AI in Communication Impacts Relationships and Language

As AI sees dramatic increases in use, it is critical to understand its impact on social life. Through two large-scale online studies, we investigate the social consequences of using AI in communication.

tl;dr

• Commercially-deployed AI changes how people interact with and perceive each other
• For a message sender, the use of AI corresponds with increased communication efficiency, more positive emotional language, and better evaluations by the the message receiver
• Consistent with common assumptions about the negative implications of AI, senders are evaluated more negatively if they are suspected to be using AI
• Even though AI can increase communication efficiency and improve interpersonal perceptions, it risks changing users’ language production and continues to be viewed negatively

My Role

• Formulated research question based on my previous work and literature review
• Screened and recruited select participants using Mechanical Turk
• Designed a web-based messaging app
• Designed and executed remote, asynchronous study sessions (N=1036)
• Linguistically and statistically analyzed results using LIWC, Python, and R, respectively
• Wrote paper sharing findings (currently under review)

Methodology

We ran a remote experiment that incorporated our own messaging app with an online survey.

We developed a web-based AI-MC platform to control and record the smart replies that participants see. This figure shows both positive and negative sentiment smart reply examples (i.e., blue and grey boxes, respectively). During actual use, users see only one of those sets.
Participants (N=1036) used our messaging app to complete a task with an anonymous partner.

• The first study investigated interpersonal effects of AI through three conditions: 1) both participants can use smart replies (i.e., suggested responses generated using the Google Reply API ), 2) only one participant can use smart replies, or 3) neither participant can use smart replies

• The follow-up investigated linguistic effects of AI through four conditions: with Google smart replies, positive smart replies (i.e., participants receive suggested responses that have positive sentiment, as rated by crowdworkers), negative smart replies (i.e., participants receive suggested responses that have negative sentiment, as rated by crowdworkers), or no smart replies

• In the first study, participants rated how much they believed their partner had used smart replies and filled out the Perceived Cooperative Communication scale and Revised Interpersonal Adjective Scale (IAS-R)

Main Findings

People use smart replies when given the opportunity, and this increases the efficiency of communication and leads to more emotionally positive language.

Average proportion of smart reply use in conversation by experimental condition (i.e., first-stage effect)
As participants think that their partner is using more algorithmic responses, they perceive them as less cooperative and feel less affiliation towards them.
Latent average treatment effect (LATE) estimates of how increased smart reply use by the self and by the partner affected conversation sentiment, communication efficiency, cooperative communication, and affiliation
As people actually use more algorithmic responses, their communication partner has more positive attitudes about them.
AI language sentiment affects the emotional content of conversation.
Mean overall conversation sentiment by experimental condition

Recommendations

While AI has the potential to help people communicate more efficiently and improve interpersonal perceptions in everyday conversation, users should be cautioned that these benefits are coupled with changes to the emotional aspects of their language and a corresponding potential loss of personal expression.

Designers of AI systems could potentially affect conversational dynamics and outcomes by carefully controlling the linguistics of smart replies that are shown to users (example).

Perceptions of AI in communication are at odds with actual effects, and future work is need to determine how developers can improve these perceptions.
• Users of messaging apps that feature smart replies currently get no information about how these messages are generated, which could be increasing their uncertainty about and negatively affecting perceptions of AI
• If users are given background information or real-time feedback on how the AI works, this could alleviate some of the negative perceptions of smart reply use that we have seen
• The amount and type of information that needs to be disclosed for such a system to be considered sufficiently transparent warrants further investigation

You might also want to see: