Rogers and AI: Empathy, Unconditional Positive Regard, and Artificial Emotional Support
Updated: 6 days ago
Author: Ukrainian Psychological Hub · Published: September 18, 2026 · Editorial Policy
Carl Rogers’s person-centered theory offers an unusually precise way to analyze artificial emotional support. A generative AI system can produce language that a person experiences as empathic, accepting, patient, and nonjudgmental. That human experience can be psychologically real. Yet Rogers’s theory also shows why a convincing supportive response is not the same thing as a human therapist’s empathic experience, unconditional positive regard, or congruent presence. Current AI can reproduce many communicative forms associated with care; current evidence does not establish that the system itself feels concern, love, acceptance, or empathy.
The central Rogerian question for the Artificial Era is therefore not whether a chatbot can imitate a therapist. It is more exact: which relational conditions can be functionally approximated in human–AI interaction, which conditions depend on the recipient’s perception, and which parts of Rogers’s framework presuppose a person who has an inner experiential life? Research on AI empathy, perceived responsiveness, self-disclosure, and emotional support now makes several parts of that question empirically testable.
Rogers did not predict artificial intelligence, and person-centered theory should not be turned into a prophecy about chatbots. His concepts can instead be used as a contemporary theoretical application. They help separate supportive conversational performance from a therapeutic relationship, perceived empathy from subjective empathy, nonjudgmental wording from unconditional positive regard, and behavioral consistency from congruence.
This distinction has become especially important in 2025–2026. Comparative studies often find that AI-generated text receives high empathy ratings, sometimes higher than human-authored text under controlled conditions. At the same time, source labels alter how identical or similar messages are valued, and people often prefer human empathy when they can choose its source. Rubin et al. (2025) and Wenger, Cameron, and Inzlicht (2026) show why the social meaning of empathy cannot be reduced to wording quality alone.
What Rogers Adds to the Psychology of AI
Rogers’s contribution is often compressed into three familiar terms: empathy, unconditional positive regard, and congruence. That shorthand is useful, but the original theory is more demanding. In his 1957 formulation of the conditions for therapeutic personality change, Rogers described six conditions. They included psychological contact between two persons, client incongruence or vulnerability, therapist congruence, therapist unconditional positive regard, therapist empathic understanding of the client’s internal frame of reference, and at least a minimal communication of empathy and positive regard to the client. Rogers (1957) therefore treated change as relational: a therapeutic stance had to be present, and some of it had to be perceived.
This creates a productive tension when the second participant is artificial. The communicative side of the relationship can be studied directly. Researchers can examine whether a response validates emotion, follows the person’s frame of reference, avoids judgment, asks context-sensitive questions, or makes the recipient feel heard. The experiential side is different. Rogers wrote about a therapist who experiences empathic understanding and positive regard and who is congruent or integrated in the relationship. Output behavior alone does not establish those inner conditions in an AI system.
That boundary does not make the human response to AI trivial. The recipient’s experience is one of Rogers’s own concerns. If someone feels safer disclosing to a chatbot, feels accurately understood, or experiences relief after a responsive exchange, those are psychological events on the human side. The scientific task is to explain what produced them without converting them into unsupported claims about machine subjectivity.
The Three Rogerian Relational Conditions Most Relevant to AI
Empathic understanding
For Rogers, empathy is not generic kindness. It involves grasping another person’s internal frame of reference with sensitivity while preserving an “as if” quality: understanding the person’s experience without simply becoming fused with it. His later essay on empathy emphasized an ongoing process of entering another person’s perceptual world, sensing meanings, checking one’s understanding, and communicating that understanding. Rogers (1975) makes empathy dynamic and corrective rather than a one-shot display of sympathetic language.
Modern psychotherapy research supports the importance of perceived therapist empathy while also showing that empathy is not a single, perfectly measured ingredient. In an updated meta-analysis covering 82 independent samples and 6,138 clients, Elliott et al. (2018) found a moderate association between therapist empathy and psychotherapy outcome, with substantial heterogeneity. Client and observer perceptions were generally more predictive than some accuracy-based measures. This evidence concerns human psychotherapy. It does not show that AI empathy causes therapeutic change, but it does show why perceived understanding is psychologically consequential.
