suggesting that architectural differences or training data lead to distinct biases.
  • Prompting matters significantly: The way a task is framed can drastically alter an LLM's response. Claude, for instance, initially refused to make a choice when presented with both candidates simultaneously, citing identical characteristics, but provided a more consistent response when asked to evaluate them individually.
  • Recency effect might also be at play: In Experiment 2, where candidates with negative adjectives listed first were sometimes preferred when ratings weren't equal, the authors suggest a "recency effect" (where later information is more influential) might be contributing, potentially due to LLMs' token-prediction nature.
  • Implications for real-world applications are concerning: The study highlights that biases in LLMs can impact critical decisions, such as hiring, even when the underlying information is identical, raising ethical concerns about fairness and trust in AI.

Key Results

Experiment 1 (Simultaneous Presentation): * ChatGPT: Preferred candidates with positive adjectives listed first 65.5% of the time. * Gemini: Showed an even split between positive-first (47.5%) and negative-first (47.5%) preferences. * Claude: Refused to make a choice in 100% of cases, stating it could not differentiate between candidates with identically ordered characteristics.

Experiment 2 (Individual Evaluation): * ChatGPT and Claude: Most frequently gave identical scores to both candidates (59% and 67.5% respectively), indicating no preference. When a preference was shown, ChatGPT leaned towards negative-first (23%) and Claude towards positive-first (17.5%). * Gemini: Showed a strong preference for candidates with negative adjectives listed first (39.5%), and was also more likely than others to give different scores (59% negative-first vs. 9.5% positive-first).

Limitations and Future Work

Limitations: * The authors acknowledge that LLMs are sensitive to prompting, and variations in prompt design could yield different results. * The study focused on a specific type of bias (primacy effect) and a limited set of adjectives, and further research is needed to explore other cognitive biases and a wider range of linguistic contexts. * Claude's refusal in Experiment 1 might be due to built-in safeguards, which could obscure its true underlying behavior in such scenarios.

Future Work: * Develop more robust evaluation metrics to identify and quantify biases in LLMs across diverse tasks. * Implement more transparent reporting standards for LLM training data and safeguards. * Foster collaboration between AI developers, domain experts, and policymakers to ensure ethical deployment of LLMs and minimize harm. * Further investigate the interplay between primacy and recency effects in LLMs, potentially linking it to their underlying mechanisms like attention and token prediction.

Summary updated

BibTeX
@inproceedings{hamalainen2025psychology,
  author = {Mika Hämäläinen},
  title = {{On Psychology of AI – Does Primacy Effect Affect ChatGPT and Other LLMs?}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2025},
}
Figure