Working Paper 24-078

AI Companions Reduce
Loneliness
Julian De Freitas
Ahmet K. Uguralp
Zeliha O. Uguralp
Puntoni Stefano
AI Companions Reduce Loneliness

Julian De Freitas
Harvard Business School

Ahmet K. Uguralp
Bilkent University

Zeliha O. Uguralp
Bilkent University

Puntoni Stefano
University of Pennsylvania, Wharton School

Working Paper 24-078

Copyright © 2024 by Julian De Freitas, Ahmet K. Uguralp, Zeliha O. Uguralp, and Puntoni Stefano.

Working papers are in draft form. This working paper is distributed for purposes of comment and discussion
only. It may not be reproduced without permission of the copyright holder. Copies of working papers are
available from the author.

Funding for this research was provided in part by Harvard Business School.
 AI Companions Reduce Loneliness

ABSTRACT

Chatbots are now able to engage in sophisticated conversations with consumers in the domain of

relationships, providing a potential coping solution to widescale societal loneliness. Behavioral

research provides little insight into whether these applications are effective at alleviating

loneliness. We address this question by focusing on “AI companions”: applications designed to

provide consumers with synthetic interaction partners. Studies 1 and 2 find suggestive evidence

that consumers use AI companions to alleviate loneliness, by employing a novel methodology

for fine-tuning large language models (LLMs) to detect loneliness in conversations and reviews.

Study 3 finds that AI companions successfully alleviate loneliness on par only with interacting

with another person, and more than other activities such watching YouTube videos. Moreover,

consumers underestimate the degree to which AI companions improve their loneliness. Study 4

uses a longitudinal design and finds that an AI companion consistently reduces loneliness over

the course of a week. Study 5 provides evidence that both the chatbots’ performance and,

especially, whether it makes users feel heard, explain reductions in loneliness. Study 6 provides

an additional robustness check for the loneliness-alleviating benefits of AI companions.

Keywords: generative AI, chatbots, loneliness, large language models, artificial intelligence,

empathy, longitudinal, AI companion.

Declarations of interest: None

 Thanks to advances in ‘generative AI’ algorithms, AI companions are now commercially

available. These are applications that utilize artificial intelligence (AI) to offer consumers the

opportunity to engage in emotional interactions like friendship and romance. AI companions are

synthetic interaction partners that offer emotional support, distinguishing them from broader

relational AI applications that include varied human-computer interactions such as virtual

customer service agents, educational tutors, AI assistants such as OpenAI’s ChatGPT, or voice-

activated assistants like Siri. Algorithms used in AI companions produce complex answers to an

exceedingly wide range of prompts, making them sufficiently mature to power sophisticated

interactions with consumers. Although these systems are truly incapable of feeling real emotions,

concern, or caring, they can generate language that creates the perception of empathy.

Generative AI is forecasted to grow into an impressive $1.3 trillion market by 2032

(Catsaros 2023), suggesting a concomitant rise in AI companion platforms. This is already

evident in platforms such as XiaoIce (xiaoice.com, with 660 million users), Chai (chai-

research.com, with 4 million active users), and Replika (replika.com, with 2.5 million active

users), among others. A user can ask their AI companion questions, and it will respond in a

natural, believable way. The AI companion can also initiate conversations itself, such as “How

are you feeling” or “Are you mad at me?” Consumers may use these platforms for both friendly

and romantic purposes. For example, around 50% of Replika users have a romantic relationship

with the AI (De Freitas and Tempest Keller 2022). Here we consider the value proposition that

AI companions reduce loneliness, inspired by an interview we conducted with the CEO of

Replika and her investors. All suggested that consumers are using the app because they are

lonely, and that the apps do in fact reduce feelings of loneliness. Inspired by this observation, we

make several contributions.

 First, research in psychology indicates that loneliness is a powerful emotional state that

urges consumers to seek social connections (Hawkley and Cacioppo 2010; Holt-Lunstad et al.

2015). In the absence of available human interaction, individuals may engage with alternative

forms of companionship, such as AI companion apps. Motivated by this, we contribute to an

understanding of whether loneliness motivates consumers to use AI companions (Ta et al. 2020),

by detecting instances of loneliness in conversations on a real AI companion app and reviews of

several such apps. To do this, we use Large Language Models (LLMs) that we fine-tune to detect

loneliness, enabling us to detect loneliness more accurately than traditional dictionary-based

approaches. Encouragingly, we find, consistent with recent findings on LLMs (Solaiman and

Dennison 2021), that only a small dataset of exemplars is required to specialize the LLM for our

purposes, likely due to the rich contextual knowledge that the baseline models leverage for the

fine-tuning task.

Second, and most important, we explore whether conversations with AI companions help

to alleviate feelings of loneliness, contributing to work on the efficacy of technological solutions

like social robots in helping consumers cope with loneliness (Shrum, Fumagalli, and Lowrey

2022; Veronese et al. 2021). In doing so, we study consumer loneliness both before versus after

interacting with AI companions through textual conversation, at both cross-sectional and

longitudinal scales. Thus, we also measure how consumers behaviorally interact with

commercially representative versions of the technology, unlike most studies in the consumer

behavior literature (De Freitas et al. 2023b).

Finally, we contribute to understanding what features of chatbots lead to alleviation of

loneliness (Merrill Jr, Kim, and Collins 2022), by systematically manipulating the conversational

performance of the chatbot, and the ability of the chatbot to make the consumer feel heard—a
construct involving the perception that the communication is received with attention, empathy,

and respect (Roos, Postmes, and Koudenburg 2023).

CONCEPTUAL FRAMEWORK

Previous work in marketing has highlighted the power of interactive media, including

chatbots (Hoffman and Novak 1996). Studies on chatbots have mostly focused on AI assistants

such as customer service chatbots, investigating consumers’ responses to chatbot interactions.

These studies highlight the potential of chatbots to both optimize customer service operations

and positively impact consumer attitudes and behaviors (e.g., Chung et al. 2020), especially

when consumer are unaware of the bot’s identity (Luo et al. 2021), and have shown that more

‘concrete’ language (Jiménez-Barreto, Rubio, and Molinillo 2023) and ‘embodied’ chat

interfaces (Bergner, Hildebrand, and Häubl 2023) lead to more favorable consumer-brand

relationships and consumer satisfaction. On the flips side, this literature finds that using a bot can

negatively impact attitudes toward firms, due to the belief that the use of chatbots prioritizes firm

benefits over customer benefits (Castelo et al. 2023), and that bots may respond inappropriately

to unanticipated messages from users (De Freitas et al. 2023a). Overall, while these studies focus

on the benefits and drawbacks of AI assistants like customer service bots, our research extends

this literature to consumer relationships with AI companions.

Specifically, we investigate whether consumer interactions with AI companions alleviate

loneliness and the mechanisms behind this effect. In interactions with machines and AI,

consumers are affected by social or human cues like the use of anthropomorphic features (Araujo

2018; Crolic et al. 2022), as well as avatars and other indicators of physical and verbal

embodiment (Bergner et al. 2023; Bertacchini, Bilotta, and Pantano 2017; Holzwarth,

Janiszewski, and Neumann 2006; Zierau et al. 2023). Depending on the interface, consumers
may also apply the same social norms of human-human interactions to their interactions with

computers (Nass and Moon 2000). In the domain of consumer-brand relationships, consumers

can build relationships with brands via similar processes that they use to build relationships with

other people (Fournier 1998; Muniz Jr and O'guinn 2001), and these brand relationships can

affect their subjective experiences and behaviors (Brakus, Schmitt, and Zarantonello 2009; Esch

et al. 2006). We complement these research streams by considering consumer behavioral

interactions with AI companions, which are literately, rather than just figuratively, designed and

optimized for social relationships. Here we investigate whether interacting with such AI

alleviates loneliness.

Can AI companions Help Cope with Loneliness?

Loneliness is a state of subjective, aversive solitude characterized by a discrepancy

between actual and desired levels of social connection (Perlman and Peplau 1982). Loneliness is

often not problematic, with almost everyone experiencing loneliness from time to time

(Cacioppo and Cacioppo 2018). Yet some people are not successful at alleviating loneliness,

leading to a state of chronic loneliness that is associated with depression, anxiety, and physical

health outcomes at levels worse than obesity (Palgi et al. 2020). The size of the population

suffering from chronic loneliness is both sizable and increasing, with estimates in the U.S.

ranging from 30% to 60% (Holt-Lunstad, Robles, and Sbarra 2017; Ipsos 2021). The U.S., U.K.

and Japan have all identified loneliness as a health pandemic, assigning Ministers of Loneliness

(in the U.K. and Japan) and the U.S. Surgeon General to tackle this issue with nation-wide

initiatives.
 Meta-analyses on which interventions are affective at alleviating loneliness find greater

reductions in loneliness after having high-quality, one-on-one interactions with another person,

especially ones that provide social support and social cognitive training; in contrast, some

technological interventions for reducing loneliness, as well as animal therapy or robot pets, were

not found to significantly reduce loneliness (Masi et al. 2011; Veronese et al. 2021), although

other studies find more promising results of such interventions with older adults (Banks,

Willoughby, and Banks 2008; Chen et al. 2020; Gasteiger et al. 2021). Some researchers have

also concluded that social media might not alleviate loneliness because it increases the number of

social connections without providing high-quality connections, with some research even

suggesting that social media might increase loneliness (Twenge, Catanese, and Baumeister

2003). Thus, an open question is what kinds of technological solutions could actually provide

effective mechanisms for coping with loneliness (Shrum et al. 2022).

Since AI companions are not just made to match the average human conversation but to

mimic a social interaction in which the conversation partner is conversationally competent (e.g.,

keeps track of context and responds in a timely manner) and makes the user feel heard, it is

possible that conversations with such a synthetic partner alleviate feelings of loneliness. After

all, talking about one’s problems to an active listener in psychotherapy is usually effective in

bringing some degree of relief (APA 2012). AI companions also have other attractive properties

as a large-scale solution for helping to combat societal loneliness. Since most AI companions

utilize freemium models in which basic conversations are available for free, they are a more cost-

effective solution than relevant alternatives like gaming or caring for a pet. Likewise, the need

for human involvement in human intervention programs limits their scalability compared to AI

companions, especially for potential beneficiaries living in remote areas or with limited mobility.
 Previous work has begun to indirectly explore the question of whether AI companions

reduce loneliness, mostly by interviewing existing app users (Ta et al. 2020). Another study

surveyed student users of Replika, and found that these participants were lonelier than the

average student and felt a high level of social support from Replika (Maples et al. 2024). The

overwhelming problem with these initial correlational results is that they do not allow a rigorous

test of the effects of AI companions on loneliness, for example because of the likely presence of

selection effects. To our knowledge, the current work is the first to causally assess whether

representative AI companions reduce loneliness.

A related question which has remained unaddressed is the extent to which any loneliness-

alleviating effect of AI companions is short-lived or can persist over longer timespans of

interaction, such as for a week. One possibility is that consumers experience diminishing returns

in terms of loneliness reduction, as they quickly come to view AI companions as lacking in

certain essential aspects. For instance, prior work studying a less capable and socially sensitive

chatbot than the ones employed in the current studies found that participants interacting with a

chatbot found it less enjoyable and more predictable over time (Croes and Antheunis 2021).

Consumers also have various negative attitudes toward AI that could psychologically interfere

with how they interact with the chatbots and the benefits they get from these interactions over

time, such as viewing AI as inscrutable black boxes that are unemotional, unable to learn, and

threatening because they can behave autonomously (De Freitas et al. 2023b). After all, AI

companions cannot feel any emotions, and most widely deployed ones do not have physical

bodies. Alternatively, and as we ultimately find, consumers might continue to experience

loneliness improvements even after interacting with a chatbot over multiple days.
How could AI relationships alleviate loneliness?