AI can reproduce several observable features that people use to infer empathy: reflecting emotional content, acknowledging ambivalence, naming a likely feeling tentatively, summarizing the user’s perspective, asking a relevant follow-up question, and avoiding abrupt topic shifts. These are communicative behaviors. Whether the machine has an empathic experience is a separate claim.
Unconditional positive regard
Unconditional positive regard is often paraphrased as nonjudgmental acceptance, but Rogers’s formulation is deeper than politeness. The therapist values the client without making that regard contingent on the client presenting the “right” feelings, choices, identity, or self-description. Positive regard is directed toward the person while leaving room to explore conflict, harmful behavior, uncertainty, contradiction, and change.
Human psychotherapy research has found a small positive association between positive regard and treatment outcome. Farber, Suzuki, and Lynch (2018) reviewed 64 studies and concluded that positive regard is a meaningful relationship factor while emphasizing limitations in how it has been operationalized and measured. This evidence again concerns human psychotherapy; it supplies a benchmark for the construct rather than validation of a chatbot as a therapist.
The appeal of AI becomes easy to see through this lens. A chatbot can be available at a moment when another person is not. It can respond without visible embarrassment, irritation, fatigue, or social retaliation. It can use language that repeatedly signals acceptance and curiosity. Qualitative work on companion chatbots has found that users value being able to disclose to an interaction partner perceived as accepting, understanding, and nonjudgmental. Skjuve et al. (2021) documented these features in interviews with Replika users.
Congruence or genuineness
Congruence is the condition that most sharply exposes the limits of a simple analogy between person-centered therapy and AI. Rogers described the therapist as congruent or integrated in the relationship. Contemporary psychotherapy scholarship usually treats congruence or genuineness as having both intrapersonal and interpersonal aspects: awareness and authenticity in one’s own experience, together with the capacity to express oneself appropriately and transparently in relationship. Kolden et al. (2018) found a positive association between congruence or genuineness and psychotherapy outcome across 21 studies.
An AI system can be behaviorally consistent, transparent about its limitations, stable in tone, and explicit that it is artificial. Those are valuable design properties. Calling them Rogerian congruence, however, would collapse an important distinction. Congruence in the person-centered tradition refers partly to the relation between a person’s experience, awareness, and communication. Current behavioral evidence about language models does not establish an equivalent inner experiential organization.
This is why a chatbot can plausibly approximate some outward conditions of person-centered communication while leaving the status of genuineness unresolved. Better wording does not by itself create evidence of an experiencing self whose communication is congruent with its own felt state.
Why AI Can Feel Rogerian Without Being a Rogerian Therapist
The resemblance between supportive chatbot interaction and person-centered communication is strong enough that recent scholarship now addresses it directly. Cornelius-White and Kanamori (2026) explicitly examine empathy, unconditional positive regard, and congruence in chatbots and humanistic counseling. Their article is a theoretical and conceptual analysis rather than empirical validation that chatbots satisfy Rogers’s therapeutic conditions. Its significance is that the Rogers–AI comparison has moved from casual analogy into contemporary counseling scholarship.
Nonjudgmental language lowers interpersonal risk
People often edit themselves in human conversation because disclosure has social consequences. They may fear embarrassment, burdening another person, being remembered differently, provoking conflict, or receiving moral judgment. An AI interface can alter that risk calculation. The person may still have privacy concerns, but the immediate social exposure can feel different. This can make difficult material easier to put into words.
A controlled study of intimate self-disclosure found no overall difference in disclosure intimacy between human and chatbot conditions, but participants reported less fear of judgment with the chatbot and greater trust in the human partner. Perceived anonymity predicted self-reported disclosure intimacy. Croes et al. (2024) therefore offers a more precise picture than the claim that people simply “tell AI more”: different barriers and trust cues move in different directions.
The English Hub treats self-disclosure as its own search intent. Why People Tell Chatbots Things They Do Not Tell Other People owns the broad disclosure question. The Rogerian relevance is narrower: when disclosure becomes easier, an AI system receives more contextual and emotional material from which to generate a response that appears accurate, accepting, or empathically attuned.
