AI emotional support is better only when chosen, but shifts preferences even when it is not

• A meta-analysis estimated AI’s superiority in providing short empathic responses at Cohen’s d = 0.55 [19].Source: AI and Emotional Support · p. 4
• Most prior studies concealed the support source and removed choice, while labeling responses AI-generated reduced satisfaction and perceived empathy [16,18,20,21].Source: AI and Emotional Support · p. 4
• The paper first maps beliefs about human and AI emotional support and examines their association with choice.Source: AI and Emotional Support · p. 5
• Because limited access can lead people to receive support from a nonpreferred source, the paper tests how congruence or incongruence shapes interaction-quality evaluations.Source: AI and Emotional Support · p. 6
• The paper examines path dependence—whether support experiences shift future source preferences—amid the possibility that widespread AI availability creates incidental exposure.Source: AI and Emotional Support · pp. 6–7

The Current Studies

• Studies 1–4 examined beliefs and choices about human versus AI support, effects of choice congruence on support experiences, and preference changes after AI interaction.The program comprised a survey, two experiments, and a longitudinal study.Source: The Current Studies · p. 7
• Study 1 asked participants to recall a recent significant emotional experience, rate human and AI support beliefs on four dimensions, and choose a sharing partner.The dimensions were emotional support, nonjudgment, confidentiality, and advice quality.Source: The Current Studies · p. 7, Experimental Design · p. 10

Experimental Design

• Studies 2 and 3 randomly assigned participants, after their choice, to a five-minute text conversation with a human or an empathically instructed AI, congruent or incongruent with that choice.Study 2 used ChatGPT and Study 3 used Claude.Source: The Current Studies · p. 8, Experimental Design · p. 10
• After the conversation, participants rated their experience on the same four dimensions and reported empathy, enjoyment, satisfaction, desire to continue, and future partner preference.Future preference was measured with a binary choice and continuous ratings.Source: The Current Studies · p. 8, Experimental Design · p. 11

The Current Studies

• Study 4 used a 28-day collaboration with OpenAI to examine preference changes after daily five-minute AI conversations on personal, non-personal, and open-ended topics.The study also examined the role of conversational behaviors in predicting future preferences.Source: The Current Studies · p. 8
• Studies 1–3 recruited 1,951 participants; after preregistered exclusions, the Studies 2–3 conversation sample included 944 sharers, 415 in the human condition and 529 in the AI condition.The human condition involved paired sharer-listener interactions, and listeners were excluded from the primary analyses.Source: Emotional Support · pp. 8–9
• Combined regression analyses used mixed-effects models with a random intercept for study to allow baseline outcomes to vary across studies.The models accounted for between-study heterogeneity.Source: Emotional Support · p. 9

How do people perceive sharing emotional experiences with a human and an AI?

• Across Studies 1–3, humans were rated higher in emotional support (mean difference, 0.79) and judgment (1.32); AI was more confidential (0.39), while advice quality did not differ (P = 0.743).Source: How do people perceive sharing… · p. 12, Figure 1 · p. 13
• Emotional-support differences were the strongest predictors of partner choice, closely followed by nonjudgment differences.Source: How do people perceive sharing… · pp. 13–14, Table 1 · p. 14
• AI was chosen by 75.5% of Study 1 participants, 69.6% in Study 2, and 63.2% in Study 3 when choosing a partner for sharing an emotional experience.Source: How do people perceive sharing… · p. 14
• Participants reported greater comfort and willingness to share with AI than with a human (standardized mean difference, 0.51 SD; Cohen’s d = 0.36).Source: How do people perceive sharing… · p. 15
• Beliefs tracked choice: AI choosers saw AI as less judgmental, more confidential, and better at advice, with comparable support; human choosers saw equal nonjudgment but rated humans higher in support, confidentiality, and advice.Source: How do people perceive sharing… · p. 16
• Perceptions include support capabilities, advice quality, and confidentiality.Source: How do people perceive sharing… · p. 16

Experienced Quality of the Conversation

• The study examined whether choice-partner congruence influenced conversation quality and whether this relationship differed between human and AI interactions.Source: Experienced Quality of the Conversation · p. 16
• Studies 2 and 3 assigned five-minute conversations with human or AI partners regardless of preference; the authors hypothesized congruence would improve sharing experience (H2a).Study 2 used GPT-4o and Study 3 used Claude Opus 4.6.Source: Experienced Quality of the Conversation · p. 17

How does the congruence between choice and chat condition influence

• Participants whose assigned partner matched their initial choice reported greater empathy, desire to continue, enjoyment, and satisfaction.These were positive main effects of choice–assignment congruence across the four outcomes.Source: How does the congruence between choice… · p. 17
• Congruence improved all four outcomes among participants assigned to AI, but had no significant effect among those assigned to a human.This moderation indicates that the overall congruence effects were driven primarily by participants who chose and received AI.Source: How does the congruence between choice… · pp. 17–18
• AI emotional support was rated more favorably than human support only by participants who had initially chosen AI.The choice-by-partner interactions covered empathy, enjoyment, desire to continue, and satisfaction.Source: How does the congruence between choice… · p. 17
• Choice congruence did not change the empathy level in AI responses, suggesting it altered subjective experiences rather than the responses themselves.The estimated effect on coded AI empathy was β = 0.0005; 95% CI, (−0.04, 0.04).Source: How does the congruence between choice… · pp. 18–19

Post-Conversation Perception and Choice

• After an AI chat, perceived emotional support rose versus pre-conversation beliefs (mean difference, 1.07; 95% CI, (0.97, 1.17)), as did perceived nonjudgment (mean difference, 0.52; 95% CI, (0.43, 0.61)).These changes were reported regardless of initial sharing choice; Fig. 3 presents the pre- to post-conversation changes.Source: Post-Conversation Perception and Choice · p. 20, Figure 3 · p. 21
• Among participants who initially chose a human, chatting with AI raised perceived AI emotional support to the pre-chat level of those who initially chose AI.The groups’ pre-conversation beliefs were not significantly different (mean difference, -0.06; 95% CI, (-0.28, 0.16)).Source: Post-Conversation Perception and Choice · p. 20
• After a human chat, participants perceived human partners as more nonjudgmental, but perceived human emotional support did not increase significantly.The support change was not significant (mean difference, -0.15; 95% CI, (-0.31, 0.02)).Source: Post-Conversation Perception and Choice · p. 20, Figure 3 · p. 21
• Participants who initially chose a human reported receiving less emotional support from their human partner than they had anticipated.The mean difference was -0.69; 95% CI, (-0.98, -0.41).Source: Post-Conversation Perception and Choice · p. 20

How does the interaction change people’s future choice on who to share emotions

• 70.1% of participants assigned to chat with AI chose AI for future emotional sharing, compared with 45.5% assigned to chat with a human.The difference followed a five-minute emotional conversation.Source: How does the interaction change… · p. 22
• The findings suggest that emotional-support choices are path-dependent: sharing with a partner increases the likelihood of choosing that same partner in the future.The main effect was consistent among participants who initially chose either AI or a human.Source: How does the interaction change… · p. 22
• Among initial AI choosers, human (vs. AI) chat predicted future human choice (β = 0.70; 95% CI, (0.50, 0.91)); among initial human choosers, AI (vs. human) chat predicted future AI choice (β = 0.61; 95% CI, (0.31, 0.90)).Both effects were significant at P < 0.001.Source: How does the interaction change… · p. 22, Figure 4 · p. 23

Study 4: Examining Effects of Using AI on Future Preference in a Longitudinal Study

• The 4-week randomized trial asked participants to have at least a 5-minute ChatGPT conversation daily and assigned them to Personal, Non-personal, or Open-ended topics.Participants reported partner preferences before and after the 28 days; conversations could use text or voice, but modality was outside this paper’s focus.Source: Experimental Design · pp. 24–25
• Emotional Support & Empathy made up 43.1% of conversations in the Personal condition, 9.6% in Open-ended, and a negligible share in Non-personal.The topic classification check showed that the conversation-topic manipulation produced the expected differences.Source: Experimental Design · pp. 25–26

personal matters?