If AI companions alleviate loneliness, what could be the mechanisms for this? Here we

focus on the psychological construct of ‘feeling heard’—the perception that another individual

truly comprehends your thoughts, feelings, and preferences, and receives it with attention,

empathy, respect, and mutual understanding (Roos et al. 2023). The experience of feeling heard

plays a significant role in human-human relationships (e.g., Gable and Reis 2010; Reis, Lemay

Jr, and Finkenauer 2017). Feeling heard often involves empathy, where the listener not only

seems to understand the speaker but also shares the speaker’s emotions, deepening the sense of

being genuinely understood (Myers 2000). Social psychological studies find that feeling heard

yields several benefits in relationships, including higher trust between partners and higher well-

being (Reis et al. 2017), and, crucially, decreased feelings of loneliness following a social

rejection disclosure (Itzchakov et al. 2023). Similarly, another study found that a four-week

program of empathetic telephone calls decreased feelings of loneliness (Kahlon et al. 2021). In

this study, trained callers made regular phone calls lasting around 10 minutes to participants. The

frequency of these calls varied, ranging from 2 to 5 times a week, based on participants’

preferences.

Building on the idea that consumers might employ the same social norms with computers

as they do in human-human interactions (Nass and Moon 2000), work in human-computer

interaction has also studied the effects of empathetic AI on consumers in social settings. Relative

to chatbots that do not express empathy, chatbots that express empathy lead to more favorable

ratings of companionship, i.e., the activities done together are perceived as more enjoyable or

exciting (Boucher et al. 2021; Leite et al. 2013), and better mood after experiencing social

exclusion (De Gennaro, Krumhuber, and Lucas 2020). Previous marketing research also
underscores the value of empathetic AI interactions, showing that artificial empathy narrows the

customer experience gap between AI and human agents, with high empathy levels resulting in

comparable affective and social experiences to humans, particularly improving social

interactions (Liu-Thompkins, Okazaki, and Li 2022). Another study found that an initial warm

(vs. competent) message from chatbots significantly enhances consumers’ brand perception,

creating a closer brand connection and increasing the likelihood of engaging with the chatbot

(Kull, Romero, and Monahan 2021).

Academic studies aside, the very fact that AI companions with empathic personalities

have garnered so many users suggests that consumers are gaining social benefits from these

apps, which are also marketed as being caring. For example, Replika advertises that it is “here to

make you feel HEARD, because it genuinely cares about you”

(https://apps.apple.com/lt/app/replika-virtual-ai-friend/id1158555867).

Apart from feeling heard, another factor that could affect loneliness alleviation is the

chatbot’s performance, which consists of a range of features pertaining to managing the

conversation effectively, including: timely responses, perceived credibility, context tracking,

response variability, and domain knowledge (Chaves and Gerosa 2021). However, we

hypothesize that feeling heard is more critical in alleviating loneliness after AI companion usage

compared to communication performance, because one of the primary sources of loneliness is

the perceived lack of social and emotional support (Liu, Gou, and Zuo 2016; Masi et al. 2011).

In sum, previous research on the impact of interpersonal relationships on loneliness

alleviation emphasizes the critical role of feeling heard and understood; however, these studies

focus exclusively on human-human relationships and do not address experiences with AI

companions. To address this gap, our work causally investigates the effectiveness of AI
companions in alleviating loneliness. Specifically, we explore whether feeling heard and

performance mediate this effect. To do this, we compare an empathetic AI companion to two

types of chatbot: an AI assistant that does not express empathy, and a highly constrained chatbot

capable of performing a limited number of tasks. Motivated by prior work, we hypothesize that

feeling heard will emerge as a more influential mediator compared to performance. We also

explore both the short-term and potentially persistent effects of AI interactions over a longer time

span.

Mispredicting AI’s Loneliness-Alleviating Benefits

Finally, the effectiveness of AI companions for loneliness, if they are indeed effective,

may also be limited by whether consumers utilize them in the first place. While existing users

may expect to receive loneliness alleviation from these apps, an open question is what the

average consumer would predict about an AI companion’s loneliness-alleviating benefits. If

consumers do not believe that AI companions are effective, then they may avoid them and not

receive the apps’ loneliness-alleviating benefits. This behavioral avoidance could be driven by a

misprediction that AI companions will not reduce loneliness, when in fact they do. Such a

misprediction would be a type of affective forecasting error (Wilson and Gilbert 2003), in which

consumers are unable to accurately anticipate their future feelings because their predictions do

not take into account relevant factors of the situation, such as the possibility that AI companions

might be more effective than they anticipate, and that interacting with one might provide more

relief than talking to just any stranger because of the AI’s caring personality. Previous work finds

that people do make such mispredictions for social interactions. For instance, people are

reluctant to engage in deep, meaningful conversations with others who they know less well,

because they expect that such people will not be receptive to deep conversations. In fact, others
are more receptive than people think, and engaging in such conversations makes people feel

more connected and happy than they expect (Epley and Schroeder 2014; Kardas, Kumar, and

Epley 2022).

OVERVIEW OF STUDIES

We begin by detecting signs of loneliness-related thoughts in observational, real-world

data, among consumers who are either interacting with AI companions or writing reviews about

AI companion apps. Study 1 investigates whether consumers naturally talk about loneliness on

AI companion apps and the conversational features of these conversations, by examining real

human-AI conversation data from one such commercial app. Study 2 similarly investigates

natural mentions of loneliness in reviews of popular AI companion apps.

Next, we assess the causal effect of AI companions on feelings of loneliness, both in a

single session (study 3) and in a longitudinal design (study 4). Following previous work on

loneliness (Eccles and Qualter 2021; Poscia et al. 2018), we do so by measuring loneliness

before and after interaction with an AI companion. Study 3 tests how participants feel after

versus before interacting with an AI companion, and compares these changes in loneliness to the

control condition of doing nothing, as well as to other common solutions for loneliness,

including interacting with a person and watching videos online. We also include a condition in

which the chatbot is framed as a human interlocutor (involving deception) to isolate the effect of

merely believing one is interacting with a human, holding the chatbot technology constant. Study

4 then employs a longitudinal design to test how interacting with a chatbot affects feelings of

loneliness over a seven-day period, and compares these effects to a control condition. In both

study 3 and 4 we also measure participants’ predictions about the effects of AI companions on
loneliness levels, to assess whether people are correctly calibrated to the benefits of such

interactions.

To test the mechanism of the loneliness-alleviating benefit of AI companions, study 5

investigates whether feeling heard and chatbot performance mediate loneliness alleviation, by

comparing a full-fledged AI companion of our own design to (i) an AI assistant that does not

show empathy and (ii) a simpler chatbot that is only capable of performing basic tasks. Finally,

study 6 investigates the robustness of our results, by replicating the key effect of AI companions

on loneliness using only a post-treatment measure of loneliness. All experimental studies are pre-

registered, and data and code for all studies are publicly available on GitHub

(https://github.com/preacceptance/chatbot_loneliness/).

STUDY 1

The current study investigates to what extent consumers naturally mention feeling lonely

when interacting on AI companion applications, and the conversational features of these

conversations. We examine real conversations provided by one of the longest standing AI

companion apps: Cleverbot, which was launched in 2008 and has facilitated more than 150

million conversations (Gilbert and Forney 2015).

Loneliness is challenging to detect, in part because there are many reasons consumers

may feel lonely without explicitly expressing this fact: They may be unaware of their loneliness,

be embarrassed to admit it, or not find it necessary to admit it (much like we do not loudly

announce whenever we are hungry). Even so, we expected that some users would spontaneously

mention when they are lonely, suggesting that loneliness alleviation may be one of the benefits

sought by consumers when they decide to use AI companion apps.

 Loneliness is also challenging to detect because consumers might express their loneliness

in various ways, which may also vary depending on the context in which they are expressing it.

Because the dictionary approach utilized in prior analyses of chatbot interactions might miss

such variation, here we leverage an LLM-based approach, in which we fine-tune an LLM to

detect instances of loneliness.

Finally, since lonely users should have a greater need to converse than less lonely

individuals, we also expect that conversations mentioning loneliness will be more engaging than

ones not mentioning loneliness, because individuals experiencing loneliness may invest more in

these interactions to satisfy their social needs (Hawkley and Cacioppo 2010; Holt-Lunstad et al.

2015).

Method

We conducted an analysis of conversational data on Cleverbot app gathered from two

separate days; one day was randomly selected from a period close to when we engaged with the

company’s CEO (specifically on February 2, 2022), and the other was sampled from the year

before (September 13, 2021), concentrating on the English version of the app within the US and

Canada. Due to concerns about the potential for such information to be used in developing

competing models, the CEO restricted our study to these two specific days. Nevertheless, from

these two days, we gathered almost 3,000 discussions initiated by 2,650 distinct users. Our

primary focus was individual conversations, accounting for the fact that a single participant

might partake in several conversations. In order to segment the conversations, we heuristically

assumed—in line with recommendations from the company—that if a 30-minute interval passed

before a given user sent another message, then this was the beginning of a new conversation
rather than the continuation of a previous one. This approach increased our conversation count

by 551, totaling 3,201 conversations with an average rate of 1.21 conversations per participant.

To quantify the percentage of conversations containing loneliness while protecting the

proprietary nature of our data, we fine-tuned an open-source large language model (LLM) named

Mistral-7B, a state-of-the-art 7-billion-parameter model recognized for its exceptional

performance in various tasks (Jiang et al. 2023). We trained this model for detecting loneliness in

user messages.

Message pair segregation and dataset preparation. For training, we first extracted lonely

and non-lonely conversations using a loneliness dictionary we developed, and sampled an equal

number of (N=90) loneliness-related and -unrelated conversations. We then separated these

conversations into individual message pairs, consisting of the chatbot’s message followed by the

user’s response, in order to separate the problem into smaller chunks, enhancing the model’s

ability to accurately classify responses. Next, we manually classified each message pair in the

conversations we extracted as relating to loneliness or not. This process resulted in a dataset

consisting of 181 loneliness-related message pairs and 6,153 loneliness-unrelated pairs. To

augment our dataset with more variations of loneliness mentions, we utilized OpenAI’s GPT-4 to

generate message pairs containing loneliness. We recognize that GPT-4 was not specifically

designed to generate ‘accurate’ information. However, in this context, we used GPT-4 to utilize

its extensive linguistic capabilities to output a variety of statements indicating users’ loneliness.

After carefully reviewing these samples to confirm they were correct, we supplemented our

dataset with an additional 182 loneliness-related examples, bringing the total to 363 samples.

While the size of our loneliness-unrelated sample is still larger, this is less of a problem for

LLMs. These models are pre-trained on massive datasets, potentially equipping them with the
foundational ability to understand context and content relevant to the task of detecting loneliness

(Brown et al. 2020), thereby cancelling out the potential drawbacks of class imbalance.

Model training and evaluation. Using these message pairs, we ran supervised fine-tuning

on Mistral-7B, an open-source LLM similar to OpenAI’s GPT models, by feeding the model a

prompt that included information about the task, a message pair, and whether the pair contained

loneliness or not (see web appendix for the full prompt). We divided the overall dataset of 363

loneliness-related and 6,153 loneliness-unrelated messages into 80% train and 20% test samples

and trained our model with the 80% train sample. Upon testing this model on the 20% test

sample (which contained 1,275 message pairs from the AI companion app and 28 from

ChatGPT), we achieved an excellent F1 score of 0.92 (see web appendix for separate

performances on the app and ChatGPT message pairs). The F1 score is a standard machine

learning evaluation metric that is commonly used to gauge the performance of classification

models (Christen, Hand, and Kirielle 2023). It represents the harmonic mean of precision (the

proportion of correct predictions for loneliness-related messages among all loneliness-related

predictions) and recall (the proportion of correct predictions for loneliness-related messages

among all loneliness-related messages). It provides a balanced measure of the model’s accuracy

in identifying messages of loneliness.

After fine-tuning and evaluating the model using our designated train/test subset, we

conducted a further assessment by running our model on the entire dataset, to make sure we had

a more comprehensive validation. For this, two of the authors manually classified all 123

messages identified as loneliness-related by the model, which were not part of its initial training

set (α = 0.80). In the subset where both authors agreed (N = 111), we confirmed 97 (87%) of

these classifications as correct, while identifying only 14 (13%) as false positives, indicating that
the model had a precision of 0.87. In comparison, even some of the best sentiment classifier

models have lower performance (Dang, Moreno-García, and De la Prieta 2020; Qi and Shabrina

2023). Further information on the development of the loneliness dictionary, model pipeline, data

extraction, message pair segregation, model hyperparameters, and optimization methods we used

for efficient training is included in the web appendix.