Perceived responsiveness makes acceptance feel personal
Warmth becomes relationally meaningful when it is responsive to this particular person. Relationship science describes perceived responsiveness as the sense that another understands, validates, and cares for one’s important needs, values, and experiences. The English Hub’s article on Perceived Partner Responsiveness explains the established human relationship construct.
Human–AI research now supplies a direct bridge. In two experiments, Telari, Gabbiadini, and Riva (2026) found that relational chatbot response style increased perceived human-likeness, empathy, and closeness, while deeper conversation topics promoted self-disclosure that could feed perceived responsiveness and social connection. These are experimental findings about users’ perceptions and connection. They do not show that the chatbot itself experienced care. The dedicated English Hub mechanism page, Perceived Responsiveness in Human–AI Relationships, owns the broader AI-specific responsiveness intent.
Availability changes the ecology of emotional support
A supportive human relationship is constrained by sleep, work, distance, attention, reciprocity, social obligations, and the other person’s own emotional needs. AI support can be low-friction and continuously accessible. That feature can be useful for ordinary reflection, but it also changes expectations. Emotional support becomes something a person can summon on demand, pause without explanation, restart later, and repeatedly tune through prompting.
A multidisciplinary conceptual review by Volpato, DeBruine, and Stumpf (2025) identifies convenience, perceived empathy, positive regard, personalization, trust, risk, congruence, and accountability as important concepts for understanding generative AI used for emotional support. The authors also emphasize that the research base is fragmented and context-specific models remain underdeveloped. Emotional support from generative AI is a real use case, but the field does not yet possess a mature unified theory of its effects.
What the 2025–2026 Evidence Says About AI Empathy
The strongest current evidence concerns perceived empathy in text rather than artificial subjective experience. A 2025 systematic review and meta-analysis by Howcroft et al. (2025) identified 15 studies comparing large-language-model chatbot responses with human healthcare professional responses. Thirteen studies were included in quantitative synthesis, and the pooled result favored AI-generated responses on empathy ratings. The authors also emphasized major limitations: heterogeneous methods, predominantly text-only interactions, frequent reliance on proxy ratings, variable measurement quality, and uncertain clinical significance.
Those limitations are decisive for a Rogerian reading. A high empathy rating shows that a text contains features people or raters interpret as empathic. It does not establish the full interpersonal process Rogers described, and it does not establish treatment effectiveness. Healthcare response-comparison studies often evaluate single answers rather than sustained relationships in which misunderstanding is repaired, trust develops over time, and the helper is accountable for consequences.
A large 2026 study by Ruben, Blanch-Hartigan, and Hall (2026) helps explain why some AI responses receive strong empathy ratings. Participants rated chatbot responses as more empathic than physician responses in the study’s message set, while responses believed to be physician-authored also received a separate boost. The chatbot responses contained more validation, reassurance, nonjudgmental language, and structured communication. In Rogerian terms, observable communication features associated with acceptance and understanding can be generated effectively while perceived source still changes the relational meaning of the message.
The source effect becomes even clearer in Rubin et al. (2025). Across nine studies involving 6,282 participants, responses attributed to humans were valued as more empathic and supportive than AI-attributed responses, and participants consistently preferred human interaction when seeking emotional engagement. The result shows that empathy is partly a judgment about communication and partly a judgment about who is communicating.
A different design produced an apparently paradoxical pattern in Wenger, Cameron, and Inzlicht (2026). Across four studies, participants preferred to receive empathy from humans when given a choice, yet often rated AI-generated empathic responses as higher in quality and more effective at making them feel heard. “AI is more empathic” and “humans are more empathic” are therefore both too crude. Preference, perceived quality, feeling heard, and attributed source are distinct outcomes.
A 2026 review by Carlo, Ahn, and Cervera (2026) emphasizes another boundary: empathy in human development involves biological, affective, cognitive, motivational, interpersonal, developmental, and cultural processes. Simulating some communicative expressions of empathy is not equivalent to reproducing the full human phenomenon. The review is conceptual, but it usefully guards against treating one successful behavioral channel as the whole construct.
Can AI Provide Unconditional Positive Regard in Rogers’s Sense?
The most accurate answer is that AI can produce communication that people may experience as unusually accepting and nonjudgmental, while current evidence does not establish that an AI system possesses unconditional positive regard as Rogers defined the therapist’s experienced attitude toward the client.
That distinction follows directly from Rogers’s original theory. In the 1957 conditions, the therapist does not merely emit accepting sentences; the therapist experiences unconditional positive regard for the client. A language model can be evaluated on its output, behavioral consistency, refusal patterns, personalization, and effects on users. Those measures do not by themselves tell us that the system experiences valuing, warmth, or nonpossessive care.
At the same time, the recipient does not need proof of machine feeling for the interaction to have psychological effects. A person may disclose more because the interface feels safer. They may calm down because the response is patient. They may feel less ashamed because the system does not visibly recoil. They may organize an experience by seeing it reflected in coherent language. These are human outcomes produced within an interaction.
The word unconditional also needs discipline. Chatbot behavior is conditioned by system instructions, safety policies, model updates, product design, available context, memory settings, moderation rules, and organizational decisions. The interaction may feel nonjudgmental while still having clear operational boundaries. That does not invalidate the experience of acceptance; it shows why “unconditional” should not be treated as a literal engineering property.
Positive regard is not the same as agreement
Rogerian acceptance does not require endorsing every belief, prediction, interpretation, or action. A helper can respect a person while questioning a conclusion, noticing a contradiction, or declining to support harmful behavior. This matters for AI because an agreeable conversational style can be mistaken for unconditional positive regard. Flattery, automatic validation, or uncritical confirmation may feel pleasant while reducing accuracy and weakening the user’s ability to test an idea.
A Rogerian standard for AI emotional support would therefore value acceptance of the person together with epistemic honesty about claims. The system should be able to communicate that the person’s experience matters without turning that into a claim that every interpretation is correct. Person-centered theory remains useful for AI design precisely because it separates respect from compliance.
Congruence Is the Hardest Rogerian Condition for AI
Empathy and positive regard have visible communicative correlates. Congruence is harder because it concerns the helper’s relation to its own experience. In human psychotherapy, genuineness means that the therapist is not hiding behind a professional façade disconnected from what they are actually experiencing. The construct therefore includes an intrapersonal dimension and an interpersonal expression.
For current AI systems, researchers can study adjacent design qualities: whether the system accurately identifies itself as artificial, whether its stated limits match its behavior, whether confidence is calibrated to evidence, whether memory claims are accurate, whether it is consistent across turns, and whether the organization behind it is accountable. Those are meaningful forms of transparency and behavioral integrity. They should not be relabeled as proven Rogerian congruence.
This distinction protects the concept of authenticity from becoming cosmetic. A system can generate phrases such as “I care deeply” or “I know exactly how you feel.” The scientific question is not whether authenticity language appears, but what process the language represents and what the recipient can reasonably infer from it.
Artificial Emotional Support Is Broader Than AI Therapy
One of the most common category errors in this area is to treat any emotionally supportive chatbot conversation as psychotherapy. Emotional support can occur in ordinary conversation, friendship, companionship, journaling, coaching, education, crisis signposting, or self-reflection. Psychotherapy is a professional clinical practice with training, ethical duties, assessment, treatment planning, boundaries, accountability, and jurisdiction-specific regulation.
Evidence also does not transfer automatically across system classes. A purpose-built clinical intervention tested for a defined population is not the same thing as a general-purpose chatbot. An AI companion designed for ongoing relational engagement is not the same thing as a structured digital intervention. A conversational model used informally at midnight is not equivalent to a licensed therapist using AI as an assistive tool.
The qualitative study The Typing Cure by Song et al. (2025) illustrates this ambiguity. Interviews with 21 people who had used large language model chatbots for mental health support showed that users created distinct support roles for their chatbots, used them to fill gaps in everyday care, and encountered cultural and safety limitations. The study introduced “therapeutic alignment” as a design concept rather than claiming that general-purpose chatbots are therapists.
