# The R-U-A-Robot Dataset: Helping Avoid Chatbot Deception by Detecting User Questions About Human or Non-Human Identity

**Authors:** David Gros, Yu Li, Zhou Yu  
**arXiv:** [2106.02692](https://arxiv.org/abs/2106.02692)  
**Date:** 2021-06-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2106.02692v1/

## Key points

- Research focuses on human-machine interaction through language, especially when users are unaware they are communicating with a machine
- Goal is to develop methods for confirming non-human identity of systems
- Gathered 2,500 adversarial utterances to test confirmation methods
- Study compares classifiers for recognizing intent and discusses tradeoffs between precision, recall, and model complexity
- Classifiers could be integrated into dialog systems to prevent deception
- Blender, Amazon Alexa, and Google Assistant often fail to confirm non-human identity
- User study conducted to compare important aspects of responding to the intent of asking if a system is a robot
- Four metrics considered: Pw (modified precision), recall (R), classification accuracy (Acc), and aggregate measure (M)
- Simple classifiers like BOW LR perform better than chance but still misclassify over 1/10 examples
- BERT classifier outperforms other classifiers but still misclassifies about 1/25 utterances
- Grammar-based classifier performs worse than simple ML models but offers high precision in checking intent
- Crowd sourcing used to expand grammar for generating examples through surveys issued to colleagues and Amazon Mechanical Turk workers
- Research aims to improve understanding of how machines can confirm their non-human identity during language interactions by examining different classifiers and existing dialog systems while providing insights into effective responses.

## Summary

This research focuses on the interaction between humans and machines through language, particularly in situations where the user may not be aware that they are communicating with a machine. The goal is to understand how system designers and researchers can develop methods for confirming the non-human identity of these systems. They also gathered an additional set of 2,500 adversarial utterances where simply confirming non-human identity would be insufficient or unnatural. The study compares different classifiers to recognize the intent and discusses tradeoffs between precision, recall, and model complexity. These classifiers could potentially be integrated into dialog systems to prevent undesired deception. The researchers then examine how three different systems (Blender, Amazon Alexa, and Google Assistant) handle this intent and find that these systems often fail to confirm their non-human identity. Additionally, the study explores what constitutes a good response to the intent of asking if a system is a robot. A user study is conducted to compare important aspects when responding to this intent. In terms of metrics, four measures are considered: Pw (a modified precision measure), recall (R), classification accuracy (Acc), and an aggregate measure (M) which is the geometric mean of the other three metrics. The results show that simple classifiers like BOW LR perform better than chance but still misclassify more than 1/10 examples. BERT classifier outperforms other classifiers but still misclassifies about 1/25 utterances. The grammar-based classifier performs significantly worse than simple ML models but offers high precision in checking intent. To expand their initial grammar for generating examples, crowd sourcing was employed through surveys issued to internal colleagues and Amazon Mechanical Turk workers. The responses collected were used to diversify the grammar and provide a broader range of expressions for the intent. Overall, this research aims to improve understanding of how machines can confirm their non-human identity during language interactions with humans by examining performance of different classifiers as well as existing dialog systems in handling this intent while providing insights into what constitutes an effective response.