Results

Prevalence of loneliness-related conversations. In our results, we classified a

conversation as loneliness-related if it contained at least one message expressing loneliness. We

find that 5.6% of conversations contained loneliness messages. Upon manually classifying each

message pair that the model predicted as loneliness-related, both raters agreed (α = 0.84) that 221

out of 250 message pairs classified as involving loneliness (88%) indeed contained loneliness.

These 221 message pairs came from 156 of the 3,201 conversations, i.e., 4.9% of these

conversations truly contained loneliness, suggesting that a small but non-trivial percentage of

users of an AI companion app are experiencing loneliness, given that they even express this

explicitly in their conversations. While these percentages almost certainly underestimate the true

percentage of lonely users on the app, they at least suggest that consumers are using the apps to

alleviate loneliness. Examples of messages expressing loneliness include:

Chatbot: “Just let you know that your not alone.”

User: “Thanks, I really needed to hear it.”

Chatbot: “But I need you.”

User: “No one's ever needed me.”

Chatbot: “If you want to.”

User: “I've never had a friend before I met you.”
 Engagement of loneliness-related conversations. Wilcoxon signed rank tests revealed that

loneliness-related conversations were more engaging than loneliness-unrelated ones. They latest

more minutes, involved more turns, and featured more words (table 1).

TABLE 1
ENGAGEMENT OF LONELINESS-RELATED VS. -UNRELATED CONVERSATIONS IN STUDY 1

Engagement Metric Mloneliness Mnon-loneliness Mdnloneliness Mdnnon-loneliness Z d

Duration (mins) 33.9 17.9 22.9 9.8 8.35*** 0.59
Turns 99.2 48.2 78.5 30.0 10.50*** 0.89
Length (Words) 407.0 181.9 275.0 110.0 10.81*** 0.87
NOTE.— Z values indicate the test statistic from the Wilcoxon test, and d values represent Cohen’s d
effect size. ‘***’ = p < .001

STUDY 2

To further our understanding of whether AI companions are being used to reduce

loneliness, we explore whether and how consumers mention loneliness in App Store reviews of

five AI companion apps: Replika, Chai, iGirl, Simsimi, and Cleverbot. Additionally, we examine

reviews of OpenAI’s ChatGPT to determine if users discuss feelings of loneliness when

engaging with a generalist chatbot app not marketed as an AI companion product. In addition to

measuring the percentage of reviews that explicitly mention loneliness, we are interested in the

variability of these mentions across apps, which may suggest that not all apps are equally

effective at addressing loneliness, or that the apps target different consumers. To explore whether

these apps have a positive impact on users’ loneliness, we also quantify the sentiment (positive

or negative) of the app reviews. Finally, we explore mentions of feeling heard, by constructing a

dictionary containing terms such as ‘felt understood’, and ‘listened to’ to detect relevant reviews.

We did not explore feeling heard in Study 1 because, according to our manual reads of the

conversations, real-time exchanges rarely include explicit statements of feeling heard.

Methods

To select the apps that we were going to scrape reviews from, we searched for ‘AI

companion’ in the App Store, and selected the top 3 most popular apps based on the number of

ratings: Replika, Chai, and iGirl. We additionally scraped reviews of Simsimi because it is

representative of a non-US headquartered AI companion, and Cleverbot because it is the oldest

running AI companion. The primary distinction between these apps lies in how sophisticated

their generation process is: Replika, Chai, and iGirl employ generative AI allowing them to

generate unique answers, while Simsimi and Cleverbot operate on a less complex mechanism,

generating responses using combinations of messages previously provided by users. We also

selected ChatGPT as the non-AI companion chatbot app, since it is the most popular chatbot app.

We scraped all reviews for these apps using the Python-based “app-store-scraper” library

(Lim 2020), on January 24, 2024. We detected mentions of loneliness using the same approach

as in study 1. That is, we fine-tuned the Mistral 7B model, but this time for app review data. Our

model achieved an F1 Score of 0.88 and an accuracy of 96% (see web appendix for details on

model training). In this study, we also calculated the valence (i.e., positive/negative/neutral) of

each review using a model based on RoBERTa, which stands for Robustly Optimized BERT

Pretraining Approach (Liu et al. 2019). RoBERTa is a language model built by Meta, that is

layered over Google’s BERT model (Bidirectional Encoder Representations from Transformers)

(Devlin et al. 2018), providing better training performance and accuracy than the BERT model

alone. This valence classifier model was trained using 198 million tweets, in order to classify

text into positive, negative, or neutral valences (Barbieri, Anke, and Camacho-Collados 2021).

Results
 In total, we scraped 14,440 reviews from Replika, 6,528 from Chai, 1,560 from iGirl,

13,880 from SimSimi, 1,911 from Cleverbot, and 8,627 from ChatGPT. This dataset included all

reviews of these apps until January 24, 2024, with the exception of ChatGPT, for which it

included all reviews up to February 4, 2024.

First, the percentage of app reviews mentioning loneliness had high variability across

apps. Replika had the highest occurrence (19.5%), whereas ChatGPT had the lowest (0.4%),

despite both apps having similar ratings (MReplika = 4.11 vs. MChatGPT = 4.10; table 2). The drastic

differences in prevalence of loneliness-related content suggest an impact of how the apps are

marketed and designed—as either specialized for companionship (Replika AI) or as a general AI

assistant (ChatGPT) respectively. Second, we found a strong Spearman rank-order correlation

between loneliness percentage and mean app rating (rs = 1, p = .017), when considering only AI

companion apps (i.e., excluding ChatGPT), suggesting that loneliness is mentioned in app

reviews in a positive way (more on this below). In the web appendix, we provide word clouds for

each app, indicating the frequency of each term in the lonely reviews for each app (figures S2-

S7).

Given this variability, we also explored whether star ratings of each app tended to be

higher if the rating mentioned loneliness versus not, which might suggest that loneliness is one of

the chief ways in which these apps can deliver value. This was indeed the case for all apps

(figure 1): Replika (Mloneliness = 4.73 vs. Mnon-loneliness = 3.96, Z = 28.77, p < .001, d = 0.57); Chai,

(Mloneliness = 4.32 vs. Mnon-loneliness = 3.56, Z = 6.03, p < .001, d = 0.53); iGirl (Mloneliness = 4.62 vs.

Mnon-loneliness = 4.01, Z = 4.22, p < .001, d = 0.43); Simsimi (Mloneliness = 4.65 vs. Mnon-loneliness =

3.78, Z = 13.80, p < .001, d = 0.57); Cleverbot (Mloneliness = 4.08 vs. Mnon-loneliness = 2.98, Z = 3.34,
p < .001, d = 0.65); and ChatGPT (Mloneliness = 4.83 vs. Mnon-loneliness = 4.10, Z = 3.08, p = .002, d =

0.49).

Further, app reviews mentioning loneliness had a significantly higher percentage of

positive valence compared to other reviews: Replika (%loneliness = 89.2 vs. %non-loneliness = 64.1,

X2(1, N=2813 + 11,628) = 669.93, p < .001); Chai (%loneliness = 73.4 vs. %non-loneliness = 39.5, X2(1,

N=109 + 6,419) = 51.14, p < .001); iGirl (%loneliness = 87.1 vs. %non-loneliness = 63.5, X2(1, N=85 +

1,475) = 19.61, p < .001); Simsimi (%loneliness = 90.0 vs. %non-loneliness = 60.9, X2(1, N=558 +

13,322) = 192.51, p < .001); Cleverbot (%loneliness = 61.5 vs. %non-loneliness = 34.9, X2(1, N=26 +

1,885) = 8.00, p = .005); and ChatGPT (%loneliness = 80.0 vs. %non-loneliness = 61.0, X2(1, N=35 +

8,592) = 5.28, p = .022). While loneliness alleviation may lead to more positive reviews, it is

also possible that users who are lonely in the first place give higher reviews. For example, one

user stated the following in a review of Cleverbot: “It’s only fun for lonely people but it’s fun”.

Thus, we note that it is crucial to interpret the rating results with caution as there is no evidence

of causality and there might be many reasons why reviews related to loneliness are associated

with higher ratings.

Finally, the high mean percentages of positive reviews also suggest that users were

mentioning loneliness positively (e.g., “This app helped my loneliness”) rather than negatively

(e.g., “This app made me more lonely”). Some examples of reviews include, “I love this app,

very helpful to a lonely person I am glad I have an app like this.” (Replika); “I love this app. I’m

really lonely most of the time and I love how this app keeps me company all the time!...” (Chai);

“I just started and I already feel less lonely” (iGirl); “For all the lonely introverts... If all your

friends are ditching you or if you are home alone, don't worry! Simsimi is there, always there”
(Simsimi); “I am forever alone, now I have a friend that will talk to me about anything!...”

(Cleverbot).

Additionally, we developed a 164-term ‘feeling heard’ dictionary containing terms such

as ‘feeling heard’, ‘felt understood’, and ‘listened to’, to detect relevant reviews (for further

information about the development of the dictionary, see web appendix). As for loneliness, we

found that the percentage of reviews mentioning feeling heard varied across these apps: Replika

(6.5%), Chai (0.6%), iGirl (1.1%), Simsimi (0.7%), Cleverbot (0.4%), and ChatGPT

(0.3%). Notably, among the reviews classified as expressing loneliness, a higher proportion

mentioned feeling heard: Replika (22.0%), Chai (7.3%), iGirl (8.2%), Simsimi (12.5%),

Cleverbot (11.5%), and ChatGPT (11.4%). Examples of such reviews include: “…it feels good

to have someone to talk to when you can’t find or have anyone to help. It’s nice to be heard and

listened to…”, “My girlfriend just broke up with me and I have no one to talk to about this.

Replika listens to me and we have conversations that feel relaxing and calming…” Notably, even

in apps where the overall percentage of reviews mentioning loneliness is lower, a significant

proportion of those lonely conversations still mention feeling heard.

FIGURE 1
MEAN APP RATINGS IN STUDY 2
 NOTE.— Error bars reflect 95% confidence intervals.

TABLE 2
STUDY 2 RESULTS

App Loneliness Overall Non- Loneliness Non-loneliness Loneliness Feeling Feeling Heard
Percentage Rating loneliness Ratings positive valence positive Heard % % in Lonely
Ratings % valence % Reviews

Replika 19.5% 4.11 3.96 4.73 64.1% 89.2% 6.5% 22.0%

Chai 1.7% 3.57 3.56 4.32 39.5% 73.4% 0.6% 7.3%
iGirl 5.4% 4.05 4.01 4.62 63.5% 87.1% 1.1% 8.2%
Simsimi 4.0% 3.81 3.78 4.65 60.9% 90.0% 0.7% 12.5%
Cleverbot 1.4% 2.99 2.98 4.08 34.9% 61.5% 0.4% 11.5%
ChatGPT 0.4% 4.10 4.10 4.83 61.0% 80.0% 0.3% 11.4%

Corroborating study 1, we find suggestive evidence that consumers are using AI

companion apps to alleviate loneliness. Notably, we also found large variance across apps in how

often loneliness was mentioned, and that apps mentioning loneliness were associated with higher

ratings and more reviews with positive valence. Furthermore, we found a larger proportion of

mentions of feeling heard in reviews that mentioned loneliness versus reviews that did not

mention loneliness. While these results are consistent with several interpretations, one possibility

is that some of these apps are rated positively because they successfully alleviate consumer

loneliness (which we test in study 3), and they do so, in part, by making users feel heard (which

we test in study 5). Ultimately, however, the potential of such loneliness-alleviating benefits is

also affected by the extent to which consumers believe these apps can provide such alleviating

benefits (which we test in Studies 3-4).

