Source-linked AI summary
Artificial intelligence in communication impacts language and social relationships
Jess Hohenstein, Dominic DiFranzo, Rene F. Kizilcec, Zhila Aghajari, Hannah Mieczkowski, Karen Levy, Mor Naaman, Jeff Hancock, Malte Jung
TL;DR
AI-generated smart replies are widely used, but their effects on communication and social relationships remain insufficiently understood. Across two randomized experiments, the paper tests how smart replies alter interaction and perception. The results show greater efficiency, more positive emotional language, and improved evaluations with actual use, alongside more negative evaluations when use is suspected and potential changes to personal expression.
Problem
The social effects of integrating AI-generated messages into human-to-human communication remain understudied despite AI’s growing role in communication.
Method
Two randomized experiments examine how displaying and using smart replies affects real-time text-based interaction and interpersonal perception.
Results
Actual smart reply use increases communication efficiency, positive emotional language, and partners’ interpersonal evaluations, while suspected use receives more negative evaluations.
Takeaways & Limitations
AI can improve communication and interpersonal perceptions while altering emotional language and carrying risks for personal expression and social evaluation.
Takeaways & Limitations
The one-time interactions with anonymous crowdworkers do not establish how smart replies affect perceptions over time or in socially intimate relationships.
Abstract
from arXiv · showhide
Artificial intelligence (AI) is now widely used to facilitate social interaction, but its impact on social relationships and communication is not well understood. We study the social consequences of one of the most pervasive AI applications: algorithmic response suggestions ("smart replies"). Two randomized experiments (n = 1036) provide evidence that a commercially-deployed AI changes how people interact with and perceive one another in pro-social and anti-social ways. We find that using algorithmic responses increases communication efficiency, use of positive emotional language, and positive evaluations by communication partners. However, consistent with common assumptions about the negative implications of AI, people are evaluated more negatively if they are suspected to be using algorithmic responses. Thus, even though AI can increase communication efficiency and improve interpersonal perceptions, it risks changing users' language production and continues to be viewed negatively.
Introduction
AI-generated messages are increasingly embedded in human communication, but their social consequences remain understudied. This paper examines how smart replies shape communication and interpersonal evaluations.
- Research gap: Research has largely emphasized AI’s technical aspects while overlooking how AI-generated messages affect human-to-human communication.The authors frame this as an under-investment in research on AI’s social implications.
- Why communication matters: Communication shapes perceptions of others, social relationships, and cooperative outcomes, making AI-mediated language socially consequential.
- AI-mediated communication: Smart replies are a highly visible AI application designed to help users compose text with one tap.They generate suggested responses from general and user-specific text corpora.
- Study aim: The study uses randomized controlled experiments to test how displaying and using smart replies affects interaction and interpersonal perception.
- Study aim: The authors find that AI influences communication efficiency, emotional tone, and interpersonal evaluations in both positive and negative ways.
Results
The experiments show that smart replies increase communication efficiency and can improve sentiment and interpersonal evaluations, while perceived use is associated with more negative judgments. Actual and perceived AI use therefore have contrasting social consequences.
- Smart reply use: 14.3% of sent messages used smart replies on average, showing that making them available strongly encouraged adoption.The availability manipulation generated experimentally induced variation in use.
- Communication efficiency: Increased self-use improved communication efficiency, measured by the number of messages the self sent per minute, whereas partner use did not.
- Perceived use: Perceived partner use correlated weakly with actual use (Pearson’s r=0.22) and predicted lower ratings of cooperation and affiliation.These associations remained after controlling for the partner’s actual smart reply use.
- Actual use: Actual increased partner use improved the self’s ratings of the partner’s cooperation and affiliation despite negative judgments of perceived use.The cooperation estimate was t(167)=2.23, p=0.0273, and the affiliation estimate was t(167)=2.54, p=0.0120.
- Conversation sentiment: More smart reply use by either conversational partner produced more positive conversation sentiment.The authors link this pattern to changes in language introduced by the AI system.
- Emotional content: Positive and Google smart replies produced more positive emotional content than negative or no smart replies.Google smart replies had a similar effect to positive, but not negative, smart replies, indicating a positive sentiment bias.
Discussion
Commercially deployed AI reshapes communication in both beneficial and harmful ways: it can improve efficiency and interpersonal perceptions while altering emotional language and inviting negative judgments about suspected use.
- Discussion: AI use increases communication efficiency and produces more emotionally positive language, but suspected algorithmic use lowers perceived cooperativeness and affiliation.These opposing effects show that actual use and perceived use can produce different interpersonal consequences.
- Discussion: Actual smart reply use is associated with more positive partner attitudes, despite negative perceptions of using AI to communicate.The authors conclude that perceived stigma does not match the observed interpersonal effects of actual use.
- Discussion: AI-generated responses alter emotional expression, raising concerns about potential long-term effects on personal communication styles and expression.The paper notes that AI-mediated communication is already widespread while its long-term implications remain poorly understood.
- Discussion: Smart replies offer developers a way to influence conversational dynamics by controlling the linguistic qualities of suggested responses.The same mechanism creates risk because changes in users’ language remain consistent with the wording characteristics of the algorithmic replies.
- Discussion: AI’s efficiency and interpersonal benefits are coupled with altered emotional language and a potential loss of personal expression.The authors therefore frame smart replies as simultaneously useful and socially consequential.
Methods
The study used a web-based real-time messaging platform to manipulate smart replies, recruit participants online, and measure communication and interpersonal outcomes. Analyses combined experimental assignment, participant surveys, and sentiment scoring, with follow-up attention to contextual limitations.
- Platform: The web application enabled two participants to text chat in real time while researchers controlled and recorded smart replies.The platform supported browser-based participation and was designed for online interpersonal communication tasks.
- Experimental conditions: Four messenger conditions provided positive, negative, Google-generated, or no smart replies, with manual typing required in the control condition.Positive and negative suggestions were selected from predefined examples, while the Google condition used Google’s Reply model.
- Participants: 438 Mechanical Turk participants, aged 18–68, were recruited for the study.Participants received monetary payment; mean age was 34.15 years (SD=10.1).
- Procedure and measures: Participants discussed unfairly rejected work with an anonymous partner, then reported perceived smart-reply use, cooperative communication, and interpersonal traits.The study measured cooperation using a 7-item scale and affiliation and dominance using selected IAS-R adjectives.
- Sentiment analysis: Positive and Google smart replies increased conversation affect, whereas negative smart replies reduced it relative to no-smart-reply conversations.The positive and Google conditions had statistically similar effects; VADER findings were confirmed using LIWC.