• AI preference for discussing personal issues rose from 25.4% to 31.9% after 28 days of AI interaction (t(980) = 3.88; P < 0.001; d = 0.12), while human preference shifted in the opposite direction.The human-preference share began at 82.1%; the supplied passage does not give its endpoint.Source: personal matters? · p. 26
• AI preference rose by 11.6% in Personal (31.6% → 43.2%) and 9.5% in Open-ended (24.9% → 34.4%) conversations, but fell 1.2% in Non-personal conversations (20.2% → 19.0%).Human preference declined in Personal (82.6% → 72.3%, −10.3%) and Open-ended (80.5% → 75.1%, −5.4%) conditions, but changed little in Non-personal (83.1% → 81.6%, −1.5%).Source: personal matters? · pp. 26–27
• Compared with Non-personal conversations, future AI preference increased more in Open-ended (OR = 2.25, 95% CI, (1.55, 3.25), Tukey-adjusted P < 0.001) and Personal conditions (OR = 3.03, 95% CI, (2.09, 4.38), Tukey-adjusted P < 0.001).The decrease in human preference was significantly larger in Personal than Non-personal conversations (OR = 0.56, 95% CI, (0.38, 0.83), Tukey-adjusted P = 0.010); the Open-ended decrease was not significant (Tukey-adjusted P = 0.15).Source: personal matters? · p. 27
• Models predicted post-study preference from behavior rates while controlling for pre-study preference, conversation topic, and modality, with correction across 13 tests per outcome.Source: What conversational behaviors predict… · p. 28
• Twelve of 13 conversation behaviors significantly predicted both higher odds of choosing AI and lower odds of choosing a human for personal issues.Figure 7 presents the behavior-preference associations.Source: What conversational behaviors predict… · p. 28, Figure 7 · p. 29
• AI feeling validation (OR = 1.98) and affection (OR = 1.80) predicted choosing AI; AI empathy most strongly predicted lower odds of choosing a human (OR = 0.55).All three associations had P < 0.001.Source: What conversational behaviors predict… · pp. 28–29
• Sharing problems (OR = 1.91), seeking support (OR = 1.85), and disclosing feelings (OR = 1.80) predicted greater future preference for AI and lower odds of choosing a human.Each association had P < 0.001 for AI preference; all predicted lower odds of choosing a human (all OR < 1, P < 0.01).Source: What conversational behaviors predict… · p. 29
• Only AI inquiry into personal information predicted neither outcome (both P > 0.13); results point to emotionally engaging AI behaviors as relevant to future support preferences.Source: What conversational behaviors predict… · p. 29

Discussion

• People who preferred humans believed humans offered better emotional support and were similarly nonjudgmental, whereas AI-preferrers saw equal support ability but greater AI nonjudgmentality.The authors suggest these differing beliefs shape the choice of emotional support partner.Source: Discussion · p. 30
• AI was rated higher than humans only when participants had initially preferred and been assigned to AI.The authors conclude that AI superiority in emotional support is not a general effect, but is concentrated among participants who initially preferred AI.Source: Discussion · p. 30
• Regardless of initial choice, interacting with AI about emotional or personal experiences updated beliefs about its support capacity and increased the likelihood of choosing AI again.In the longitudinal study, five-minute daily ChatGPT conversations were conducted for 28 days.Source: Discussion · pp. 30–31
• 70.1% of participants assigned to AI later chose AI again, compared with 45.5% of those assigned to a human.Participants were more likely to choose whichever partner they had initially been assigned to.Source: Discussion · p. 31
• 25.4% to 31.9%: AI preference for personal conversations rose, while human preference fell from 82.1% to 76.5% over 28 days.Open-ended and personal-issue conversations also increased willingness to use AI for personal issues and decreased likelihood of choosing a human.Source: Discussion · p. 31
• The extent of the shift remains uncertain: AI may move preferences only so far, or may lead some people to stop preferring human connection.The authors propose that systems detecting emotional interactions could encourage users to reach out to supportive humans, potentially mitigating a shift toward AI.Source: Discussion · p. 32

Limitations and future directions

• The lab studies compared AI with a human stranger, so future research should examine choices and experiences when friends, partners, family members, or professionals are available.The authors note that results may differ when people choose between AI and a friend.Source: Limitations and future directions · p. 32
• Emotional support may emerge incidentally during task-oriented use of general-purpose AI platforms [37,38], leaving unclear whether effects observed in more structured settings generalize across diverse real-world scenarios.Source: Limitations and future directions · p. 33
• Daily AI use for a month in the open-ended condition increased preference for discussing personal matters with AI and decreased preference for human partners.This condition was the closest to typical everyday use.Source: Limitations and future directions · p. 33
• The authors call for naturalistic research on how people experience emotional support, how choice influences it, and how support-partner preferences evolve.They specifically recommend testing the Study 4 results in naturalistic contexts.Source: Limitations and future directions · p. 33
• Because preference change was measured by self-report rather than actual support-seeking, future work should track real-world behavior and assess effects on relationships and well-being over time.Whether reported preferences translate into changes in whom people actually approach for support remains open.Source: Limitations and future directions · p. 33

