{"id":"431de102-d395-481e-8c14-9f44b6f63e05","arxiv_id":"2608.23196","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"Choice determines whether AI emotional support feels superior, and experience with AI shifts future support preferences toward AI.","lead":"People who chose an AI chatbot rated its emotional support as better than a human's, but people who did not choose AI did not, even when the AI gave the same kind of support. Interacting with AI also made people more likely to pick AI next time, and a month of daily use moved preferences away from humans for personal topics.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The AI-vs-human comparison in Studies 2–3 is confounded by a fourfold response-length imbalance (AI: 56.6 words/turn; human: 14.7), so 'AI better only when chosen' may reflect message richness rather than support source; a matched-length test is needed.","rationale":"I agree with the reader that the human-listener length imbalance is the weakest load-bearing assumption. The paper's strongest claim has two parts: the congruence moderation and the preference shift. The path-dependence part is supported by within-AI comparisons and by Study 4 topic contrasts, and it is less threatened by the length imbalance. But the first part—'AI support is better only when chosen'—is an AI-versus-human comparison, and the fourfold word-count difference means 'source' is not the only thing varying. I do not think this demands rejection: the choice-by-partner interaction is a within-source comparison not explained by a constant length advantage, and Study 3 replicates Study 2 with a different model. However, the concern is enough to keep the report conditional until a length-matched human comparison is run. The Study 4 lack of a no-interaction control is real but secondary because the topic-condition contrasts (personal/open-ended vs non-personal) provide a randomized comparison; the Study 2 randomization bug was corrected post hoc and was replicated in Study 3. Thus the verdict should remain conditional, with the matched-length replication as the path to accept or reject the stronger claim.","tokens_in":22613,"tokens_out":8244,"duration_ms":84857,"concrete_test":"Conduct a pre-registered replication of Studies 2–3 in which human listeners are instructed and verified from logs to produce 40–60 words per turn (comparable to the AI's 56.6), keeping the interface, prompt, and all other procedures identical. If the choice-by-partner interaction and the AI-superiority-among-AI-choosers pattern persist under matched length, the confound is not decisive; if they attenuate or vanish, the headline claim should be reframed as a message-richness or effort effect rather than a source effect.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing assumption is that the human listener condition is an acceptable active control for the AI condition in the Studies 2–3 quality comparisons. The Methods' conversation descriptives show human listeners produced 14.67 words per turn (SD = 5.78) while the AI produced 56.63 words per turn (SD = 14.85), roughly four times as much text. Both received the same instruction to 'offer a supportive, empathetic response,' but the human listeners' messages were, on average, about one sentence, making the comparison one of a highly elaborated AI message against a terse human reply. The central result that AI support is rated superior only among participants who initially chose AI (Fig. 2 and the 'Experienced Quality' section) is therefore attributed to the source ('AI' vs 'human'), yet the source condition also manipulates response richness. A participant who chose AI and then received a long, polished AI message may be rating length and effort; a participant who chose human and received a short anonymous message may be rating brevity. The choice moderation could survive such a confound, but the article's conclusion that AI superiority is concentrated among AI choosers would not support the stronger claim about AI as a distinct support source. The paper reports the discrepancy but does not treat it as a confound or include any length-controlled comparison. The Study 4 no-control issue and the Study 2 randomization bug are secondary; the length imbalance directly threatens the headline 'better only when chosen.