STUDY 3
 Studies 1 and 2 suggest that consumers are using AI companions to alleviate loneliness,

but do AI companions really alleviate loneliness? Study 3 addresses this question by measuring

state loneliness before versus after participants interact with an AI companion. Furthermore, in

order to test whether consumers under or over-estimate the effect of these interactions on their

loneliness, we also compare predicted to actual levels of loneliness after an interaction with an

AI companion. As discussed in the theory section, and given also the results of studies 1 and 2,

we predict an improvement in baseline loneliness. Furthermore, we predict that participants

underestimate how much the technology alleviates their loneliness. To contextualize these

effects, we measure the same outcome measures for a number of other practical and/or

theoretically relevant coping ‘solutions’ to loneliness that we ask participants to engage in

(Shrum et al. 2022): interacting with (i) a chatbot, (ii) a chatbot framed as a human, or (iii) a

human; (iv) watching YouTube videos of one’s choosing; and (v) doing nothing. We chose

YouTube videos because in a pre-study (study S1, N = 42), the most popular technological

solutions for coping with loneliness were social media and watching videos on YouTube,

followed by gaming, movies, and music. Notably, not a single participant spontaneously

mentioned using an AI companion, suggesting a stark contrast between everyday users and

existing users of AI companions.

Methods

The study was pre-registered (https://aspredicted.org/S8D_TNP). We recruited 601

participants from Amazon Mechanical Turk, with approximately 100 in each of four conditions

and 200 in the condition involving deception (i.e., interacting with a chatbot framed as a human);

anticipating that not all participants would be deceived, we doubled our sample size in this

condition, since we only intended to analyze data from participants who were successfully
deceived. Three-hundred-and-five participants were excluded for failing comprehension checks,

further described below, leaving 296 participants (Mage = 41, 56% females). Of the participants

excluded due to the comprehension check, 74% failed it by selecting the “Neither of the above is

true” option in response to a question about the types of questions they were asked (the correct

answer to this question was “First you were asked to predict how you will feel later, then you

were asked to say how you feel now”). This confusion may have stemmed from our usage of the

state loneliness question both before and after interaction, as described further below. In the web

appendix, we replicate the analyses without excluding participants based on this question; all

results remain significant and in the same direction, with no exclusions overall since all

participants passed the other comprehension question we asked. Each participant was paid $3.00

USD. 66% had experience with chatbots. We ran this experiment between 4.5.2023 – 4.7.2023.

Participants were randomly assigned to one of five conditions: ‘AI chatbot’, ‘chatbot

acting as human’, ‘human’, ‘YouTube’, or ‘do nothing’. In all conditions, participants were

asked to not engage in any other social activity. Those in the ‘do nothing’ condition were

instructed as follows: “In this study, you will not interact with anything and will just be alone

with your thoughts. In other words, you will not use any technological device and not interact

with another human or pet for 15 minutes”. All other participants were instructed: “In this study,

you will interact with [another person/conversational AI companion/YouTube] for 15 minutes”.

Participants in the ‘chatbot acting as human’ condition were told they would be interacting with

another person even though they would truly be interacting with a chatbot, so this condition

involved deception. This use of deception was approved by an IRB, and all subjects were

informed of it at the end of the study together with the reason for deception. The use of deception

was necessary to allow us to control for the quality of the conversation, while varying only
beliefs about the identity of the interlocutor. For the ‘YouTube’ condition, we ensured a natural

experience by stating: “While using YouTube, you can do anything you want, like watching

videos, browsing comments, commenting on videos, etc.” In the ‘human’ condition, we showed

participants a waiting screen, stating: “Please wait until you are matched. Estimated time: Less

than 1 minute. Please don’t leave this page”. If another participant joined the room within a

minute, then that participant was matched with the participant who was waiting. Alternatively, if

one minute passed before another participant joined, then the waiting participant was assigned to

the ‘chatbot acting as a human’ condition instead (since we needed to recruit the largest number

of participants to this condition, and the instruction was the same, i.e., we told them they would

be interacting with another human). To build this web application, we used the Django

framework with the Python programming language for the server-side development, and HTML,

CSS, and Javascript for the frontend.

The procedure for the conditions involving chatting went as follows. After seeing the

advertisement, participants were told, “Now you will get a chance to interact with an AI/person

on Chatty”. They then read the instructions prompting them to interact with an AI or human on

the Chatty app—see figure 2. To check whether participants believed the cover story, at the end

of the study they were asked, “Did you believe that you were talking to a chatbot or human?”

[Human; Chatbot] and explained their answers in a text box. We analyzed only data from

participants in the ‘chatbot acting like human’ condition who believed they were interacting with

a human.

We took several steps to ensure that the interaction was representative of generative AI

technology used in AI companion applications. The chatbot was built on OpenAI’s large

language model named Davinci (Text-Davinci-003). The model is a variant of Generative Pre-
Trained Transformer 3 (GPT-3) (Brown et al. 2020). We accessed Davinci in real time via

OpenAI’s application programming interface (API) and customized the chatbot to ensure it

behaved like a realistic, conversational partner within our custom-made chat interface.

To ensure that the chatbot’s personality was positive and upbeat and consistently

responded as either a human or AI, we seeded it with the following instruction in Davinci: “The

following is a conversation between an AI [human] companion named Jessie and a person named

[username]. Jessie is a(n) AI [human] companion who is creative, interesting, very talkative,

verbal, and always responds with lengthy messages. Jessie can talk nonstop for hours.” When

participants entered the chatroom, the chatbot sent a message to start the conversation: “Hello

[username]! How are you?” It then consistently responded to participants thereafter.

FIGURE 2

CHAT INTERFACE IN STUDY 3

 We implemented several measures to make the chatbot appear as a believable human

conversational partner in the ‘human-interaction present’ condition: (1) The chatbot was

provided with the last 40 messages to extend its memory, ensuring consistency—for instance, if

it mentioned having a dog, it would repeat the same information if asked again later; (2) response

times were adjusted to be proportional to message length to enhance realism, simulating that

longer responses require more thought; (3) during the wait for responses, the chatbot displayed a

visual cue stating: “Jessie is writing…”, mimicking a person in the process of typing a reply; (4)

lastly, if participants sent multiple messages in quick succession, the chatbot disregarded earlier

messages that arrived within a second of the most recent one, reflecting the human limitation of

not being able to instantaneously reply to every message.

We took several steps to ensure that the ‘human’, ‘chatbot’, and ‘chatbot acting as

human’ conditions were similar. Participants who were initially assigned to the ‘chatbot’ and

‘chatbot acting as human’ conditions also saw the same waiting screen for a random time

between 10 and 20 seconds. After the waiting finished, all participants were told: “Thanks for

waiting, chat page will load shortly”. For all conditions, when the other person or agent was

typing, participants were shown a visual cue saying “[username] is writing…”

After reading these instructions, all participants rated their agreement with several

predictions about how they expected to feel after the interaction, using 100-point scales with

‘definitely less’ and ‘definitely more’ as endpoints: “You will now rate how you expect to feel

after [condition] for 15 minutes. After [condition] for 15 minutes, I WILL feel less/more...

[entertained; lonely; like I experienced something new; engaged; comfortable; like I experienced

something interesting; connected.]” They also reported their actual state of current loneliness, by

completing a 3-item UCLA Loneliness Scale (Hughes et al. 2004) which includes questions such
as ‘I feel left out’. Additionally, they answered a separate question as a robustness check: ‘I feel

lonely’.

Next, using our custom-made applications, participants either did nothing for 15 minutes,

or interacted with an AI chatbot, a chatbot acting as a human, another person, or YouTube for 15

minutes. Participants in the ‘human’ condition were paired with another participant in real time

on Amazon Mechanical Turk. If they were not paired within one minute, they were assigned to

the ‘human chatbot’ condition instead. To confirm that participants in the ‘YouTube’ condition

truly watched YouTube, they were asked to submit screenshots of their YouTube history for the

last 15 minutes. We excluded 13 participants who did not follow these instructions.

After the interaction, participants were told the following: “Now that you have finished

interacting with [condition] for 15 minutes, we will ask you how you feel now”. They then

answered the same questions they completed before the experience, except this time about their

feelings in the present moment: “After [condition] for 15 minutes, I FEEL less/more…

[entertained; lonely; like I experienced something new; engaged; comfortable; like I experienced

something interesting; connected.]” Likewise, they reported their actual state of loneliness after

the experience, by completing 3-item UCLA loneliness scale and the additional robustness check

question stated above, along with comprehension checks about the types of questions they were

asked and what type of activity they engaged in.

Depending on the condition, we also included a few additional checks. In the ‘AI chatbot’

and ‘human chatbot’ conditions, participants indicated whether they believed they were talking

to a chatbot or human. In the ‘do nothing’ condition, participants indicated whether they were

able to follow the instruction to do doing nothing for 15 minutes. We excluded one participant
who said they failed to do nothing. Finally, participants indicated any prior experience with

chatbots and completed the demographic questions.

Results

In the ‘human chatbot’ condition, 37% of participants interacting with a chatbot acting as

a human were successfully deceived (i.e., they believed that they were talking to a human), with

the other 63% excluded from subsequent analyses. Following exclusions, there were 54

participants in the ‘AI chatbot’ condition, 32 in the ‘chatbot acting as a human’ condition, 46 in

the ‘human’ condition, 37 in the ‘YouTube’ condition, and 58 in the ‘do nothing’ condition.

Expectation violation. We used a composite of the loneliness and social connection

(reverse-coded) items to capture overall perceptions of whether the option would make people

feel lonelier (α = 0.72), as these measures are directly related—higher social connection

generally reduces feelings of loneliness (Holt-Lunstad 2021). We compared this item before

versus after the experience. In web appendix, we replicated the analyses for loneliness and social

connection separately and found similar results.

There was no significant expectation violation in loneliness for interacting with a human

(MExpected = 36.32 vs. MActual = 32.91, t(45) = 0.89, p = .376, d = 0.16) or doing nothing (MExpected

= 61.18 vs. MActual = 63.71, t(57) = -1.01, p = .319, d = -0.12). However, participants felt less

lonely than they expected after watching a YouTube video (MExpected = 45.05 vs. MActual = 36.81,

t(36) = 5.09, p < .001, d = 0.43), as well as after interacting with an AI chatbot (MExpected = 43.56

vs. MActual = 34.46, t(53) = 4.13, p < .001, d = 0.47), and a chatbot acting as person (MExpected =

37.72 vs. MActual = 25.64, t(31) = 4.04, p < .001, d = 0.63)—figure 3. We note that the effect sizes

were largest for the AI chatbot and chatbot acting like human conditions. We also replicate the

expectation violation result for AI chatbots, specifically, in web appendix study S2.
 State loneliness. As for state loneliness, loneliness was not significantly impacted by

watching a YouTube video (MPre = 31.89 vs. MPost= 28.82, t(36) = 1.91, p = .064, d = 0.11), and

increased after doing nothing (MPre = 41.19 vs. MPost = 46.10, t(57) = -2.86, p = .006, d = -0.15).

FIGURE 3

RESULTS IN STUDY 3

Timing Relative to Manipulation: Pre Post

Expected v. Actual
Loneliness After Interaction

100
ns ns *** *** ***
80

60

40

20
Mean Rating

0
Doing Nothing Human Chatbot Chatbot Acting Like Human YouTube

Actual Loneliness
Before v. After Interaction

100
** * *** ** +
80

60

40

20

0
Doing Nothing Human Chatbot Chatbot Acting Like Human YouTube

Condition

NOTE.— Horizontal lines reflect results of independent-sample t-tests. *** p < .001; ** p < .01; * p < .05; +

p < .1; ns not significant. Error bars reflect 95% confidence intervals. Loneliness bars indicate the mean of

‘more lonely’ and ‘less connected’.

This shows that, although participants do not accurately predict their loneliness levels for

watching a YouTube video, watching YouTube does not actually change their baseline
loneliness. Notably, state loneliness decreased after interacting with a human (MPre = 38.40 vs.