Ethical statement

• Studies 1–3 received Harvard IRB approval and Study 4 received WCG IRB approval; participants in all studies gave informed consent.Source: Ethical statement · p. 33

Preregistrations

• All studies were preregistered; the passage provides the Study 1 registration link.Source: Preregistrations · p. 34

Participants

• Study 1 recruited 250 Prolific participants and retained 229 after attention-check and consent exclusions.Source: Participants · p. 34
• Studies 2–3 recruited U.S.-resident, fluent-English CloudResearch Connect users with at least 90% approval; final post-conversation samples were 487 and 457, respectively.Source: Participants · pp. 35–37
• Study 4’s final participant set comprised 981 people.Source: Participants · p. 38
• 463 participants (47.2%) had used ChatGPT text at least a few times a week.Source: Participants · p. 38
• 683 participants (69.6%) had never used ChatGPT voice.The passage suggests this may be because voice was a new feature.Source: Participants · p. 38
• The survey reported use of assistant-type chatbots more than a few times a week (e.g., Google’s Gemini, Anthropic’s Claude), while most participants (696, 70.9%) had never used companion chatbots.Source: Participants · p. 38

Studies 1–3: Procedure

• Participants first recalled a significant emotional experience from the past month, described it in 4–5 sentences, and rated its valence and eight discrete emotions.They were told they would later discuss the experience with another participant or an AI chatbot.Source: Procedure · p. 38
• Study 1 asked participants to choose a human or AI partner before rating both options, whereas Studies 2–3 reversed this order to address potentially choice-motivated ratings.Studies 2–3 also measured the importance of emotional support, judgment concerns, confidentiality, and advice quality.Source: Procedure · p. 39
• In Studies 2–3, participants rated human and AI partners before assignment, then completed a five-minute text conversation with an empathetic AI chatbot or human listener.The AI prompt was matched to the human listener instructions; AI models were GPT-4o in Study 2 and Claude Opus 4.6 in Study 3.Source: Procedure · pp. 39–40
• After the conversation, participants rated their partner, conversation outcomes, current emotions, and future preferences for sharing emotional experiences with humans or AI.The post-conversation measures included enjoyment, satisfaction, desire to continue, and both binary and continuous future partner-preference items.Source: Procedure · p. 41

Study 4

• Study 4 randomized 981 participants across nine conditions in a 3 × 3 design crossing text or voice modality with open-ended, non-personal, or personal conversation prompts for 28 days.Participants were asked to use ChatGPT for at least five minutes daily and completed baseline, weekly, and post-study surveys.Source: Procedure · p. 42

Procedure

• Dependent variables were considered relative to baseline measures.Source: Procedure · p. 42

Main Dependent Variables

• A 14-item measure assessed beliefs about emotional sharing across perceived judgment, emotional support, advice, and confidentiality.Participants rated beliefs before the conversation and the assigned partner afterward, enabling comparison of expectations with experience; items used a 7-point scale.Source: Main Dependent Variables · pp. 42–43
• In Studies 1–3, participants chose whether they preferred an AI companion or another participant and rated their desire to talk with each.Study 1 used a 0–10 comfort scale; Studies 2–3 used a 1–7 scale for how much they wanted to talk.Source: Main Dependent Variables · pp. 43–44
• After conversations in Studies 2–3, participants rated satisfaction, enjoyment, and desire to continue with their partner on 1–7 scales.These three ratings indexed experienced interaction quality.Source: Main Dependent Variables · p. 44
• In Studies 2 and 3, future partner choice was assessed by asking participants whether they would talk about a future emotional experience with a human stranger or an AI companion.Source: Main Dependent Variables · p. 44
• In Study 4, participants reported choices across seven scenarios, including personal issues, before and after the 28-day study.For personal issues, options included chatbot, human online, human in person, or no preference; responses were also binarized as choosing AI or human versus not choosing each.Source: Main Dependent Variables · p. 44

• Preprocessed data exclude participants’ emotional experiences and chat messages.Source: Data availability · p. 45