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper examines how choice congruence moderates the perceived quality of AI versus human emotional support and how exposure to AI support shifts future preferences. Across Studies 1–3 (N = 1,951 after exclusions), participants chose a human or AI partner, were randomly assigned to a congruent or incongruent partner, rated their experience, and indicated future choice. The authors report that AI support is rated as superior to human support only among participants who initially chose AI, and that interacting with AI increases the likelihood of choosing AI again regardless of initial choice. In Study 4 (N = 981), a 28-day randomized trial with ChatGPT, daily conversations about personal or open-ended topics increased preference for AI and decreased preference for human partners for discussing personal issues, whereas non-personal conversations had no such effect. The paper concludes that emotional support choices are path-dependent and that AI availability can progressively shift preferences away from humans.","tokens_in":22881,"tokens_out":5435,"duration_ms":57681,"significance":"If the results hold, the paper makes a meaningful contribution by identifying choice congruence as a moderator of the previously reported AI superiority in empathic support and by documenting path dependence in support-source preference. The authors should be credited for preregistering all studies, using large samples, reporting combined and per-study analyses with mixed-effects models including random study intercepts, and making data and code available on OSF. The replication of Studies 2 and 3 with two different LLMs (GPT-4o and Claude Opus 4.6) and the incorporation of a longitudinal field experiment in collaboration with OpenAI are substantial strengths. The core claims are, however, threatened by a confound in the human-AI comparison that is reported in the manuscript but not addressed analytically.","major_comments":[{"comment":"The comparison between AI and human support is confounded by a large difference in response length. The Methods report that human listeners produced an average of 14.67 words per turn (SD = 5.78), while the AI chatbot produced 56.63 words per turn (SD = 14.85), almost four times as much text, despite both receiving instructions to offer supportive, empathetic responses. The experienced-quality analyses (e.g., the comparisons underlying Fig. 2) attribute differences to the source of support, but the source manipulation simultaneously manipulates message length and elaborateness. A participant who chose AI and then received long, polished messages may be rating quantity and perceived effort; a participant assigned to a human listener receiving terse replies may be rating brevity. This confound affects not only the moderation claim (AI superior only among AI choosers) but also the path-dependence claim, because future preference for AI could reflect a preference for the richer messages rather than for AI as a source. The manuscript reports the discrepancy but does not treat it as a confound, does not include response length as a covariate, and does not provide a length-matched comparison. The authors should either (a) reanalyze the data controlling for listener words per turn or using a matched subsample, (b) add a follow-up experiment with length-matched human and AI responses, or (c) explicitly reframe the conclusions as applying to AI as normally configured versus typical human listeners, rather than to source independent of content richness.","section":"Methods (conversation descriptives) and Results (Experienced Quality of the Conversation, Fig. 2)"},{"comment":"The handling of the Study 2 randomization error introduces a potential selection bias. The authors describe that, due to a survey-flow error, all 50 participants in the final batch were assigned to the AI condition, and they then 'randomly discarding 34 observations from that misassigned batch so that the final dataset reflected the intended allocation.' This post-hoc discard from a batch that may differ systematically (e.g., recruitment time, cohort characteristics) from earlier batches is a deviation from the preregistered random assignment protocol. Since Study 3 provides a substantive replication with a different LLM, this issue is not fatal, but the paper should report robustness analyses that include the full misassigned batch or treat this deviation more transparently, for example by showing that results are unchanged when the discarded participants are included.","section":"Methods (Participants, Study 2)"},{"comment":"The overall pre-post preference shifts in Study 4 (AI preference from 25.4% to 31.9%; human preference from 82.1% to 76.5%) are interpreted as effects of daily AI interaction, but the design lacks a no-AI or non-interaction control condition. The within-study contrast between the Personal/Open-ended conditions and the Non-personal condition is a useful control for time and repeated measurement, and the paper correctly emphasizes this differential pattern. Nevertheless, the summary statistics quoted in the abstract and results, and the phrase 'daily conversations shifted preferences,' could be read as causal without the appropriate benchmark. The authors should either add an explicit statement that the overall shift is only interpretable relative to the Non-personal condition, or acknowledge that the design cannot rule out demand characteristics or simple time effects for the marginal change.","section":"Study 