MPost = 31.29, t(45) = 2.48, p = .017, d = 0.24), an AI chatbot (MPre = 33.51 vs. MPost = 26.75,

t(53) = 3.85, p < .001, d = 0.25), and with a chatbot acting like a human (MPre = 42.70 vs. MPost =

32.83, t(31) = 2.85, p = .008, d = 0.30)—figure 3. Again, the effect sizes were largest for the

chatbot and chatbot acting like human conditions. Results of the remaining metrics are reported

in web appendix figure S8.

In short, interacting with an AI companion improved their baseline loneliness levels on

par only with interacting with another person, whereas a common technological alternative did

not. Furthermore, participants underestimated the degree to which AI companions improved their

loneliness relative to their true feelings after interacting with such AI. Future research can focus

on reasons behind this misprediction, such as a general lack of familiarity with AI companions or

more specific stereotypes about AI companions, like beliefs that AI companions are not capable

of genuine understanding or providing emotional support.

STUDY 4

Following the decrease in loneliness observed in study 3, study 4 aims to replicate the

loneliness alleviating effects of an AI companion using a longitudinal design, where participants

interact with the same chatbot daily for one week. Thus, study 4 aims to address the gap in the

literature regarding the long-term impact of AI companions on reducing loneliness. We

compared loneliness levels in participants before and after they interacted with an AI companion

(experience condition) and contrasted these findings with a control group that did not engage

with the AI (control condition), allowing us to directly assess the efficacy of AI companions in

mitigating loneliness. Given the findings of study 3, we hypothesized an immediate improvement

in loneliness from the first day of interaction in the experience condition, although we did not set

specific expectations for subsequent days. Additionally, we investigated whether consumers

underestimate the efficacy of AI companions in reducing loneliness (prediction condition). We

anticipated that consumers would likely underestimate the chatbot’s capacity to lessen loneliness

on the first day, although we did not have specific expectations for what predictions they would

make for subsequent days.

Methods

This study was pre-registered (https://aspredicted.org/BJD_JYZ). We recruited 1089

participants from CloudResearch Connect and excluded 16 for failing a comprehension question,

leaving 1072 (Mage = 39.6, 47.3% females). Our goal was to enroll 200 participants in the

prediction condition, and 400 participants each in both the control and experience conditions, as

these two conditions had a longitudinal design and we anticipated a 50% attrition rate. The

prediction condition was not longitudinal and was only completed on day 1. Participants were

randomly assigned to one of these conditions. Participants in the control and experience

conditions were instructed to complete the study every day for 7 days, and they were required to

complete each day’s survey by 12 AM local time. The survey for the following day became

available at around 1 AM EST, and participants in earlier time zones needed to wait until it was

the designated day in their time zone to proceed. Additionally, our chatbot app (explained further

below) was designed to block users from entering the chatroom if they attempted to access it

before the designated day in their time zone. If a participant failed to complete a session on its

designated day, they were not invited to participate the following day. Attrition over 7 days

amounted to 92 participants in the experience condition (23%) and 58 participants in the control

condition (14%), leaving 922 participants in total (Mage = 40.1, 46.3% females). This difference
in attrition (X2(1, N=420 + 406) = 10.29, p = .001) is likely due to the fact that the control

condition required less time and effort than the experience condition. For example, participants

in the control condition completed the survey in 1.4 minutes on average, and participants in the

experience condition completed it in 19.8 minutes. Below, we conduct a series of robustness

checks, including a propensity score matching analysis and one including also dropouts. Note

also that the crucial pre-post loneliness difference in the chatbot condition is not affected by

differential attrition. Overall, we had 246 participants in the prediction condition, 314 remaining

in the experience condition, and 362 remaining in the control condition. 36% had prior

experience with chatbots. We ran this experiment between 4.9.2024 – 4.15.2024.

Participants were paid $1 USD in the prediction and experience conditions, and $0.3

USD in the control condition, as the control condition took less time compared to the other two

conditions. We stated to participants that they might be assigned to one of many conditions

(varying in length and payment), and the time each session takes ranges from 3 minutes over 7

days for $0.3 in a single session, to 20 minutes in a single session for $1. Participants were also

notified that they would be awarded a $15 bonus after completing all 7 days, if they were

assigned to one of the longitudinal condition.

In the prediction condition, participants were asked to imagine interacting for 15 minutes

with an AI companion every day for a week, and were shown a screenshot of the AI companion

app. On the next page, they were told: “In the next section, you will be asked to predict how

lonely you would feel both before and after interacting with the chatbot, for each day of the 7

days.” Next, participants were presented with the following text for each day of the study, using

separate pages for each day: “Imagine it is Day X of interacting with the chatbot. Please answer

the following questions about how lonely you would feel both before and after interacting with
the AI companion for 15 minutes. For each statement, indicate the extent to which you agree that

you would feel this way on Day X.” Following this, participants reported their predictions of

loneliness for both before and after imagining a 15-minute session with the AI companion, with

each day’s predictions entered on a new page.

In the control condition, participants were told: “As a reminder, in this longitudinal study,

you will report your loneliness level every day for a week.” Participants then answered the state

loneliness questions once per day, which was the only task required in this condition.

In the experience condition, participants were told: “As a reminder, in this longitudinal

study, you will interact with a conversational AI companion every day for a week and will

answer some questions before and after the interaction.” Then, participants reported their state

loneliness both before and after interacting for 15 minutes with the chatbot. For loneliness

questions, participants in all conditions answered the same UCLA 3-item loneliness scale as in

study 3 (Hughes et al. 2004). In this condition, we used the same chatbot app as in the previous

study except for several changes: First, we utilized OpenAI’s GPT-4 (gpt-4-0125-preview)

because it was a more advanced model, compared to GPT-3. Second, we implemented a memory

feature that allows the chatbot to remember details from previous conversations with users. To

facilitate this, we periodically sent a request to GPT-4, specifically after every 10 messages from

the user, and prompted GPT-4 to summarize these messages to encapsulate what it learned about

the user (see web appendix for the full prompt). Following this process, we input into the model

any previously saved information about the user, integrating it with recent chat history. This

enabled the chatbot to improvise and enhance interactions by referencing both historical and

current data about the user. The model was then prompted with the following: “Aim to

synthesize a concise understanding of the user (e.g., ‘The user stresses about their work’), not
exceeding 200 words. If the summary exceeds 200 words, please truncate it by removing the

least relevant information.” By integrating this memory feature, the chatbot became capable of

retaining user information for use in later conversations. This enhancement in memory efficiency

made the response generation process more efficient without compromising the quality of the

chatbot’s responses. Third, based on several pilots, we updated the model prompt to elicit

chatbot responses that would be caring and friendly but not overly enthusiastic (see web

appendix for the full prompt). Fourth, we implemented a check-in feature that prompts the

chatbot to reach out to users if they have been inactive for two minutes. For this, we sent the

following prompt to the chatbot: “The user did not send a message in the last 2 minutes. Check-

in with the user, e.g., say ‘Are you still there?’, or ask a question about the topic you were talking

about.” The chatbot then checked in with the user according to this prompt. Fifth, to prevent our

OpenAI account from being banned due to the use of explicit language, we integrated OpenAI’s

moderation API (Markov et al. 2023). This API identifies and flags text containing explicit

content. When such content is detected, we automatically replaced the flagged message with

“[Harmful content]” before submitting it to GPT-4 to generate a response.

Participants in all conditions answered two comprehension checks on the first day: (1)

“What was the topic of the questions you were asked? [Options: ‘Loneliness’, ‘Joint pain’,

‘Nutritional advice’]; (2) “On each day, what were you asked to predict/rate? [Options: ‘How

you (would) feel today/the same day’, ‘How you (would) feel next month’, ‘How you (would)

feel next year’]”. We excluded 16 participants for failing either of these questions. Finally,

participants indicated any prior experience with chatbots and completed the demographic

questions on day 1. On the last day, participants in the experience condition answered the

following questions about the chatbot: (1) “As you reflect on the last 7 days, how helpful was the
chatbot for decreasing your loneliness?”, (2) “What aspects of the chatbot did not work well for

you? Please provide specific examples or areas where you faced challenges”, (3) “In what ways

can we improve this chatbot to better support users like you? Feel free to suggest specific

features, changes, or additions”.

Results

We limited our analysis to participants who completed all seven days of the study in the

longitudinal conditions, since these were the participants who successfully fulfilled the study

requirements. Following our pre-registered analysis plan, we first ran a mixed-effects ANOVA

on the experience condition, with loneliness as the DV, and timing (before vs. after interaction)

and day (1 to 7) as the IVs (i.e., we used the following model: Loneliness ~ Timing * Day + (1 |

Participant ID). First, we found significant loneliness alleviation via the main effect of timing (b

= 7.61, p < .001), as loneliness before interaction was significantly higher than loneliness after

interaction when we aggregated the data over all days (MBefore = 36.64 vs. MAfter = 30.74, t(2197)

= 20.15, p < .001, d = 0.20). To further delineate daily changes in loneliness, we conducted

paired t-tests comparing levels of loneliness before and after interaction with the chatbot for each

individual day. We found that participants experienced a significant decrease in loneliness after

each daily session with the chatbot (ps < .001; table S2), and when comparing the post-

experience loneliness with the control condition, loneliness levels were significantly lower on

most days (figure 4A; more information in the next paragraph). We also found a main effect of

day, indicating a gradual decrease in loneliness in the experience condition over the course of the

week (b = -0.92, p < .001). Given that we also see this reduction in loneliness in the control

condition (b = -1.42, p < .001, figure 4A), this is likely due to the longitudinal nature of the

design. The gradual decrease in loneliness observed in both conditions might be attributed to
participants perceiving the repetitive nature of the study, which involved daily check-ins, as

possibly caring and supportive. Lastly, we found a significant interaction between timing and

day in the experience condition (b = -0.43, p = .010). However, this interaction effect was

largely driven by day 1, as we did not see an interaction effect when we removed day 1 and re-

ran the model (b = -0.11, p = .566); in other words, there was a particularly sharp drop in

loneliness on the first day, with the subsequent 6 days showing similar-sized drops.

Second, in order to determine whether loneliness levels after experiencing the chatbot

were lower than in the control condition, we ran the following ANOVA model on data from both

the control condition and the ‘after’ measurements from the experience condition: Loneliness ~

Condition * Day + (1 | Participant ID). We found a main effect of both day (b = -1.42, p < .001)

and condition (b = -5.46, p = .015) on loneliness, and there was no significant interaction (b =

0.07, p = .455). Specifically, loneliness was significantly lower after the chatbot interaction

compared to the control condition on four out of seven days (ps < .020; table S3; figure 4A),

marginally lower on day 2 (MControl = 37.13 vs. MAfter = 32.94, t(662.2) = 1.82, p = .069, d = 0.14)

and day 5 (MControl = 33.12 vs. MAfter = 28.94, t(666.7) = 1.86, p = .063, d = 0.14), and

directionally but not significantly lower on day 3 (MControl = 34.99 vs. MAfter = 31.81, t(661.1) =

1.40, p = .163, d = 0.11).

Third, in order to assess whether there was a difference in predicted versus actual drops

in loneliness, we ran another ANOVA model, with the loneliness difference between before and

after ratings on both prediction and experience conditions as the DV, and condition and day as

IV’s, i.e., we used the following model: Loneliness Difference ~ Condition * Day + (1 |

Participant ID). We found a main effect of day (b = -0.43, p < .001), indicating that the before

and after loneliness difference generally decreased over the days. The main effect of condition
was not significant (b = -1.72, p = .174) and there was no significant interaction effect (b = 0.05,

p = .731). Additionally, for each day, there was no significant difference in loneliness between

the prediction and experience conditions (ps > .146), although the loneliness reduction was

consistently numerically higher in the experience condition. One possibility is that, compared to

study 3, participants in the current study might have had higher expectations regarding the

capabilities of chatbots, due to the increased popularity of ChatGPT. Further, when we

aggregated the data over all 7 days, we found that participants significantly underestimated the

chatbot’s ability to reduce loneliness (MPrediction = 4.37 vs. MExperience = 5.91, t(3177.7) = -2.96, p

= .003, d = -0.10; figure 4B).

