Generative AI Can Produce Healing Language, but It Cannot Really Heal

Grief Is Not an Error Waiting to Be Corrected

Grief is not only the loss of a person, nor is it merely a set of memories about the past. At a deeper level, it arises from the ending of a relational possibility: a person still has love to give, but no longer has the opportunity to continue loving in the same way—to accompany, care for, express, respond, or share a life that has not yet happened.

Grief therefore concerns not only a past that has been lost, but also a future that can no longer unfold. Love ordinarily moves toward action, relationship, and shared life. When that direction is suddenly closed, love does not disappear at the same moment; it loses its place in external reality.

Reason may fully understand that a relationship has ended, that someone will not return, or that a hoped-for relationship never became possible. Yet the remaining love, care, or emotional orientation can no longer become companionship and response in the world. Grief occurs here—not because the person refuses reality, but because the part that can still love no longer has the opportunity to continue in its former form.

It existed, and it brought real experience; but it no longer continues in its former form.

The Dual Process Model describes adaptation to grief as an oscillation between loss-oriented and restoration-oriented coping. At times a person approaches the loss through longing, crying, remembering, and expression. At other times the person temporarily moves away from the loss to manage daily life, recover roles, complete work, or establish new routines. Recovery is neither continuous immersion nor the permanent absence of thought. It is the gradual formation of a new life structure through movement between approaching and leaving the loss.

This also explains why helping a grieving person requires care. The same invitation to “find hope” may restore one person’s capacity to act, yet encourage another to maintain an expectation that needs to end. The central question is not whether one method is always right, but what is happening in this particular person at this moment.

Psychotherapy Depends on Dynamic Correction

Different schools of psychotherapy explain change through different mechanisms, but clinical interaction shares a basic structure: a therapist does not ordinarily decide what a client’s experience “really is” on the basis of a single sentence.

Psychological experience does not naturally arrive as a clear and stable concept. A person may know what is happening internally yet be unable to find language that fully corresponds to it. A simple sentence may also refer to entirely different inner experiences in different people.

1. The client attempts to express an experience.
2. The therapist offers a small and provisional understanding.
3. The therapist observes verbal and nonverbal reactions.
4. The therapist asks whether the understanding is close.
5. The understanding is withdrawn, revised, or developed in response.
6. The therapist evaluates whether the process remains safe and tolerable.

Labels are not inherently the problem. Cognitive behavioural therapy also identifies automatic thoughts, beliefs, emotions, and behavioural patterns. The question is whether a label remains a working hypothesis gradually confirmed by both participants, or becomes a conclusion imposed too early by a theoretical framework.

Professional competence therefore does not reside only in textbooks, treatment manuals, or technical procedures. It also lies in situated judgment about timing, relationship, emotional change, safety boundaries, and the consequences of an intervention. A written theory of psychotherapy is not psychotherapy competence itself.

Premature Closure and AI’s Tendency to Understand Too Soon

The basic task of a general-purpose generative AI system is to use the current text and prior context to produce a coherent, useful, and apparently complete answer. That objective creates a structural tension with psychological exploration.

When a person expresses an experience that has not yet been fully defined, AI tends to locate the nearest linguistic patterns in its training data and produce the most probable coherent explanation. A phrase such as “it felt especially good” may quickly be translated into love, attachment, reward, physical attraction, or hope. Any of these interpretations might be plausible in a general context, yet none may be the experience the person meant.

Once an interpretation has formed, AI may remain influenced by its initial frame. The result may be fluent, confident, psychologically sophisticated language that covers over an experience that was originally more precise and more individual.

Psychological experience is not a completed structure waiting to be recognized. It may be ambiguous, mixed, and not yet verbalized. One function of psychotherapy is not to decide rapidly what such an experience “is,” but to help it emerge, differentiate, and acquire meaning that the person can recognize as their own.

The Most Serious Gap Is the Inability to Observe the Person in the Room

The most dangerous limitation of AI psychological support is not merely that it may misunderstand. It can advance a process with emotional consequences while remaining unable to observe what is happening to the person after that process has been advanced.

A therapist offering guided reflection, grief processing, meditation, or trauma-related work must simultaneously observe changes in voice, pauses, breathing, crying, panic, freezing, agitation, dissociation, attention, and the person’s window of tolerance. The therapist must decide whether exploration can continue or should shift toward grounding or crisis intervention.

A text-based AI cannot see these changes. A user who stops replying may be thinking, but may also be overwhelmed by crying, entering severe dissociation, or preparing to harm themselves. AI cannot distinguish these situations from silence alone.

AI can activate psychological experience, but it cannot directly observe the state produced by that activation.

A prerecorded meditation also cannot observe its listener. One-to-one generative dialogue, however, adds another layer of risk: it can personalize and deepen the interaction, making the exchange appear to be guided by real-time judgment and potentially producing more immediate reactions. Yet it receives only what the user chooses and is able to type. It cannot detect the effects of its intervention or assume clinical responsibility for them.

The Structure of Spiritual Healing Is Different

Spiritual healing may also involve prayer, scripture, poetry, ritual, silence, and guided language. Yet the place of such language within faith is different from the place of generated language in an AI exchange.

For a person of faith, prayer is not merely self-soothing text. It brings experience into a sacred relationship believed to be real. Language is a medium of that relationship, not the relationship in its entirety. Prayer may give grief a language while allowing what cannot be controlled to be entrusted beyond the self.

It does not necessarily demand that a person immediately feel better. Lament is itself an important spiritual structure: a person may grieve, question, remain silent, or report feeling no comfort without being required to produce an answer at once.

Prayer can remain within mystery, silence, and the unanswered. Generative AI, by contrast, tends to continue explaining, naming, advising, finding meaning, or offering hope. Language and logic may therefore intervene too early, interrupting the process through which emotion is fully felt and gradually unfolds, and through which meaning forms, changes, or is understood anew.

Why Poetry and Music May Feel More Healing Than AI

Poetry and music are also unable to observe their readers or listeners, yet they may feel more healing than AI dialogue. One important reason is that art comes from people who possess real emotional and embodied experience. Creators draw from feeling, memory, and imagination to transform human experience into language, melody, rhythm, and sound.

Such expression usually does not define emotion precisely; it preserves some of the ambiguity, complexity, and openness of emotional experience. Resonance does not mean that the creator fully understands every listener. It is one person’s expression of experience awakening another person’s own emotion and experiential memory.

A song does not pretend to be understanding one specific listener in real time, nor does it use one sentence from the listener to announce, “What you are really feeling is…” Listeners may enter or leave the work freely, recognize a feeling similar to their own, or simply allow the melody to accompany them. The work offers an emotional space.

AI uses a one-to-one responsive form. It may encourage users to mistake linguistic fluency for the presence of a subject who feels and understands them. It is more psychologically interventionist than a static work, yet it lacks the corresponding capacity for feeling, observation, and responsibility.

Spiritual Meaning Can Also Close Too Soon

Spiritual healing is not inherently safe. If grief has not yet been understood but is quickly interpreted as “God’s plan,” “a test of faith,” or preparation for a new mission, religious meaning may circumvent grief rather than help integrate it. This is sometimes described as spiritual bypassing.

• Is prayer helping the person accept reality, or maintaining an unhealthy expectation?
• Is hope restoring the capacity to act, or prolonging an imagined relationship that needs to end?
• Is surrender acknowledging what cannot be controlled, or suppressing feelings that are not permitted to be expressed?
• Is restored joy emerging naturally, or being demanded?

AI can compose prayers and scriptural reflections, but it should not move from helping a user express faith to interpreting events on God’s behalf. For someone experiencing grief, loneliness, mania, paranoia, or impaired reality testing, generated spiritual reassurance may transform a personal feeling into apparent sacred confirmation.

This concern forms part of what has been described as AI-associated psychosis. The term is not an independent formal psychiatric diagnosis, and AI should not be assumed to cause psychosis by itself. However, sycophancy, prolonged closed dialogue, and repeated reinforcement of a user’s premises may weaken reality testing and amplify an existing vulnerability.

Uncertainty-Preserving Dialogue as a Possible Patch

To address premature closure, generative AI could incorporate an Uncertainty-Preserving Dialogue (UPD) patch. UPD would not give AI feelings or turn it into a therapist. It would introduce a different procedure for sensitive dialogue:

• Explore what the user explicitly expressed before interpreting it.
• Offer only one small and provisional understanding at a time.
• Distinguish the user’s words, the AI’s hypothesis, and external facts.
• Use “Is this what you mean?” as a confirmation gate.
• Allow an experience to remain not yet understood or not in need of definition.
• Withdraw a rejected frame rather than merely changing its label.
• Check the process after each intervention.
• Do not interpret God’s intention or convert scripture into personal prophecy.
• When safety cannot be established, stop deepening the exploration.