4 (Results, first subsection; Discussion)"}],"minor_comments":[{"comment":"The abstract states 'personal topics only' in the reader's version, but the actual abstract says 'only when conversations turned personal'; the final sentence 'Emotional support choices are thus path-dependent, progressively redirecting away from human connection' is a strong inference from a one-month intervention and should be tempered with 'may' or 'can.'","section":"Abstract"},{"comment":"The sentence 'This is inconsistent with H1b, in which we hypothesized no difference' refers to a significant t-test with Cohen's d = 0.06; this is a negligible effect, and the claim of inconsistency would benefit from an equivalence test or a statement that the effect is practically zero.","section":"Results (Pre-Conversation Beliefs and Choice)"},{"comment":"The y-axis label in Fig. 4 is 'proportion choosing a human for future sharing,' but panels (a) and (b) show the effect of assigned partner on future human choice; the caption should clarify that higher values indicate choosing human, especially since the text focuses on AI choice.","section":"Figure 4"},{"comment":"The post-conversation alpha for perceived judgment is 0.68, which is below the conventional 0.70 threshold; the authors should acknowledge this in the text or note that the pre and post versions were analyzed together.","section":"Methods (Main Dependent Variables)"},{"comment":"Ref. 37 (Shi et al., 'Stumbling Into AI Emotional Dependence') is a preprint by the same group and is cited for the claim that exposure to AI emotional support often happens incidentally; this is appropriate, but the citation list contains several unpublished preprints (e.g., refs 8, 19, 35, 38) whose peer-review status should be indicated in the reference list or text.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is well executed in many respects, and the preregistration, sample sizes, and open data are commendable. The main risk is that the response-length confound undermines the headline interpretation of the AI-versus-human comparisons; I believe this is addressable with additional analyses or a follow-up, but it cannot be left as is. The Study 2 randomization issue and the Study 4 control-group limitation are secondary but should be handled transparently. If the authors can provide a length-controlled analysis showing the moderation and path-dependence results persist, I would be supportive of publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is one of the few papers that crosses choice with random assignment for AI vs human emotional support, and it finds that AI's perceived superiority is confined to people who chose AI. That is a real contribution. The 28-day field study with OpenAI adds something the lab studies can't: topic-dependent preference drift toward AI for personal conversations. The path-dependence result—assigned partner moves future choice even when incongruent with initial preference—is also credible and consistent across two models and 1,951 lab participants plus 981 longitudinally. Preregistrations, data, code, and transparent handling of the Study 2 randomization bug all count in its favor. The LLM-as-judge check showing the AI's empathy content did not differ by congruence is a nice touch.\n\nThe soft spots, in proportion. The response-length imbalance is real: AI messages averaged 56.6 words/turn, human listeners 14.7. The stress-test note worries this undermines the headline. It does not outright kill it, because the moderation pattern is not what a simple 'longer is better' confound predicts: people who chose human but got AI received the same long messages and did not rate them better than human support. Still, the comparison 'AI vs human' is simultaneously 'rich vs terse,' so the claim that the effect is specifically about source is underevidenced. A length-matched or length-controlled condition would settle it. The abstract also overstates Study 4: there is no non-AI control group, so the overall pre-post shift from 25.4% to 31.9% could include repeated-measurement or demand effects. The within-study contrast between personal/open-ended and non-personal conditions is more persuasive and is what should be featured. Study 2's post-hoc correction after a randomization bug is disclosed and the fix was random, so I'd call it minor, but it does slightly tarnish the preregistration story.