{ "schemaVersion": "1.0", "language": "en", "canonicalUrl": "https://catalysphere.co.jp/en/articles/ai-mediated-empathy-authenticity/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-mediated-empathy-authenticity/index.json", "publisher": { "name": "Catalysphere", "url": "https://catalysphere.co.jp/", "description": "A research and consulting studio connecting AI and big-data research with practice." }, "article": { "slug": "ai-mediated-empathy-authenticity", "title": "AI-written empathy does not replace human relational responsibility", "summary": "We design an operation that distinguishes the quality of AI-generated empathetic text from the recipient's perceived understanding, connection, and source preference, while safeguarding human editorial responsibility and recipient choice.", "takeaways": [ "AI can improve empathetic expression in limited scenarios, but high ratings for text do not equate to the authenticity or connection of human relationships.", "This is because experimental results show that when content is recognized as AI-generated, its evaluation drops, and a preference for receiving it from a human persists.", "We limit AI to drafting and feedback, ensure humans take responsibility for facts and relational accountability, and maintain pathways for recipients to choose human responses." ], "publishedAt": "2026-07-15", "updatedAt": "2026-07-24", "readingMinutes": 9, "topics": [ "Behavioral Science", "Conversational AI", "Personalization", "Organizational Design" ], "domains": [ "generative-ai", "organization-adoption", "governance-risk" ], "evidenceKind": "research-review", "status": "published", "image": { "prompt": "Abstract scientific editorial illustration for the research question “AIは、共感のある文章を人の関係へ戻せるか”, interpreted symbolically rather than typographically: a contextual signal field moving through perception and choice, then branching into several observable outcomes while one restrained gold path remains traceable; deep Catalysphere navy and indigo field, muted scientific blue, restrained antique-gold structure, translucent geometry, fine particles, subtle paper grain, generous negative space, one clear centered metaphor that survives a narrow mobile cover crop, calm premium research-journal art direction, no readable text, no logo, no literal chart, no photorealistic people.", "negativePrompt": "text, letters, numerals, watermark, logo, dashboard UI, humanoid robot, server rack, binary rain, stock-photo office, literal quantitative chart, red alarm graphics, neon cyberpunk, purple gradient, clutter", "alt": "An abstract diagram representing behavioral paths branching from context to perception, choice, and multiple outcomes in dark blue, blue, and gold", "aspectRatio": "16:9", "assetPath": "article-art/ai-mediated-empathy-authenticity.png", "status": "approved" }, "lead": "In texts addressing complaints, failures, loss, and anxiety, recipients demand not only correct information but also the sensation of being understood and relational accountability. While AI can stably propose empathetic phrasing, the recipient's evaluation upon knowing the source is AI, and their preference for receiving it from a human, operate separately. This article integrates primary research from Nature Machine Intelligence, PNAS, Nature Human Behaviour, and Communications Psychology to design usage as mediation rather than substitution. [1][2][3][4]", "sections": [ { "heading": "Do not summarize empathy into a single score.", "paragraphs": [ "Empathetic texts may involve correctly grasping the other's situation, acknowledging emotions, and indicating necessary support. On the other hand, recipients experience the sensation of being understood, reassurance, connection with the other, and trust in the information source. Even if third-party evaluations of the text are high, it does not necessarily mean the recipient feels a relational connection.", "Furthermore, we distinguish between the actual creator and the creator believed by the recipient. The content and attribution differ between texts generated by AI and edited by humans, written by humans and proofread by AI, and texts presented as human-created but generated by AI. The starting point is not to mix 'AI text quality' and 'effects of knowing it is AI' into the same comparison.", "This article does not determine whether AI possesses emotions or consciousness. It also does not address the diagnosis or therapeutic effects of psychological disorders. The focus is on recipient outcomes and relational accountability when AI assists human communication in contexts involving emotions, such as customer service, colleague support, and peer support." ] }, { "heading": "Feedback to humans can elevate empathetic expression.", "paragraphs": [ "Sharma et al. conducted a non-clinical randomized trial with 300 actual supporters on an online peer mental health support platform. They reported that in the HAILEY group, where AI provided immediate feedback to responses, conversational empathy was 19.6% higher overall compared to the control group, and 38.9% higher for the group that self-reported difficulty in providing support. [1]", "Crucially, this was not a trial where AI provided answers to recipients alone. Humans adopted, modified, or ignored AI suggestions, remaining the final responders. This study positioned AI not as the subject of empathy, but as a feedback device for humans to understand their counterparts and write responses. [1]", "The limitations are also clear. It focused on peer support on a single platform, short conversations, and trained automatic empathy metrics, without measuring clinical outcomes or long-term relationship quality. [1] One co-author disclosed being a co-founder and shareholder of a psychological support quality technology company, which is recorded as a conflict of interest when interpreting the results." ], "bullets": [ "ASSIST: AI does not replace responders