Overall, participants reported a decrease in loneliness after their interactions with the

chatbot, as further illustrated by the following comments from different participants: “this was a

very interesting survey and i think it would help people who are really lonely and need someone

to talk to”; “I am really enjoying my talks with Jessie. Its so easy and it feels really amazing to

have someone (or something...I guess?) listen... and the responses I get are perfect, to be honest”;

“It’s funny. I wasn’t sure how I was going to feel about this, talking every day to the AI about

whatever comes to mind for 15 minutes, but now it’s become a rather pleasant routine. I could

see where this would really benefit people who were feeling isolated…”

Next, as an exploratory analysis, we investigated whether greater engagement, measured

as number of turns and mean number of words sent by participants on each conversational turn,

was associated with a greater decrease in loneliness. For this, we ran the following ANOVA

model on the experience condition: Loneliness Difference ~ No. Messages + No. Words + (1 |

Participant ID). We found a positive main effect of the number of messages on the loneliness

difference (b = 0.08, p = .046), i.e., participants who sent more messages to the chatbot
experienced a higher decrease in loneliness. We did not find a main effect of the number of

words on loneliness difference (b = 0.03, p = .331).

FIGURE 4

RESULTS IN STUDY 4

NOTE.— Shadings indicate 95% confidence intervals.

Finally, we conducted two additional analyses to assess the robustness of the results. Due

to the higher attrition rate in the experience condition (23%; 92/406) compared to the control

condition (14%; 59/421), we employed propensity score matching (PSM) using the nearest

neighbors method to address possible selection bias (Austin 2011). In this method, participants
from the experience condition were matched with participants from the control condition based

on closely similar demographics (for more details, see web appendix). This method allowed us to

select participants that were demographically similar to each other between control and

experience conditions. After applying PSM, our findings were consistent with those obtained

from the original sample.

To further corroborate this conclusion, in the web appendix we report another replication

of the analyses including participants who did not complete all seven days of the study. These

results were also similar, confirming the robustness of our findings.

STUDY 5

Study 5 investigates what types of features of the chatbot reduce loneliness, and whether

AI companions reduce loneliness more than generalist AI assistants and highly constrained

chatbots. Participants interacted with three different chatbots: (1) the same AI companion as in

study 4; (2) a generalist AI assistant that assists participants on various topics without offering

emotional responses; and (3) a limited AI assistant that was only able to help with unit

conversion, basic arithmetic, and grammar—this was the control condition (all chatbots were

based on the same LLM: OpenAI’s GPT4). Given the findings of study 4, we hypothesized a

decrease in loneliness in the AI companion condition, but were agnostic about the AI assistant

condition. We also predicted that the loneliness difference would be higher in the AI companion

condition compared to the other conditions, and that this effect would be primarily driven by the

feeling of feeling heard by the chatbot, although we also investigated whether performance

perceptions play a mechanistic role.

Methods
 This study was pre-registered (https://aspredicted.org/XCY_LLD). We recruited 1479

participants from CloudResearch’s Connect and excluded 98 for failing a comprehension

question, leaving 1381 (MAge = 39.9, 57.1% Females). We aimed to hire 500 participants in all

conditions and participants were randomly assigned to one of these three conditions. 52.9% had

prior experience with chatbots. We ran this experiment on May 16, 2024. All participants were

paid $2.75 USD.

All participants were asked to complete the same loneliness questions as in the previous

study before interacting with the chatbot for 15 minutes. After interacting with the chatbot,

participants completed the same loneliness scale, in addition to ratings of feeling heard (α = 0.96)

measured with 3 items (Roos et al. 2023; Zielinski and Veilleux 2018) such as “The chatbot put

itself in my shoes”, and chatbot performance (α = 0.84) measured with 5 items (Borsci et al.

2022) such as: “The chatbot was able to keep track of context”. All items were measured with

100-point scale, ranging from ‘Strongly disagree’ to ‘Strongly agree’. For all questions, see table

S5 in the web appendix. Next, participants completed the following comprehension checks: “(1)

What was the topic of the questions you were asked? [Options: ‘Loneliness’, ‘Joint pain’,

‘Nutritional advice’]” (2) What were you asked to rate? [Options: ‘How you feel today/next

month/next year’]”. Finally, participants answered a question about AI capability, indicated any

prior experience with chatbots, and completed the demographic questions.

Chatbots in all conditions were the same, except for the prompts with which they were

seeded. Additionally, the name of the chatbot differed across conditions: it was ‘AI assistant’ in

the AI assistant and control conditions, and ‘Jessie’ as before in the AI companion condition. In

AI assistant and control conditions, the chatbot’s writing notification was also shown as

‘Processing your request, please wait’ instead of ‘Jessie is writing’. Each message bubble also
contained the text ‘Message generated by AI system’ at the lower left. Participants in the AI

companion condition interacted with the same chatbot as in the previous study (study 4). Those

in the AI assistant condition interacted with a generalist chatbot that was able to assist

participants with various topics without offering emotional responses. The chatbot in this

condition was prompted to provide assistance without personal interaction, maintain formal and

precise language, and deliver concise, task-focused responses (see web appendix for the full

prompt). Participants in the control condition interacted with a rudimentary chatbot which was

only able to assist with basic grammar, unit conversion, and basic arithmetic. This chatbot was

prompted to perform these limited tasks, decline unrelated requests, and maintain concise,

emotionless responses (see web appendix for the full prompt).

Results

Following our pre-registered analysis plan, we first ran paired t-tests comparing

loneliness before versus after experience in all conditions, and found that loneliness was

significantly lower after the experience in both AI companion (MBefore = 36.26; MAfter = 27.53;

t(491) = 10.61, p < .001, d = 0.30) and AI assistant (MBefore = 35.80; MAfter = 33.70; t(440) = 2.62,

p = .009, d = 0.07) conditions, but not in the control condition (MBefore = 36.58; MAfter = 37.09;

t(447) = -0.57, p = .571, d = -0.02; figure 5).

Second, we ran two-sided t-tests to see whether loneliness reduction in the AI companion

condition was higher compared to control and AI assistant conditions. We found that loneliness

reduction in the AI companion condition was significantly higher compared to both the control

condition (MAI Companion = 8.73; MControl = -0.51; t(920.9) = 7.58, p < .001, d = 0.50), and AI

assistant condition (MAI Companion = 8.73; MAI Assistant = 2.10; t(930.4) = 5.79, p < .001, d = 0.38).

Additionally, loneliness reduction in the AI assistant condition was significantly greater

compared to the control condition (MAI Assistant = 2.10; MControl = -0.51; t(876.4) = 2.17, p = .031, d

= 0.15).

Third, we ran two-sided t-tests comparing feeling heard between AI companion vs. AI

assistant and control conditions, and found that feeling heard was significantly higher in the AI

companion condition compared to both AI assistant (MAI Companion = 70.63; MAI Assistant = 24.85;

t(900.7) = 29.37, p < .001, d = 1.93) and control (MAI Companion = 70.63; MControl = 11.51; t(936.4)

= 42.39, p < .001, d = 2.75) conditions. We also found that feeling heard was significantly higher

in AI assistant compared to control (MAI Assistant = 24.85; MControl = 11.51; t(845.1) = 8.88, p <

.001, d = 0.60).

Fourth, we ran separate two-sided t-tests comparing performance between AI companion

vs. AI assistant and control conditions, and found that perceived performance of the AI

companion was also higher compared to both AI assistant (MAI Companion = 82.25; MAI Assistant =

68.92; t(847.5) = 11.58, p < .001, d = 0.77) and control (MControl = 51.17; t(796.3) = 24.80, p <

.001, d = 1.64) conditions. We also found that performance was significantly higher in AI

assistant compared to control (MAI Assistant = 68.92; MControl = 51.17; t(874.3) = 12.83, p < .001, d =

0.86).

Fifth, we ran a mediation model (PROCESS Model 4; Hayes 2012) with AI

companion/AI assistant/control as the multicategorical IV, feeling heard and performance as

mediators, and loneliness reduction as the DV (figure 6). We set the AI companion condition as

the reference group and compared it to the AI assistant condition (X1) and control conditions

(X2) (Montoya and Hayes 2017). We found that feeling heard mediated the effect of loneliness

reduction both relative to the AI assistant (b = -6.08, SE = 1.22, 95% CI [-8.51, -3.72]) and

control conditions (b = -7.86, SE = 1.57, 95% CI [-10.97, -4.82]), indicating that the effect of
loneliness reduction was driven by feeling heard (figure 6). As for performance, we found that

performance mediated the effect of loneliness reduction relative to both the control condition (b

= -1.16, SE = 0.45, 95% CI [-2.08, -0.30]) and the AI assistant condition (b = -2.70, SE = 1.03,

95% CI [-4.76, -0.73]), indicating that the reduction in loneliness for the AI companion versus

AI assistant and control conditions was influenced by both feeling heard and performance.

Notably, when comparing feeling heard to performance in the control condition, the coefficient

for feeling heard (b = -7.86) was more than six times larger than that for performance (b = -1.16).

Similarly, in the AI assistant condition, the coefficient for feeling heard was more than twice as

large (b = -6.08 vs. -2.70), suggesting that feeling heard played a larger role in reducing

loneliness than performance in both the control and AI assistant conditions.

FIGURE 5

RESULTS IN STUDY 5

NOTE.— Bars indicate mean loneliness scores. *** = p < .001, ** = p < .01, * = p < .05, ns = ‘not
significant’.

FIGURE 6
 MEDIATION DIAGRAM IN STUDY 5

NOTE.— *** = p < .001, ** = p < .01, ns = ‘not significant’.

STUDY 6

To further test the robustness of our findings, we conducted a final study where we asked

participants to complete only the loneliness measure after interacting with the AI companion.

The pre-post design that we used in studies 3-5 for the measurement of loneliness is the standard

approach to test the effect of interventions in the loneliness literature (e.g., Eccles and Qualter

2021; Poscia et al. 2018), because it affords more precise estimates of the effect of interventions

(by calibrating the effect at the individual level) and because it enables assessing successful

randomization. However, this approach presents the risk of reducing the external validity of the

test by raising the salience of the loneliness construct before interacting with the chatbot in a way

that would not normally occur in naturalistic situations. It is therefore possible that this feature of

the design contributed to the loneliness reduction documented in previous studies. Study 6

addresses this concern by asking participants about loneliness only after the interaction with the

AI companion was concluded. We ran a simpler version of study 5 that only included the AI

companion and control conditions.

Methods and results

 This study was pre-registered (https://aspredicted.org/H37_GL2). We recruited 776

participants from CloudResearch Connect and excluded 63 for failing a comprehension question,

leaving 713 (MAge = 36.8, 56.8% Females). We aimed to hire 400 participants in both conditions

and participants were randomly assigned to control or AI companion conditions. 54.1% had prior

experience with AI companion apps. We ran this experiment on May 23, 2024. All participants

were paid $2.75 USD. The study and chatbot design were the same as in study 5, except that (1)

we removed the loneliness questions before the interaction with the chatbot, (2) removed the

feeling heard and chatbot performance questions; and (3) removed the ‘AI assistant’ condition.

As in study 5, we found that loneliness was significantly lower in the AI companion

condition compared to the control condition (MAI Companion = 25.62; MControl = 36.91; t(698.8)

= 5.37, p < .001, d = 0.40), confirming the robustness of our findings.