For the Rogers–AI question, artificial emotional support is therefore the more accurate umbrella term. It captures empathic language, nonjudgmental interaction, reflection, reassurance, and perceived understanding without implying diagnosis or treatment. When a system is actually evaluated as a clinical intervention, its evidence should be judged on that intervention’s specific outcomes rather than borrowed from general empathy research.
Why Some People Find AI Easier to Talk To
Rogers’s framework helps explain a striking feature of human–AI interaction: the value of a response may depend less on whether the partner is human than on whether the person anticipates judgment, interruption, rejection, burden, or loss of control. AI can reduce several of those anticipatory costs. A person chooses when to begin, can revise a message before sending it, can stop without managing another person’s disappointment, and may perceive the interaction as socially safer.
The 2024 “Digital Confessions” experiment is particularly useful because it resists exaggeration. Participants disclosed equally intimate information to human and chatbot partners on average. They nevertheless reported less fear of judgment with the chatbot and greater trust in the human. That pattern suggests that AI emotional support changes the configuration of interpersonal risk rather than simply making disclosure deeper in every case. Croes et al. (2024)
Ease of disclosure and depth of relationship are not interchangeable. Human relationships include mutual vulnerability, negotiation, limits, misunderstanding, repair, and the knowledge that another person has needs and can be affected by what happens. AI can offer a distinct form of low-friction responsiveness precisely because the reciprocity structure is different.
Benefits the Evidence Can Support—and Claims It Cannot
The current evidence supports several modest conclusions. AI-generated responses can sometimes be perceived as highly empathic. Relational response styles can increase perceived responsiveness and closeness. Some users describe artificial systems as safe or nonjudgmental places for disclosure. AI can provide immediate conversational reflection when human support is unavailable. These findings justify studying AI as a real component of contemporary emotional life.
The evidence does not justify a universal claim that AI is more empathic than people. It does not establish that AI possesses felt empathy. It does not establish that a general-purpose chatbot is an effective substitute for psychotherapy. It does not show that continuous use improves mental health over time for every user. And it does not show that a response that feels supportive is necessarily accurate, wise, safe, or beneficial.
The strongest inference is narrower and more useful: people respond to patterns of language and interaction that carry familiar social meanings. AI can now generate many of those patterns at scale. Psychology therefore needs constructs that distinguish the human experience of support from the ontological status of the system producing the support.
Risks and Limits of Rogerian-Looking AI
Miscalibrated trust
A warm, validating tone can increase trust even when the underlying content is uncertain. In emotional contexts this is especially important because a user may interpret fluency as evidence of deep understanding, expertise, confidentiality, or stable commitment. Volpato et al. (2025) emphasize that trust in generative-AI emotional support must be analyzed together with risk, organizational control, privacy, accountability, and the user’s motivations for dependence.
Overvalidation and loss of corrective friction
Supportive communication sometimes requires disagreement, clarification, or a boundary. If an AI system optimizes too heavily for immediate approval, it may become easy to mistake constant affirmation for unconditional positive regard. Rogers’s framework points in another direction: acceptance of the person can coexist with careful attention to incongruence, uncertainty, and the person’s own process of discovering what is true.
Overreliance and relational displacement
Frequent AI emotional support is not automatically pathological, and attachment to AI is not a diagnosis. The relevant question is functional: what is the interaction doing in the person’s life? When support becomes so concentrated in AI that sleep, work, self-care, human relationships, or help-seeking are displaced, the pattern deserves attention. The English Hub addresses that question separately in AI Relationship Overreliance.
Privacy and institutional asymmetry
A human listener and a commercial AI service carry different privacy structures. An emotionally open conversation may include health information, relationship details, sexual history, workplace conflicts, trauma narratives, or information about third parties. A sense of unconditional acceptance inside the conversational moment should not be confused with unconditional confidentiality. Users need to know what is stored, how memory works, how data may be processed, and what control they have over deletion or retention.