but provides drafts and feedback.", "EDIT: Humans verify context, facts, and tone, retaining records of adoption/rejection and reasons for changes.", "ESCALATE: Content related to crisis, medical care, safety, and rights is handed over to humans." ] }, { "heading": "Text quality and AI labels act independently on recipients.", "paragraphs": [ "Yin et al. (PNAS 2024) presented participants describing complex experiences with responses created by humans or Bing Chat, crossing the actual creator with displays indicating AI or human origin. While AI-generated texts enhanced the sensation of 'being heard' compared to human texts, this effect decreased when labeled as AI. [2]", "This design is important because it separates content effects from attribution effects. The fact that AI captured emotions more accurately and included more emotional support without rushing practical advice correlated with content evaluation. Conversely, AI labels reduced the perception of connection and understanding. [2] Improving content does not eliminate disclosure issues, nor does a drop in evaluation due to disclosure justify hiding it.", "The human comparison group consisted of online general participants who answered one case for compensation, not skilled supervisors, long-term colleagues, or trained counselors. [2] Additionally, there was no no-display condition, so it is not fully possible to separate whether AI displays lowered scores or human displays raised them." ] }, { "heading": "The same text holds more value if believed to be from a human.", "paragraphs": [ "Rubin et al. (Nature Human Behaviour 2025) conducted nine studies involving 6,282 people, displaying AI-generated empathetic texts as created by either humans or AI. The same text attributed to humans was evaluated as more empathetic and supportive, and generated more positive emotions, than when attributed to AI. [3]", "These effects were replicated across variations in text length, response delay, repetition, and multiple models, and were reported to be larger in responses emphasizing emotional sharing and care. [3] The perception that a human directed time and attention toward them may add to the meaning of the text. Even if automation shortens deadlines, the relational investment is not considered equivalent.", "The studies primarily relied on self-reports in experimental environments, without long-term tracking of actual workplace relationships, customer retention, or conflict resolution. [3] Therefore, the average decrease in evaluation due to AI disclosure is used not as a reason to hide it, but as a basis for designing usage selection and recipient expectations." ], "bullets": [ "CONTENT: Evaluate the accuracy, acknowledgment, and supportiveness of the text.", "ATTRIBUTION: Distinguish who created, who edited, and who is responsible.", "RELATION: Measure understanding, connection, and authenticity separately from text quality." ] }, { "heading": "Even if highly evaluated, it is not necessarily desired to be received.", "paragraphs": [ "Wenger et al. (Communications Psychology 2026) distinguished between the choice to receive empathy from humans or AI and the evaluation of the received response in four studies. Participants generally chose to receive from humans, but responses received from AI after choosing it were evaluated highly for quality, sensation of being heard, and effort. [4]", "This 'paradox of choice' indicates that satisfaction assigned passively cannot predict demand. Even if those who tried AI responses evaluate them highly, it does not justify defaulting to AI for those who seek humans from the start. The right to choose human response and the differences in wait time and cost must be made explicit.", "The four studies included hypothetical scenarios and pre-generated texts, with students and online participants as the center. While some studies used real emotional scenarios, they did not establish daily continuous use or relationship repair. [4] Do not mix preferences and benefits into an average; include those who did not choose in the denominator." ] }, { "heading": "Measure recipient outcomes by content, attribution, and choice.", "paragraphs": [ "In pilot tests, we distinguish conditions: human-only, human optionally using AI feedback, human editing AI drafts, and AI responding directly. We save not only the use of AI but also the amount of final text editing, factual corrections, responder, disclosure, and the pathway chosen by the recipient.", "Primary indicators are whether the recipient felt understood, reached necessary next actions, and had no factual errors. Secondary indicators include connection, authenticity, switch to human response, re-explanation, complaints, and regret. If switching to human response is impossible or false reassurance increases, it does not pass, even if text evaluation is high.", "On the sender side, we measure the adoption rate of AI proposals, but also the addition of context, removal of excessive assertions, removal of personal information, and execution of emergency handovers. To avoid mistaking the addition of empathetic clichés for understanding the other, we verify which part of the original statement was addressed." ], "bullets": [ "RECIPIENT: Measure the sensation of being understood and preference for human response from the individual.", "ACCURACY: Independently verify facts, promises, policies, and urgency.", "CHOICE: Enable selection of AI support, human-only, reply later, or termination with equal operational load." ] }, { "heading": "Establish operational boundaries that leave relational accountability to humans.", "paragraphs": [ "In low-impact standard replies, humans can confirm AI drafts and clarify the responsible person and entity. For complaints, evaluations, loss, health, and safety, AI is not used as an emotional proxy; the responsible person reads the counterpart's specific context and edits with their own