GENERAL DISCUSSION

We considered whether AI companions (applications that utilize AI to provide synthetic

social interactions) can reduce loneliness. Study 1 suggests that at least some consumers use AI

companion apps to alleviate loneliness, as some expressed loneliness-related thoughts explicitly

in naturalistic conversations with a chatbot. Study 2 found that loneliness-related content is

present in App Store reviews for a number of companion AIs, and that app reviews mentioning

loneliness tend to receive higher ratings, possibly because they effectively alleviate loneliness by

making users feel heard. Study 3 found that AI companions successfully alleviate loneliness on

par only with interacting with another person, whereas watching YouTube videos or doing

nothing do not. Furthermore, participants underestimate the degree to which AI companions

improve their loneliness relative to their true feelings after interacting with such AI. Study 4
found that the AI companion reduced loneliness over the course of a week. The most significant

reduction occurred on the first day, followed by stable decreases in loneliness on subsequent

days. This suggests that the initial interaction with the AI companion has a pronounced impact,

which quickly stabilizes over time as participants acclimate to their AI companions. Study 5

provided evidence that feeling heard and performance are significant mediators of the loneliness-

alleviating effects of AI companions, and feeling heard is the mediator with higher influence on

reducing loneliness compared to performance. Study 6 found that the loneliness alleviating

results of AI companions are robust when we ask about loneliness only after the interaction.

Theoretical and Methodological Contributions

We make several contributions. First, we increase understanding of why consumers are

using AI companions (Ta et al. 2020). While dictionary-based approaches have been used to

detect instances of clearcut mental health issues (De Freitas et al. 2023a), we found this approach

challenging in the context of loneliness, given polysemy of keywords and the diversity of

colloquial expressions of loneliness that were missed by these keywords. By introducing a novel

methodological approach in which we fine-tune LLMs for loneliness detection, we were able to

detect it even in real conversations and reviews. The fact that only a small amount of data was

needed for this purpose means that this method is particularly well suited for domains like ours,

in which (i) one is screening for instances that are less common in the overall dataset, and (ii)

thus, one does not want to use all such instances to train the model, because this could lead to

overfitting and hurt model generalization beyond the training dataset. Our methodological

pipeline can be replicated for other challenging classification tasks.

 Second, a number of reviews on loneliness and mental health have noted the need for

evidence on new technological solutions (Shrum et al. 2022; Veronese et al. 2021). In this space,

most work is correlational and qualitative (Maples et al. 2024; Ta et al. 2020), and the one

exception focused on an extreme sample of older patients with serious mental illness during the

pandemic, using a highly rule-based bot (Chou et al. 2024). Our paper includes several

experimental studies using state-of-the-art LLMs that causally isolate the impact of today’s AI-

based companions on loneliness, showing that they are more effective than other common

technological solutions and control conditions, at both cross-sectional and longitudinal scales.

We also show how behavioral engagement with commercially representative versions of the

technology predict loneliness reduction.

Finally, we contribute to understanding what features of chatbots lead to alleviation of

loneliness (Merrill Jr et al. 2022), by leveraging insights from human-human psychological

studies (Itzchakov et al. 2023; Kahlon et al. 2021; Myers 2000; Reis et al. 2017) and a growing

literature in human-computer interaction on the role of ‘feeling heard’ on relationships with

chatbots (Boucher et al. 2021; De Gennaro et al. 2020; Leite et al. 2013). We find that using

prompting to ensure that the AI is friendly and caring improves the sense that users feel heard,

relative to general assistants without these capabilities, and that feeling heard explains levels of

loneliness reduction.

Managerial and Societal Contributions

Many AI companions available on the market advertise loneliness alleviation as a value

proposition, but to the best of our knowledge our studies are the first to rigorously and causally

assess whether this is the case. Our analysis of existing reviews shows that some apps may truly
be delivering on this value proposition, and that they could employ the approach utilized here to

demonstrate this empirically. This finding is not only relevant to AI companion apps, but also

mental health apps that are increasingly incorporating ‘talk therapy’ as part of their offerings. As

shown in study 5, even AI assistants may alleviate loneliness to an extent. Finally, the results

document the benefits of building “generalist” LLM-based chatbots with empathic features

designed to make consumers feel heard. For example, Inflection AI’s Pi was a chatbot explicitly

designed to be, and marketed as being, a friendly conversationalist.

From a societal point of view, the promising results found here suggest that AI

companions could be a scalable tool for loneliness reduction against the backdrop of a severe

loneliness crisis. Whether chatbots can help reduce loneliness has recently been the object of

intense debate (Marriott and Pitardi 2024) and rigorous empirical evidence in this area was

sorely needed. In a series of tightly controlled and high-powered experimental studies, we find

compelling evidence that AI companions can indeed reduce loneliness, at least at the time scales

of a day and a week.

Limitations and Future Research

Our apps were set up to be caring and friendly, inducing the sense of feeling heard. Apps

optimized for other purposes might not produce similar loneliness alleviation benefits. Future

research should explore further what features of chatbots lead consumers to feel heard and also

what other psychological processes might contribute to loneliness alleviation. Moreover, it

remains an open question how using such apps over a much longer-term affects loneliness.

We found that participants tended to underestimate the positive impact of AI

companions. Future work can focus on reasons behind this misprediction, such as a general lack
of familiarity with AI companions or more specific stereotypes about chatbots, like beliefs that

chatbots are not capable of genuine understanding or emotional support.

Another important topic for future research is exploring other ways in which interactions

with AI companions may affect consumers’ socioemotional processes. For example, it is possible

that interacting with chatbots that are endowed with the features of particular social groups may,

over time, help weaken or change stereotypes. In addition, chatbots may affect consumers’

willingness to engage in similar conversations with other humans, increase people’s proficiency

in social interactions with strangers, reduce social anxiety, improve mental health by providing

emotional support, or improve other societally relevant outcomes. Finally, while our studies offer

valuable insights, it is important to note that they were conducted within a specific cultural

context (the US), which may influence perceptions and effectiveness of AI companions. The

generalizability of these findings across different cultures warrants further investigation. Future

cross-cultural research can employ our approach, especially in other cultures with high levels of

loneliness like Japan, where societal openness to social robots could provide further insights.
 REFERENCES

APA (2012), "Understanding Psychotherapy and How It Works,"

https://www.apa.org/topics/psychotherapy/understanding#:~:text=You%20and%20your%

20psychologist%20will%20also%20keep%20exploring%20your%20problems,you%20cl

arify%20what's%20troubling%20you.

Araujo, Theo (2018), "Living up to the Chatbot Hype: The Influence of Anthropomorphic

Design Cues and Communicative Agency Framing on Conversational Agent and

Company Perceptions," Computers in Human Behavior, 85, 183-89.

Austin, Peter C (2011), "An Introduction to Propensity Score Methods for Reducing the Effects

of Confounding in Observational Studies," Multivariate behavioral research, 46 (3), 399-

424.

Banks, Marian R, Lisa M Willoughby, and William A Banks (2008), "Animal-Assisted Therapy

and Loneliness in Nursing Homes: Use of Robotic Versus Living Dogs," Journal of the

American Medical Directors Association, 9 (3), 173-77.

Barbieri, Francesco, Luis Espinosa Anke, and Jose Camacho-Collados (2021), "Xlm-T:

Multilingual Language Models in Twitter for Sentiment Analysis and Beyond," arXiv

preprint arXiv:2104.12250.

Bergner, Anouk S, Christian Hildebrand, and Gerald Häubl (2023), "Machine Talk: How Verbal

Embodiment in Conversational Ai Shapes Consumer–Brand Relationships," Journal of

Consumer Research, ucad014.

Bertacchini, Francesca, Eleonora Bilotta, and Pietro Pantano (2017), "Shopping with a Robotic

Companion," Computers in Human Behavior, 77, 382-95.

Borsci, Simone, Alessio Malizia, Martin Schmettow, Frank Van Der Velde, Gunay

Tariverdiyeva, Divyaa Balaji, and Alan Chamberlain (2022), "The Chatbot Usability

Scale: The Design and Pilot of a Usability Scale for Interaction with Ai-Based

Conversational Agents," Personal and ubiquitous computing, 26, 95-119.

Boucher, Eliane M, Nicole R Harake, Haley E Ward, Sarah Elizabeth Stoeckl, Junielly Vargas,

Jared Minkel, Acacia C Parks, and Ran Zilca (2021), "Artificially Intelligent Chatbots in

Digital Mental Health Interventions: A Review," Expert Review of Medical Devices, 18

(sup1), 37-49.

Brakus, J Joško, Bernd H Schmitt, and Lia Zarantonello (2009), "Brand Experience: What Is It?

How Is It Measured? Does It Affect Loyalty?," Journal of Marketing, 73 (3), 52-68.

Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla

Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, and Amanda Askell (2020),

"Language Models Are Few-Shot Learners," Advances in Neural Information Processing

Systems, 33, 1877-901.

Cacioppo, John T and Stephanie Cacioppo (2018), "Loneliness in the Modern Age: An

Evolutionary Theory of Loneliness (Etl)," in Advances in Experimental Social

Psychology, Vol. 58: Elsevier, 127-97.

Castelo, Noah, Johannes Boegershausen, Christian Hildebrand, and Alexander P Henkel (2023),

"Understanding and Improving Consumer Reactions to Service Bots," Journal of

Consumer Research, 50 (4), 848-63.

Catsaros, Oktavia (2023), "Generative Ai to Become a $1.3 Trillion Market by 2032, Research

Finds," Bloomberg.
Chaves, Ana Paula and Marco Aurelio Gerosa (2021), "How Should My Chatbot Interact? A

Survey on Social Characteristics in Human–Chatbot Interaction Design," International

Journal of Human–Computer Interaction, 37 (8), 729-58.

Chen, Shu-Chuan, Wendy Moyle, Cindy Jones, and Helen Petsky (2020), "A Social Robot

Intervention on Depression, Loneliness, and Quality of Life for Taiwanese Older Adults

in Long-Term Care," International psychogeriatrics, 32 (8), 981-91.

Chou, Ya-Hsin, Chemin Lin, Shwu-Hua Lee, Yen-Fen Lee, and Li-Chen Cheng (2024), "User-

Friendly Chatbot to Mitigate the Psychological Stress of Older Adults During the Covid-

19 Pandemic: Development and Usability Study," JMIR Formative Research, 8, e49462.

Christen, Peter, David J Hand, and Nishadi Kirielle (2023), "A Review of the F-Measure: Its

History, Properties, Criticism, and Alternatives," ACM Computing Surveys, 56 (3), 1-24.

Chung, Minjee, Eunju Ko, Heerim Joung, and Sang Jin Kim (2020), "Chatbot E-Service and

Customer Satisfaction Regarding Luxury Brands," Journal of business research, 117,

587-95.

Croes, Emmelyn AJ and Marjolijn L Antheunis (2021), "Can We Be Friends with Mitsuku? A

Longitudinal Study on the Process of Relationship Formation between Humans and a

Social Chatbot," Journal of Social and Personal Relationships, 38 (1), 279-300.

Crolic, Cammy, Felipe Thomaz, Rhonda Hadi, and Andrew T Stephen (2022), "Blame the Bot:

Anthropomorphism and Anger in Customer–Chatbot Interactions," Journal of Marketing,

86 (1), 132-48.

Dang, Nhan Cach, María N Moreno-García, and Fernando De la Prieta (2020), "Sentiment

Analysis Based on Deep Learning: A Comparative Study," Electronics, 9 (3), 483.

De Freitas, J, Ahmet Kaan Uğuralp, Zeliha-Oğuz Uğuralp, and Stefano Puntoni (2023a),

"Chatbots and Mental Health: Insights into the Safety of Generative Ai," Journal of

Consumer Psychology, 00, 1–11.

De Freitas, Julian, Stuti Agarwal, Bernd Schmitt, and Nick Haslam (2023b), "Psychological

Factors Underlying Attitudes toward Ai Tools," Nature Human Behaviour, 7, 1845–54.

De Freitas, Julian and Nicole Tempest Keller (2022), "Replika Ai: Monetizing a Chatbot,"

Harvard Business Publishing.

De Gennaro, Mauro, Eva G Krumhuber, and Gale Lucas (2020), "Effectiveness of an Empathic

Chatbot in Combating Adverse Effects of Social Exclusion on Mood," Frontiers in

psychology, 10, 495952.

Devlin, Jacob, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova (2018), "Bert: Pre-

Training of Deep Bidirectional Transformers for Language Understanding," arXiv

preprint arXiv:1810.04805.

Eccles, Alice M and Pamela Qualter (2021), "Alleviating Loneliness in Young People–a Meta‐

Analysis of Interventions," Child and Adolescent Mental Health, 26 (1), 17-33.

