# "AI enhances our performance, I have no doubt this one will do the same": The Placebo effect is robust to negative descriptions of AI

**Authors:** Agnes M. Kloft, Robin Welsch, Thomas Kosch, Steeven Villa  
**arXiv:** [2309.16606](https://arxiv.org/abs/2309.16606)  
**Date:** 2023-09-28  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2309.16606v1/

## Key points

- Study on human-AI interactions focused on impact of user expectations
- Participants performed better with sham-AI regardless of AI description, indicating placebo effect
- Cognitive modeling showed participants gathered more information, leading to improved performance
- Replication study confirmed negative AI descriptions did not alter expectations, suggesting biased but robust performance expectations with AI
- Questionnaires like UEQ-S and SUS used to assess user experience
- Electrodermal activity (EDA) recording and pre-processing techniques employed using standard electrodes and data acquisition tools
- Experiment conducted on Linux laptop with specific hardware specifications for optimal performance
- Custom software used for precise control over variables during experiment
- Detailed procedure followed after consent, including attaching EDA electrodes and providing demographic information

## Summary

In a study on human-AI interactions, researchers explored the impact of user expectations on performance outcomes. The participants were informed that an AI system would either enhance or decrease their performance in a letter discrimination task. However, no actual AI was present in any condition. Surprisingly, participants performed better regardless of the AI description when a sham-AI was introduced. This indicates high expectations and a placebo effect at play. Cognitive modeling revealed that participants gathered more information, leading to improved performance. A replication study confirmed that negative AI descriptions did not alter expectations. This suggests biased but robust performance expectations with AI. To assess user experience, questionnaires such as the UEQ-S and SUS were utilized to measure pragmatic and hedonic quality as well as usability of the AI system. Electrodermal activity (EDA) recording and pre-processing techniques were employed using standard electrodes and data acquisition tools. Verbal descriptions provided to participants varied between positive and negative portrayals of the AI system's capabilities and reliability. The experiment was conducted on a Linux laptop with specific hardware specifications for optimal performance. Custom software was used for running the experiment locally, ensuring precise control over variables. Participants underwent a detailed procedure after consenting, including attaching EDA electrodes and providing demographic information. Overall, the study highlighted the robust nature of the placebo effect in human-AI interactions, emphasizing the importance of managing user expectations for improved performance outcomes. The findings contribute valuable insights into evaluating and enhancing user experiences with AI technology in real-world scenarios.