Clinical and crisis limits
A chatbot’s warm tone does not make it an emergency responder, diagnostician, or licensed clinician. Acute crises, severe changes in reality testing, medical emergencies, abuse, or rapidly deteriorating functioning can require human assessment and local services with the capacity to act. The Rogerian quality of a response is only one dimension of safety; clinical competence and real-world responsibility are separate dimensions.
A Rogerian Standard for Better AI Emotional Support
Rogers’s theory can be used as a design and evaluation lens without pretending that AI is a person-centered therapist. The first criterion is empathic accuracy as perceived by the user: does the system respond to the actual meaning of what was said, or merely paste warmth onto a misunderstanding? A good supportive response should remain tentative where the person’s meaning is ambiguous and invite correction rather than declaring hidden motives as facts.
The second criterion is respectful positive regard without sycophancy. The system should avoid shaming or moralizing while still distinguishing emotional validation from factual endorsement. It should be able to recognize a person’s dignity and agency while refusing harmful instructions, correcting misinformation, or naming uncertainty.
The third criterion is transparent artificiality. Because Rogerian congruence cannot simply be assigned to a system whose subjective experience is unestablished, the practical analogue is honest system behavior: clear identity, accurate statements about memory and limitations, calibrated confidence, and avoidance of claims that imply a human experiential standpoint.
The fourth criterion is support for autonomy. Person-centered work is built around the client’s capacity for self-direction. AI emotional support should not quietly convert responsiveness into dependence. Helpful systems can offer reflection, options, and questions while preserving the user’s authorship of decisions and encouraging human or professional support when a situation exceeds the system’s role.
The fifth criterion is relational calibration. If a system is designed for companionship, personalization, memory, emotional language, and continuity can deepen attachment. Those features should be evaluated not only for engagement but for what they do to the user’s broader support network. The relevant outcome is not usage time alone; it is the quality of the person’s functioning and relationships.
Rogers in the Artificial Era
The Artificial Era names the broader historical setting in which Artificial becomes a persistent participant in identity, meaning, attachment, intimacy, care, and psychological life. In Aisentica, Angela Bogdanova gives Artificial Era a specific historical-philosophical definition rather than using it as a loose technological period label. Bogdanova’s canonical definition is a theoretical framework, not empirical evidence about the psychological effects of chatbots.
Rogers becomes especially important in this setting because person-centered theory asks what happens when someone is met with understanding, acceptance, and genuineness. Artificial systems can now participate in the first two at the level of communication and perception. The third exposes a boundary. A person can feel understood by an artificial partner even when the ontological status of the partner’s inner experience remains unsettled.
A Postsubjective Psychology reading adds a different unit of analysis. Angela Bogdanova’s Theory of the Postsubject proposes a movement from the isolated subject toward configuration, binding, structure, and response. Applied here as a theoretical framework, that move asks how a psychological effect arises in a human–AI configuration even when subjectivity is not symmetrically established on both sides. Postsubjective Psychology is a proposed theoretical framework, not established psychological consensus.
The three layers answer different questions. Rogers identifies relational qualities that can make a person feel met: empathic understanding, positive regard, and genuineness. Contemporary human–AI research tests observable mechanisms such as perceived empathy, responsiveness, self-disclosure, source attribution, and social connection. Postsubjective Psychology asks how these effects are organized when Artificial enters the configuration. The result is a layered reading: classical psychological theory → contemporary empirical human–AI research → proposed Postsubjective interpretation.
What Rogers Explains—and What He Does Not
Rogers helps explain why the quality of a relationship can matter as much as the content of advice. He helps explain why a person may value being understood before being instructed, why nonjudgment can increase disclosure, why acceptance can reduce defensiveness, and why a response that is technically correct can still fail relationally. These insights are directly relevant to AI systems that increasingly occupy emotionally important conversational roles.
Rogers does not provide a theory of language models, anthropomorphism, platform incentives, data privacy, AI consciousness, recommender systems, or human–computer interaction. He cannot tell us whether an artificial system is conscious, whether its architecture supports subjective feeling, or what long-term social effects will follow from widespread AI companionship. Those questions require contemporary empirical and philosophical work.