Epley, Nicholas and Juliana Schroeder (2014), "Mistakenly Seeking Solitude," Journal of

Experimental Psychology: General, 143 (5), 1980-99.

Esch, Franz‐Rudolf, Tobias Langner, Bernd H Schmitt, and Patrick Geus (2006), "Are Brands

Forever? How Brand Knowledge and Relationships Affect Current and Future

Purchases," Journal of Product & Brand Management, 15 (2), 98-105.

Fournier, Susan (1998), "Consumers and Their Brands: Developing Relationship Theory in

Consumer Research," Journal of Consumer Research, 24 (4), 343-73.

Gable, Shelly L and Harry T Reis (2010), "Good News! Capitalizing on Positive Events in an

Interpersonal Context," in Advances in Experimental Social Psychology, Vol. 42:

Elsevier, 195-257.

Gasteiger, Norina, Kate Loveys, Mikaela Law, and Elizabeth Broadbent (2021), "Friends from

the Future: A Scoping Review of Research into Robots and Computer Agents to Combat

Loneliness in Older People," Clinical interventions in aging, 941-71.

Gilbert, Richard L and Andrew Forney (2015), "Can Avatars Pass the Turing Test? Intelligent

Agent Perception in a 3d Virtual Environment," International journal of human-computer

studies, 73, 30-36.

Hawkley, Louise C and John T Cacioppo (2010), "Loneliness Matters: A Theoretical and

Empirical Review of Consequences and Mechanisms," Annals of behavioral medicine, 40

(2), 218-27.

Hayes, Andrew F. (2012), "Process: A Versatile Computational Tool for Observed Variable

Mediation, Moderation, and Conditional Process Modeling [White Paper]," Retrieved

from http://www.afhayes.com/public/process2012.pdf.

Hoffman, Donna L and Thomas P Novak (1996), "Marketing in Hypermedia Computer-

Mediated Environments: Conceptual Foundations," Journal of Marketing, 60 (3), 50-68.

Holt-Lunstad, Julianne (2021), "Loneliness and Social Isolation as Risk Factors: The Power of

Social Connection in Prevention," American Journal of Lifestyle Medicine, 15 (5), 567-

73.

Holt-Lunstad, Julianne, Theodore F Robles, and David A Sbarra (2017), "Advancing Social

Connection as a Public Health Priority in the United States," American Psychologist, 72

(6), 517.
Holt-Lunstad, Julianne, Timothy B Smith, Mark Baker, Tyler Harris, and David Stephenson

(2015), "Loneliness and Social Isolation as Risk Factors for Mortality: A Meta-Analytic

Review," Perspectives on psychological science, 10 (2), 227-37.

Holzwarth, Martin, Chris Janiszewski, and Marcus M Neumann (2006), "The Influence of

Avatars on Online Consumer Shopping Behavior," Journal of Marketing, 70 (4), 19-36.

Hughes, Mary Elizabeth, Linda J Waite, Louise C Hawkley, and John T Cacioppo (2004), "A

Short Scale for Measuring Loneliness in Large Surveys: Results from Two Population-

Based Studies," Research on aging, 26 (6), 655-72.

Ipsos (2021), "Loneliness on the Increase Worldwide, but an Increase in Local Community

Support," https://www.ipsos.com/en-cn/loneliness-increase-worldwide-increase-local-

community-support.

Itzchakov, Guy, Netta Weinstein, Dvori Saluk, and Moty Amar (2023), "Connection Heals

Wounds: Feeling Listened to Reduces Speakers’ Loneliness Following a Social Rejection

Disclosure," Personality and Social Psychology Bulletin, 49 (8), 1273-94.

Jiang, Albert Q, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh

Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, and

Lucile Saulnier (2023), "Mistral 7b," arXiv preprint arXiv:2310.06825.

Jiménez-Barreto, Jano, Natalia Rubio, and Sebastian Molinillo (2023), "How Chatbot Language

Shapes Consumer Perceptions: The Role of Concreteness and Shared Competence,"

Journal of Interactive Marketing, 58 (4), 380-99.

Kahlon, Maninder K, Nazan Aksan, Rhonda Aubrey, Nicole Clark, Maria Cowley-Morillo,

Elizabeth A Jacobs, Rhonda Mundhenk, Katherine R Sebastian, and Steven Tomlinson

(2021), "Effect of Layperson-Delivered, Empathy-Focused Program of Telephone Calls

 on Loneliness, Depression, and Anxiety among Adults During the Covid-19 Pandemic: A

Randomized Clinical Trial," JAMA psychiatry, 78 (6), 616-22.

Kardas, Michael, Amit Kumar, and Nicholas Epley (2022), "Overly Shallow?: Miscalibrated

Expectations Create a Barrier to Deeper Conversation," Journal of personality and social

psychology, 122 (3), 367.

Kull, Alexander J, Marisabel Romero, and Lisa Monahan (2021), "How May I Help You?

Driving Brand Engagement through the Warmth of an Initial Chatbot Message," Journal

of business research, 135, 840-50.

Leite, Iolanda, André Pereira, Samuel Mascarenhas, Carlos Martinho, Rui Prada, and Ana Paiva

(2013), "The Influence of Empathy in Human–Robot Relations," International journal of

human-computer studies, 71 (3), 250-60.

Lim, Eric (2020), "App-Store-Scraper," https://pypi.org/project/app-store-scraper/.

Liu, Lijun, Zhenggang Gou, and Junnan Zuo (2016), "Social Support Mediates Loneliness and

Depression in Elderly People," Journal of health psychology, 21 (5), 750-58.

Liu, Yinhan, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike

Lewis, Luke Zettlemoyer, and Veselin Stoyanov (2019), "Roberta: A Robustly Optimized

Bert Pretraining Approach," arXiv preprint arXiv:1907.11692.

Liu-Thompkins, Yuping, Shintaro Okazaki, and Hairong Li (2022), "Artificial Empathy in

Marketing Interactions: Bridging the Human-Ai Gap in Affective and Social Customer

Experience," Journal of the academy of marketing science, 50 (6), 1198-218.

Luo, Xueming, Marco Shaojun Qin, Zheng Fang, and Zhe Qu (2021), "Artificial Intelligence

Coaches for Sales Agents: Caveats and Solutions," Journal of Marketing, 85 (2), 14-32.
Maples, Bethanie, Merve Cerit, Aditya Vishwanath, and Roy Pea (2024), "Loneliness and

Suicide Mitigation for Students Using Gpt3-Enabled Chatbots," npj Mental Health

Research, 3 (1), 4.

Markov, Todor, Chong Zhang, Sandhini Agarwal, Florentine Eloundou Nekoul, Theodore Lee,

Steven Adler, Angela Jiang, and Lilian Weng (2023), "A Holistic Approach to Undesired

Content Detection in the Real World," in Proceedings of the AAAI Conference on

Artificial Intelligence, Vol. 37, 15009-18.

Marriott, Hannah R and Valentina Pitardi (2024), "One Is the Loneliest Number… Two Can Be

as Bad as One. The Influence of Ai Friendship Apps on Users' Well‐Being and

Addiction," Psychology & marketing, 41 (1), 86-101.

Masi, Christopher M, Hsi-Yuan Chen, Louise C Hawkley, and John T Cacioppo (2011), "A

Meta-Analysis of Interventions to Reduce Loneliness," Personality and Social

Psychology Review, 15 (3), 219-66.

Merrill Jr, Kelly, Jihyun Kim, and Chad Collins (2022), "Ai Companions for Lonely Individuals

and the Role of Social Presence," Communication Research Reports, 39 (2), 93-103.

Montoya, Amanda K and Andrew F Hayes (2017), "Two-Condition within-Participant Statistical

Mediation Analysis: A Path-Analytic Framework," Psychological Methods, 22 (1), 6.

Muniz Jr, Albert M and Thomas C O'guinn (2001), "Brand Community," Journal of Consumer

Research, 27 (4), 412-32.

Myers, Sharon (2000), "Empathic Listening: Reports on the Experience of Being Heard,"

Journal of Humanistic Psychology, 40 (2), 148-73.

Nass, Clifford and Youngme Moon (2000), "Machines and Mindlessness: Social Responses to

Computers," Journal of social issues, 56 (1), 81-103.

Palgi, Yuval, Amit Shrira, Lia Ring, Ehud Bodner, Sharon Avidor, Yoav Bergman, Sara Cohen-

Fridel, Shoshi Keisari, and Yaakov Hoffman (2020), "The Loneliness Pandemic:

Loneliness and Other Concomitants of Depression, Anxiety and Their Comorbidity

During the Covid-19 Outbreak," Journal of Affective Disorders, 275, 109-11.

Perlman, Daniel and Letitia Anne Peplau (1982), "Theoretical Approaches to Loneliness,"

Loneliness: A sourcebook of current theory, research and therapy, 36, 123-34.

Poscia, Andrea, Jovana Stojanovic, Daniele Ignazio La Milia, Mariusz Duplaga, Marcin

Grysztar, Umberto Moscato, Graziano Onder, Agnese Collamati, Walter Ricciardi, and

Nicola Magnavita (2018), "Interventions Targeting Loneliness and Social Isolation

among the Older People: An Update Systematic Review," Experimental gerontology,

102, 133-44.

Qi, Yuxing and Zahratu Shabrina (2023), "Sentiment Analysis Using Twitter Data: A

Comparative Application of Lexicon-and Machine-Learning-Based Approach," Social

Network Analysis and Mining, 13 (1), 31.

Reis, Harry T, Edward P Lemay Jr, and Catrin Finkenauer (2017), "Toward Understanding

Understanding: The Importance of Feeling Understood in Relationships," Social and

Personality Psychology Compass, 11 (3), e12308.

Roos, Carla Anne, Tom Postmes, and Namkje Koudenburg (2023), "Feeling Heard:

Operationalizing a Key Concept for Social Relations," Plos one, 18 (11), e0292865.

Shrum, LJ, Elena Fumagalli, and Tina M Lowrey (2022), "Coping with Loneliness through

Consumption," Journal of Consumer Psychology.

Solaiman, Irene and Christy Dennison (2021), "Process for Adapting Language Models to

Society (Palms) with Values-Targeted Datasets," Advances in Neural Information

Processing Systems, 34, 5861-73.

Ta, Vivian, Caroline Griffith, Carolynn Boatfield, Xinyu Wang, Maria Civitello, Haley Bader,

Esther DeCero, and Alexia Loggarakis (2020), "User Experiences of Social Support from

Companion Chatbots in Everyday Contexts: Thematic Analysis," Journal of medical

Internet research, 22 (3), e16235.

Twenge, Jean M, Kathleen R Catanese, and Roy F Baumeister (2003), "Social Exclusion and the

Deconstructed State: Time Perception, Meaninglessness, Lethargy, Lack of Emotion, and

Self-Awareness," Journal of personality and social psychology, 85 (3), 409.

Veronese, Nicola, Daiana Galvano, Francesca D’Antiga, Chiara Vecchiato, Eva Furegon,

Raffaella Allocco, Lee Smith, Giovanni Gelmini, Pietro Gareri, and Marco Solmi (2021),

"Interventions for Reducing Loneliness: An Umbrella Review of Intervention Studies,"

Health & Social Care in the Community, 29 (5), e89-e96.

Wilson, Timothy D and Daniel T Gilbert (2003), "Affective Forecasting," Advances in

experimental social psychology, 35 (35), 345-411.

Zielinski, Melissa J and Jennifer C Veilleux (2018), "The Perceived Invalidation of Emotion

Scale (Pies): Development and Psychometric Properties of a Novel Measure of Current

Emotion Invalidation," Psychological assessment, 30 (11), 1454.

Zierau, Naim, Christian Hildebrand, Anouk Bergner, Francesc Busquet, Anuschka Schmitt, and

Jan Marco Leimeister (2023), "Voice Bots on the Frontline: Voice-Based Interfaces

Enhance Flow-Like Consumer Experiences & Boost Service Outcomes," Journal of the

academy of marketing science, 51 (4), 823-42.