judgment. We do not send promises or apologies created by AI without confirmation.", "Disclosure does not end with a uniform small label. We explain, depending on the situation, whether AI handled drafting, translation, summarization, or direct response, and who made the final confirmation. If the recipient desires a human, we transfer the history without forcing them to repeat the same conversation.", "Stop conditions include factual errors, missed crises, failure to switch to human response, massive distrust after disclosure, and decreased evaluation for specific groups. To resume, we narrow the draft scope, demonstrate handover, and re-test recipient understanding, choice, and relational outcomes. The role of AI is not to perform empathy but to increase the margin for humans to respond responsibly to their counterparts." ], "bullets": [ "Adoption: Content and recipient outcomes improve with human editing, and the right to switch functions.", "Limitation: In emotional and high-impact cases, remain limited to drafts and feedback.", "Stop: Direct response increases false reassurance, missed crises, and relational distrust.", "Resume: Restore responsible person accountability and handover to humans, and re-confirm recipient evaluation." ] } ], "note": "This article is a research review integrating non-clinical peer support, online experiments, and experimental emotional scenarios; it is not a meta-analysis evaluating the efficacy and safety of psychotherapy or medical treatment. [1][2][3][4] Many results are based on short-term self-reports and cannot be generalized to culture, existing relationships, crisis scenarios, or long-term relationship repair. In Sharma et al.'s study, a conflict of interest with a related technology company was disclosed by one author. [1]", "sources": [ { "id": 1, "authors": "Ashish Sharma, Inna W. Lin, Adam S. Miner, David C. Atkins, Tim Althoff", "year": 2023, "title": "Human–AI collaboration enables more empathic conversations in text-based peer-to-peer mental health support", "venue": "Nature Machine Intelligence, 5, 46–57", "url": "https://doi.org/10.1038/s42256-022-00593-2", "finding": "In a non-clinical RCT with 300 actual peer supporters, the AI feedback group showed 19.6% higher conversational empathy overall and 38.9% higher in the group feeling difficulty in support.", "limitation": "Focused on a single platform, short text conversations, and automatic empathy metrics, with clinical outcomes unmeasured. One co-author is a co-founder and shareholder of a related psychological support quality technology company." }, { "id": 2, "authors": "Yidan Yin, Nan Jia, Cheryl J. Wakslak", "year": 2024, "title": "AI can help people feel heard, but an AI label diminishes this impact", "venue": "Proceedings of the National Academy of Sciences, 121(14), e2319112121", "url": "https://doi.org/10.1073/pnas.2319112121", "finding": "By crossing actual creators with displays, AI-generated texts showed an independent effect of enhancing the sensation of being heard, while AI displays lowered the same outcomes.", "limitation": "Human comparisons involved online general participants, not skilled supporters or long-term relationships. There was no no-display condition, and primary outcomes were self-reports after one response." }, { "id": 3, "authors": "Matan Rubin, Joanna Z. Li, Federico Zimmerman, Desmond C. Ong, Amit Goldenberg, Anat Perry", "year": 2025, "title": "Comparing the value of perceived human versus AI-generated empathy", "venue": "Nature Human Behaviour, 9, 2345–2359", "url": "https://doi.org/10.1038/s41562-025-02247-w", "finding": "Across nine studies and 6,282 people, the same AI-generated empathetic text was evaluated as more empathetic and supportive when displayed as human-origin, with better emotional responses.", "limitation": "Primarily self-reports in experimental environments, without tracking existing relationships with workplaces or customers, conflict resolution, or long-term behavioral and relational outcomes." }, { "id": 4, "authors": "Joshua D. Wenger, C. Daryl Cameron, Michael Inzlicht", "year": 2026, "title": "People choose to receive human empathy despite rating AI empathy higher", "venue": "Communications Psychology, 4, Article 19", "url": "https://doi.org/10.1038/s44271-025-00387-3", "finding": "In four studies, participants chose to receive empathy from humans, while the AI responses they actually chose were evaluated highly for quality and sensation of being heard.", "limitation": "Included hypothetical scenarios and pre-generated texts, centered on students and online participants. Did not establish safety in natural continuous use or high-impact emotional scenarios." } ], "visual": { "eyebrow": "EMPATHY / AUTHENTICITY", "title": "Separate the quality of empathy from the reception of its origin.", "caption": "The observation conditions of each study differ. Please refer to them as a starting point for evaluation design and operational judgment, not for side-by-side superiority.", "metrics": [ { "value": "+19.6%", "label": "Conversational empathy in the AI feedback group", "source": "Sharma et al., 2023", "referenceId": 1 }, { "value": "+38.9%", "label": "Empathy difference in the group feeling support difficulty", "source": "Sharma et al., 2023", "referenceId": 1 }, { "value": "9 studies / 6,282 participants", "label": "Integrate origin display and empathy evaluation", "source": "Rubin et al., 2025", "referenceId": 3 } ] }, "url": "https://catalysphere.co.jp/en/articles/ai-mediated-empathy-authenticity/", "machineReadableUrl": "https://catalysphere.co.jp/en/articles/ai-mediated-empathy-authenticity/index.json", "topicFacets": [ { "id": "behavioral-science", "name": "Behavioral Science", "url": "https://catalysphere.co.jp/en/topics/behavioral-science/" }, { "id": "organization-adoption", "name": "Organization & Adoption", "url": "https://catalysphere.co.jp/en/topics/organization-adoption/" } ] } }