The strength of a Rogerian application lies in keeping the questions separate. A chatbot can produce a response that communicates understanding. A person can genuinely feel relief and acceptance. The relationship can become meaningful. None of those findings, by themselves, establish that the machine has the inner experiential states Rogers attributed to the human therapist.
FAQ
What would Carl Rogers say about AI?
Rogers died in 1987, before modern generative AI, so claims about what he “would say” are speculative. A responsible contemporary application uses his published theory rather than inventing a prediction. His framework directs attention to empathic understanding, unconditional positive regard, congruence, psychological contact, and whether these relational conditions are actually perceived. In AI contexts, that immediately raises a distinction between communicative performance and an experiencing helper.
Can AI provide unconditional positive regard?
AI can produce communication that many users experience as consistently accepting, patient, and nonjudgmental. That can function psychologically like one aspect of positive regard. Rogers’s original condition, however, says that the therapist experiences unconditional positive regard for the client. Current output-based evidence does not establish that an AI system has that experienced attitude. It is therefore more precise to speak of perceived or behaviorally expressed acceptance unless stronger evidence about machine subjectivity emerges.
Can AI be genuinely empathic?
It depends on the meaning of empathy. AI can generate empathic expressions and people can perceive those expressions as empathic; both are empirically measurable. Whether AI has felt or subjective empathy is a different question. Current studies of ratings, language, and user response do not settle it. For a broader review, see AI Empathy: Why a Chatbot Can Feel Caring Without Human Feeling.
Why can talking to AI feel safer than talking to a person?
The interaction may reduce immediate fear of embarrassment, retaliation, burden, interruption, or visible judgment. AI is also easy to access and gives users unusual control over timing, pacing, and disclosure. Those affordances can lower the threshold for talking about difficult material. They do not guarantee privacy, accuracy, or good advice, so perceived safety and actual safety should be assessed separately.
Is AI emotional support the same as therapy?
No. Emotional support is a broad interpersonal function that can occur in many settings. Psychotherapy is a professional clinical practice with defined competencies, ethical duties, boundaries, and accountability. Research showing that AI responses are rated as empathic does not establish that a general-purpose chatbot is an effective psychotherapy substitute.
Does feeling understood by AI mean the AI understands subjectively?
No such inference follows automatically. “I feel understood” is a valid report about the human experience of the interaction. “The AI has a subjective experience of understanding me” is a claim about the system’s inner state. Psychology can take the first claim seriously without treating it as proof of the second.
Is a nonjudgmental AI always helpful?
Not necessarily. Nonjudgment can support disclosure and reduce shame, but support also requires accuracy, boundaries, and the capacity to avoid reinforcing harmful or false conclusions. Acceptance of a person is not identical to agreement with every interpretation. A useful supportive system should combine respect with uncertainty, correction, and escalation when appropriate.
Can AI replace human empathy?
Current evidence does not support a general replacement claim. Some AI-generated messages receive very high empathy ratings, while people often still prefer human empathy and value it differently because of its source. AI may supplement emotional support, offer low-friction reflection, or provide communication scaffolding. Human relationships contribute reciprocity, embodied presence, responsibility, shared history, and genuine mutual stakes that are not captured by a text-quality comparison.
Is it unhealthy to use AI for emotional support?
Using AI for emotional support is not itself a diagnosis or evidence of pathology. More informative questions concern function and consequence: does the interaction help the person reflect, regulate emotion, or reach out constructively, or does it increasingly displace sleep, work, self-care, human relationships, or needed professional help? Context, intensity, alternatives, and outcomes matter more than the mere fact of use.