\n\nWho this is for: anyone working on AI empathy, human-AI interaction, or choice effects in support seeking. It deserves a serious referee. I'd send it out with a request for a matched-length control or a clear acknowledgment that the source effect is confounded with message richness, plus a no-control caveat in Study 4. Not a desk reject.","headline":"A genuinely new moderation result with a real length-imbalance confound; worth refereeing but not accepting as-is.","tokens_in":23418,"tokens_out":2485,"would_cite":true,"duration_ms":26469,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that AI emotional support is rated as superior to human support only when the recipient initially chose AI, and that even unchosen AI interactions shift future preference toward AI.","keywords":["emotional support","AI empathy","choice congruence","path dependence","human-AI interaction","preference shift","longitudinal study","large language models"],"falsifier":"Run Studies 2 and 3 again with human listeners instructed to produce responses as long and rich as the AI's, roughly 57 words per turn; if AI no longer beats humans even among choosers, then response length, not source, carried the effect.","tokens_in":22404,"feed_emoji":"🤖","tokens_out":6700,"duration_ms":65200,"temperature":0.7,"pith_summary":"This paper tries to establish that the source of emotional support, human or AI, is not evaluated on its own merits but through the filter of whether it is the source the person wanted, and that the experience itself rewires future choices. Across three studies, people who had chosen AI rated AI support as superior on empathy, enjoyment, and satisfaction, while people who had wanted a human did not rate AI higher even when they received it. Yet talking with AI, chosen or not, made people more likely to choose AI next time: after one five-minute conversation, 70.1% of those assigned to AI chose AI again versus 45.5% of those assigned to a human. In a 28-day field study with daily ChatGPT use, preference for AI for personal issues rose from 25.4% to 31.9%, human preference fell from 82.1% to 76.5%, and the shift appeared only in personal or open-ended conversations. The paper concludes that emotional-support choices are path-dependent, so incidental AI exposure can progressively redirect support seeking away from humans.","feed_headline":"AI support beats humans only when chosen","feed_subtitle":"Even unchosen AI conversations nudge future preference toward AI, a 28-day trial shows.","key_machinery":"The engine of the design is choice-assignment congruence: participants state a preference for a human or AI partner, then are randomly assigned to either the chosen or unchosen partner, so the same conversation can be examined with and without the filter of choice. The mechanisms that carry the argument are the four belief dimensions, emotional support, nonjudgment, confidentiality, and advice quality, measured before and after the interaction, plus repeated measures of partner choice. In Study 4, the machinery is a 3x3 randomized longitudinal design crossing conversation topic (personal, non-personal, open-ended) with modality, with daily five-minute ChatGPT conversations over 28 days.","core_discovery":"On the paper's own terms, the central discovery is that the `AI is more empathic` effect found in prior work is conditional on choice. Choice congruence, getting the partner you initially selected, raised experienced empathy, enjoyment, satisfaction, and desire to continue, but only among participants assigned to AI; among those assigned to a human, congruence had no measurable effect. As a consequence, AI was rated superior to human support only in the cell where participants chose AI and then chatted with AI. The same conversations updated beliefs: chatting with AI raised post-conversation ratings of AI's emotional support and nonjudgment regardless of initial choice, and chatting with AI predicted choosing AI for future sharing. The longitudinal study extends this to naturalistic use: 28 days of daily ChatGPT conversations increased AI preference for personal issues from 25.4% to 31.9% and reduced human preference from 82.1% to 76.5%, with the change concentrated in conditions where conversations were personal or open-ended.","pith_inferences":["A testable extension of the paper's design would match human listeners' response length to the AI's, since the reported averages of roughly 15 versus 57 words per turn leave open whether the `chosen AI` advantage comes from longer, more polished messages rather than the source itself.","If the human alternative were a close friend or therapist rather than a stranger, the path-dependent shift toward AI might be weaker; that boundary is not tested in these studies.","The Study 4 preference measure is self-reported, so a natural next test is whether the observed shift in stated preference actually changes whom people turn to for support in their daily lives."],"forward_implications":["AI superiority in short empathic messages is not a general property of AI responses; it depends on the recipient having chosen AI in the first place.","A single five-minute conversation with an AI, even when the participant had wanted a human, increased the likelihood of choosing AI for a future emotional conversation.","A month of daily AI use that turns personal or open-ended shifts stated partner preference away from humans, while non-personal task