# Influence of solution efficiency and valence of instruction on additive and subtractive solution strategies in humans, GPT-4, and GPT-4o

## TL;DR

This study shows that both humans and large language models like GPT-4 and GPT-4o prefer additive problem-solving strategies, with models showing a stronger bias.

## Contribution

The research identifies and compares addition bias in humans and LLMs, revealing context-dependent amplification in models.

## Key findings

- Addition bias was stronger in LLMs than in humans across both spatial and linguistic tasks.
- GPT-4 showed the opposite efficiency pattern to humans in additive choices.
- Positive instruction valence increased additive outputs in LLMs but not consistently in humans.

## Abstract

Generative artificial intelligences, particularly Large Language Models (LLMs), increasingly influence human decision-making, making it essential to understand how cognitive biases are reproduced or amplified in these systems. Building on evidence of the human “addition bias” – a preference for additive over subtractive problem-solving strategies1 – this research compared humans with GPT-4 (Study 1) and GPT-4o (Study 2) in spatial and linguistic tasks. Study 1 comprised four experiments (1a, 1b, 2a, 2b) with 588 human participants and 680 GPT-4 outputs; Study 2 included two experiments (3a, 3b) with 751 human participants and 1,080 GPT-4o outputs. We manipulated (a) solution efficiency and (b) instruction valence. Across both studies, a general addition bias emerged, more pronounced in the LLMs than in humans. Humans made fewer additive choices when subtraction was more efficient than addition (compared to when both were equally efficient), whereas GPT-4’s output showed the opposite pattern. GPT-4o’s outputs aligned with those of humans in the linguistic task but showed no efficiency effect in the spatial task. Instruction valence did not reach statistical significance for either agent in the spatial task. In the linguistic task, positive valence (compared to neutral valence) led to more additive outputs in both GPT models, but only in Study 2 for humans. These findings indicate that addition bias has been transferred to LLMs, which can replicate and, depending on context, amplify this human bias. This emphasizes the importance of further theoretical and empirical work on the cognitive and data-driven mechanisms underlying addition bias in both humans and LLMs.

Six experiments examined addition bias in humans and in outputs of large language models (GPT-4, GPT-4o). Both human and model-generated solutions showed a preference for additive over subtractive strategies, with a stronger bias in the model output.