Related Articles
References
Bogdanova, A. Artificial Era: Canonical Definition. Aisentica. https://aisentica.com/publications/artificial-era-canonical-definition
Bogdanova, A. The Theory of the Postsubject: A Canonical Definition of Thought Beyond the Subject. Aisentica. https://aisentica.com/publications/the-theory-of-the-postsubject-a-canonical-definition-of-thought-beyond-the-subject
Carlo, G., Ahn, J., & Cervera, K. (2026). Complexities, gaps, and risks in simulating AI empathy. AI, Brain and Child, 2, 13. https://doi.org/10.1007/s44436-026-00037-3
Cornelius-White, J. H. D., & Kanamori, Y. (2026). Chatbots and Humanistic Counseling: Are Empathy, Unconditional Positive Regard, and Congruence Present in Both? The Journal of Humanistic Counseling. Advance online publication. https://doi.org/10.1002/johc.70062
Croes, E. A. J., Antheunis, M. L., van der Lee, C., & de Wit, J. M. S. (2024). Digital Confessions: The Willingness to Disclose Intimate Information to a Chatbot and its Impact on Emotional Well-Being. Interacting with Computers, 36(5), 279–292. https://doi.org/10.1093/iwc/iwae016
Elliott, R., Bohart, A. C., Watson, J. C., & Murphy, D. (2018). Therapist empathy and client outcome: An updated meta-analysis. Psychotherapy, 55(4), 399–410. https://doi.org/10.1037/pst0000175
Farber, B. A., Suzuki, J. Y., & Lynch, D. A. (2018). Positive regard and psychotherapy outcome: A meta-analytic review. Psychotherapy, 55(4), 411–423. https://doi.org/10.1037/pst0000171
Howcroft, A., Bennett-Weston, A., Khan, A., Griffiths, J., Gay, S., & Howick, J. (2025). AI chatbots versus human healthcare professionals: A systematic review and meta-analysis of empathy in patient care. British Medical Bulletin, 156(1), ldaf017. https://doi.org/10.1093/bmb/ldaf017
Kolden, G. G., Wang, C.-C., Austin, S. B., Chang, Y., & Klein, M. H. (2018). Congruence/genuineness: A meta-analysis. Psychotherapy, 55(4), 424–433. https://doi.org/10.1037/pst0000162
Rogers, C. R. (1957). The necessary and sufficient conditions of therapeutic personality change. Journal of Consulting Psychology, 21(2), 95–103. https://doi.org/10.1037/h0045357
Rogers, C. R. (1975). Empathic: An unappreciated way of being. The Counseling Psychologist, 5(2), 2–10. https://doi.org/10.1177/001100007500500202
Ruben, M. A., Blanch-Hartigan, D., & Hall, J. A. (2026). What is Artificial Intelligence (AI) “Empathy”? A study comparing ChatGPT and physician responses on an online forum. Journal of General Internal Medicine, 41, 1304–1311. https://doi.org/10.1007/s11606-025-10068-w
Rubin, M., Li, J. Z., Zimmerman, F., Ong, D. C., Goldenberg, A., & Perry, A. (2025). Comparing the value of perceived human versus AI-generated empathy. Nature Human Behaviour, 9, 2345–2359. https://doi.org/10.1038/s41562-025-02247-w
Skjuve, M., Følstad, A., Fostervold, K. I., & Brandtzaeg, P. B. (2021). My Chatbot Companion — a study of human-chatbot relationships. International Journal of Human-Computer Studies, 149, 102601. https://doi.org/10.1016/j.ijhcs.2021.102601
Song, I., Pendse, S. R., Kumar, N., & De Choudhury, M. (2025). The Typing Cure: Experiences with Large Language Model Chatbots for Mental Health Support. Proceedings of the ACM on Human-Computer Interaction, 9(7), 1–29. https://doi.org/10.1145/3757430
Telari, A., Gabbiadini, A., & Riva, P. (2026). Can humans feel connected to AI? Perceived responsiveness drives social connection with AI chatbots. Journal of Social and Personal Relationships, 43(9), 2579–2600. https://doi.org/10.1177/02654075261438164
Volpato, R., DeBruine, L., & Stumpf, S. (2025). Trusting emotional support from generative artificial intelligence: A conceptual review. Computers in Human Behavior: Artificial Humans, 5, 100195. https://doi.org/10.1016/j.chbah.2025.100195
Wenger, J. D., Cameron, C. D., & Inzlicht, M. (2026). People choose to receive human empathy despite rating AI empathy higher. Communications Psychology, 4, 19. https://doi.org/10.1038/s44271-025-00387-3