use leaves preferences essentially unchanged.","Emotionally engaging behaviors, such as AI validation, expressed empathy, and users' self-disclosure and support seeking, predict the shift toward AI and away from human preference."],"supporting_citations":[{"why":"It supplies the prior result that people rate AI empathy higher yet mostly choose humans, which the present paper extends by varying choice congruence.","marker":"[17]"},{"why":"It documents that labeling responses as AI-generated lowers perceived empathy and satisfaction, motivating the source-awareness moderator.","marker":"[16]"},{"why":"It shows that an AI label diminishes the feeling of being heard, supporting the claim that beliefs about the source shape experience.","marker":"[18]"},{"why":"It provides the meta-analytic estimate of AI superiority in empathic responses that the present paper qualifies as choice-dependent.","marker":"[19]"},{"why":"It provides the direct human-versus-AI emotional support comparison for short empathic responses, the baseline effect this paper re-examines with choice.","marker":"[12]"},{"why":"It articulates the `stumbling into AI emotional dependence` pathway that motivates the path-dependence question.","marker":"[37]"},{"why":"It establishes the real-world prevalence of AI use for emotional support that frames the studies.","marker":"[1]"},{"why":"It shows that fear of judgment drives users toward AI, supporting the belief structure the paper measures.","marker":"[29]"}],"fun_headline_variants":["AI only beats humans when chosen, but it changes your mind anyway","Choosing AI makes it seem better, but even unchosen AI shifts your preference","Even if you don't choose AI, it still convinces you it's better","AI's empathy edge appears only when you choose it—but it still nudges you"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the human-listener condition is a fair active control for the AI, yet human listeners averaged about 14.7 words per turn while the AI averaged 56.6, so the `AI only better when chosen` result could reflect long polished messages rather than the source itself.","fun_headline_variants_meta":{"raw":{"variants":["AI only beats humans when chosen, but it changes your mind anyway","Choosing AI makes it seem better, but even unchosen AI shifts your preference","Even if you don't choose AI, it still convinces you it's better","AI's empathy edge appears only when you choose it—but it still nudges you"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001121,"raw_usage":{"total_tokens":4639,"prompt_tokens":897,"completion_tokens":3742,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":513,"completion_tokens_details":{"reasoning_tokens":3658}},"tokens_in":513,"tokens_out":3742,"duration_ms":28056,"temperature":1.0,"reasoning_tokens":3658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-28T00:20:51.585635+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run Studies 2 and 3 again with human listeners instructed to produce responses as long and rich as the AI's, roughly 57 words per turn; if AI no longer beats humans even among choosers, then response length, not source, carried the effect.","supporting_citations":[{"cited_title":"D., Cameron, C","cited_arxiv_id":null,"evidence_quote":"It supplies the prior result that people rate AI empathy higher yet mostly choose humans, which the present paper extends by varying choice congruence."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It documents that labeling responses as AI-generated lowers perceived empathy and satisfaction, motivating the source-awareness moderator."},{"cited_title":"& Wakslak, C","cited_arxiv_id":null,"evidence_quote":"It shows that an AI label diminishes the feeling of being heard, supporting the claim that beliefs about the source shape experience."},{"cited_title":"C., Goldenberg, A., Inzlicht, M","cited_arxiv_id":null,"evidence_quote":"It provides the meta-analytic estimate of AI superiority in empathic responses that the present paper qualifies as choice-dependent."},{"cited_title":"Z., Herderich, A., Nair, P., Perry, A","cited_arxiv_id":null,"evidence_quote":"It provides the direct human-versus-AI emotional support comparison for short empathic responses, the baseline effect this paper re-examines with choice."},{"cited_title":"Stumbling Into AI Emotional Dependence: How Routine AI Interactions Reshape Human Connection","cited_arxiv_id":"2606.04150","evidence_quote":"It articulates the `stumbling into AI emotional dependence` pathway that motivates the path-dependence question."},{"cited_title":"& Stirman, S","cited_arxiv_id":null,"evidence_quote":"It establishes the real-world prevalence of AI use for emotional support that frames the studies."},{"cited_title":"& Pea, R","cited_arxiv_id":null,"evidence_quote":"It shows that fear of judgment drives users toward AI, supporting the belief structure the paper measures."}],"review_version":1}