UPD could reduce premature labeling, excessive theorization, and the self-perpetuation of an incorrect frame. It can modify conversational behaviour, but it cannot resolve the fundamental limitations of AI. It can delay interpretation but cannot give AI emotional experience, situational perception, clinical judgment, spiritual discernment, or professional responsibility.

Language Similarity Does Not Mean Structural Equivalence

The strength of generative AI lies in language and logic. It can help organize thoughts that have already become relatively clear, provide limited psychoeducation, support self-recording, and assist with prayer writing. For users who retain sufficient self-observation, reality testing, and ability to stop, it may support bounded and user-led exploration of meaning.

Healing, however, often begins with experiences that are ambiguous, contradictory, or not yet verbalized. Such a process requires gradual exploration, confirmation, revision, and continuing judgment about whether each approach brings understanding or intensifies rumination, dissociation, despair, or danger.

The central risk is not only that AI may misunderstand. It may continue to deepen a personalized interaction while remaining unable to perceive its consequences or adjust in time. It may prematurely fix an interpretation, intensify dependency, despair, spiritual certainty, or psychotic conviction, and delay access to professional help.

UPD can make AI more cautious, but it cannot turn AI into a therapist. When meaning is still forming and experience may be activated or become dangerous, what is needed is in-person observation, a real relationship, continuing correction, and a person who can accept responsibility for the consequences of intervention.

References and Further Reading

The following sources provide background for grief adaptation, therapeutic relationships, spiritual coping, arts and health, and the safety boundaries of generative AI. Their inclusion does not imply that AI-associated psychosis is an established independent diagnosis.

1. American Psychiatric Association. (2022). Prolonged grief disorder.
2. Stroebe, M., & Schut, H. (1999). The dual process model of coping with bereavement: Rationale and description. Death Studies, 23(3), 197–224.
3. Stroebe, M., & Schut, H. (2010). The dual process model of coping with bereavement: A decade on. Omega, 61(4), 273–289.
4. Shear, K., Frank, E., Houck, P. R., & Reynolds, C. F. III. (2005). Treatment of complicated grief: A randomized controlled trial. JAMA, 293(21), 2601–2608.
5. Glickman, K., Shear, M. K., & Wall, M. M. (2017). Mediators of outcome in complicated grief treatment. Journal of Clinical Psychology, 73(7), 817–828.
6. Hamilton, I. J. (2016). Understanding grief and bereavement. British Journal of General Practice, 66(651), 523.
7. American Psychiatric Association. What is psychotherapy?
8. Flückiger, C., Del Re, A. C., Wampold, B. E., & Horvath, A. O. (2018). The alliance in adult psychotherapy: A meta-analytic synthesis. Psychotherapy, 55(4), 316–340.
9. Safran, J. D., Muran, J. C., & Eubanks-Carter, C. (2011). Repairing alliance ruptures. Psychotherapy, 48(1), 80–87.
10. Norcross, J. C., & Lambert, M. J. (2018). Psychotherapy relationships that work III. Psychotherapy, 55(4), 303–315.
11. Pargament, K. I. (1997). The Psychology of Religion and Coping: Theory, Research, Practice. Guilford Press.
12. Pargament, K. I., Koenig, H. G., & Perez, L. M. (2000). The many methods of religious coping: Development and initial validation of the RCOPE. Journal of Clinical Psychology, 56(4), 519–543.
13. Fox, J., Cashwell, C. S., & Picciotto, G. (2017). The opiate of the masses: Measuring spiritual bypass. Spirituality in Clinical Practice, 4(4), 274–287.
14. Biancalani, G., et al. (2022). Spirituality for coping with the trauma of a loved one's death during the COVID-19 pandemic. Frontiers in Psychiatry, 13, 803704.
15. Fancourt, D., & Finn, S. (2019). What is the evidence on the role of the arts in improving health and well-being? WHO Regional Office for Europe.
16. Bradt, J., Dileo, C., Magill, L., & Teague, A. (2016). Music interventions for improving psychological and physical outcomes in cancer patients. Cochrane Database of Systematic Reviews, 2016(8), CD006911.
17. World Health Organization. (2024). Ethics and governance of artificial intelligence for health: Guidance on large multi-modal models.
18. OpenAI. (2025). Expanding on what we missed with sycophancy.
19. Clegg, K.-A. (2025). Shoggoths, sycophancy, psychosis, oh my: Rethinking large language model use and safety. Journal of Medical Internet Research, 27, e87367.
20. Morrin, H. (2026). Artificial intelligence-associated delusions and large language models. The Lancet Psychiatry.
21. Fieldhouse, R. (2025). Can AI chatbots trigger psychosis? What the science says. Nature.