Oh F**k! How Do People Feel about Robots that Leverage Profanity?

TL;DR

This work investigates whether robots should use profanity to improve social perception during error repair. Across three exploratory studies—online with university students, online with a broader U.S. sample, and in-person campus deployment—the authors compare no-speech, non-expletive, and expletive error responses. Verbal acknowledgment generally improves perceived warmth, competence, humor, and interpersonal closeness, while expletives increase discomfort in some contexts; however, differences between non-expletive and expletive speech are not consistently significant, suggesting a nuanced design space. The findings highlight a largely positive reception to speech-based error responses in the U.S. context and argue for context-aware, personalized use of profanity as a potentially engaging tool for human-robot interaction, while also noting cultural sensitivities and safety considerations for minors. Overall, the work contributes replication-backed insights into how uncouth robot behavior can broaden the design space for social robots beyond conventional politeness.

Abstract

Profanity is nearly as old as language itself, and cursing has become particularly ubiquitous within the last century. At the same time, robots in personal and service applications are often overly polite, even though past work demonstrates the potential benefits of robot norm-breaking. Thus, we became curious about robots using curse words in error scenarios as a means for improving social perceptions by human users. We investigated this idea using three phases of exploratory work: an online video-based study (N = 76) with a student pool, an online video-based study (N = 98) in the general U.S. population, and an in-person proof-of-concept deployment (N = 52) in a campus space, each of which included the following conditions: no-speech, non-expletive error response, and expletive error response. A surprising result in the outcomes for all three studies was that although verbal acknowledgment of an error was typically beneficial (as expected based on prior work), few significant differences appeared between the non-expletive and expletive error acknowledgment conditions (counter to our expectations). Within the cultural context of our work, the U.S., it seems that many users would likely not mind if robots curse, and may even find it relatable and humorous. This work signals a promising and mischievous design space that challenges typical robot character design.

Figures (8)

• Figure 1: Depiction of the Hello Robot Stretch mobile manipulator responding to an accident with an expletive.
• Figure 2: Still frames from video stimuli for bumping the table (left), knocking over the cup (center), and dropping the pen (right). The speech bubbles were added for this figure.
• Figure 3: Boxplots showing the survey responses for RoSAS warmth, competence, and discomfort in the initial study. In this figure and other boxplots throughout the paper, the boxes extend from the 25th to the 75th percentiles, the center horizontal line of the box marks the median, an asterisk ('*') marks the mean, whiskers show up to 1.5 times the interquartile range, and circles indicate outliers. Brackets indicate significant pairwise differences.
• Figure 4: Survey responses for IOS interpersonal closeness and JRS humorousness in the initial study.
• Figure 5: Survey responses for Godspeed anthropomorphism and likeability in the initial survey.