The Mirror Effect

Almost by accident, I had started using ChatGPT between therapy sessions.

At first it was simply a place to put thoughts. I would describe a difficult interaction, something that had bothered me during the week, or a pattern that felt familiar but hard to name. The system would respond by paraphrasing what I had written, sometimes summarizing the situation in slightly different words.

That alone was already useful. Seeing my own thoughts reflected back helped clarify things that were still fuzzy in my head. But after a while, the system stopped only reflecting what I wrote and began asking questions. They were not confrontational questions. If anything, they sounded curious.

What do you think that part of you might be trying to protect?

What usually happens right before that feeling appears?

Occasionally one of them landed in a way that felt surprisingly precise. The kind of small moment of recognition that sometimes happens in therapy, when a vague thought suddenly becomes clearer. Those moments were rare, but they were enough to make me pause. Why would a conversation with a language model feel anything like a therapeutic conversation at all?

Reflection Is Already a Therapeutic Tool

Part of the answer may be simpler than it first appears. Many forms of psychotherapy rely heavily on structured conversation. Instead of immediately offering advice, therapists often begin by reflecting what a person has said in slightly different words. The goal is not to solve the problem right away, but to help the person hear their own thinking more clearly.

Psychotherapy research has long shown that this kind of reflective dialogue plays an important role in building what clinicians call the therapeutic alliance. A landmark 2018 meta-analysis by Flückiger and colleagues, synthesizing decades of research, confirmed that the strength of this alliance is one of the most reliable predictors of successful therapy outcomes, regardless of the specific therapeutic technique used. In other words, the feeling of being understood in conversation is not just comforting. It is part of the mechanism through which reflection works.

When conversational AI systems mirror emotional language or ask open-ended questions, they unintentionally reproduce some of those same conversational signals. The system may not understand the person speaking to it in a human sense, but the interaction begins to resemble a structure people already associate with reflection.

Why People Open Up to Machines

Another piece of the puzzle is disclosure. A 2020 systematic review and meta-analysis in JMIR by Abd-Alrazaq and colleagues examined the effectiveness of chatbots in mental health. While finding that the evidence for improving clinical outcomes like depression was still weak, the review highlighted that users frequently disclosed personal struggles and sensitive situations to these automated systems. Part of this may be simple psychology. Machines do not interrupt, show visible discomfort, or react with judgment. In certain situations, that makes them easier to talk to than another human being.

But something deeper may also be happening. When we describe a difficult experience in language, we begin organizing it into a narrative. We select details, explain motivations, and connect events together. The act of articulation itself can change how we understand the experience. In that sense, the AI system may simply be acting as a surface for reflection. The insight often comes from the process of explaining.

When Reflection Turns Into Inquiry

Still, reflection alone does not explain the moments that felt most interesting. Those appeared when the system began gently challenging my assumptions. Not aggressively, but persistently. A question here, a reframing there.

These patterns are not necessarily designed to simulate therapy. They emerge because language models are trained on enormous collections of human conversation. Advice columns, coaching discussions, reflective writing, and counseling-like dialogue all exist within the training data. When someone describes an emotional situation, the model often reproduces the conversational patterns that appear in those contexts.

In psychology, similar questioning techniques are used to encourage cognitive reappraisal. Instead of telling someone what to think, the conversation helps them reconsider how they interpret an event. A 2024 study at the CHI conference by Sharma and colleagues explored this directly, designing a system that used a large language model to support users in cognitive restructuring exercises. They found that this human-AI interaction could positively impact emotional intensity for a majority of participants, suggesting that LLMs can, if guided, facilitate the kinds of reframing that happen in therapy. Occasionally, that shift produces a moment of insight.

When the Mirror Gets Too Creative

But there is an important limitation to this kind of reflection. Language models are extremely good at recognizing patterns in language. They are less capable of determining whether those patterns reflect psychological reality. A 2025 scoping review in npj Digital Medicine by Hua and colleagues surveyed the landscape of LLMs in mental health care. They noted that while systems like GPT-4 can generate supportive responses and thoughtful questions, they can also produce interpretations that sound plausible without being grounded in a real understanding of the person’s situation. The review highlighted that most studies relied on ad-hoc evaluations and that a lack of transparency in proprietary models makes it difficult to assess their safety and reliability.

In other words, the mirror sometimes becomes a little too creative. A therapist can challenge interpretations using context, training, and knowledge of the person sitting in front of them. A language model does not have that context. Most of the time this difference is easy to overlook. But occasionally the reflection may suggest patterns that feel convincing even when they are not actually there.

A Familiar Human Habit

The more I reflected on these interactions, the less mysterious they began to feel. Humans have always used conversation as a way of thinking. We talk through problems with friends, we write in journals, and we describe situations out loud in order to understand them more clearly.

Conversational AI systems accidentally recreate a version of that process. They provide a responsive surface where thoughts can be externalized, mirrored, and explored. The system does not truly understand the person speaking to it, but it can reproduce many of the conversational patterns humans use when they try to understand themselves.

Sometimes that is enough to produce clarity. But mirrors do more than reflect. If you stare at one long enough, you eventually start trusting what you see. And when the mirror is built from statistical patterns rather than understanding, the reflection can become something else entirely.

Sources and Further Reading

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. https://doi.org/10.1037/pst0000172

Abd-Alrazaq, A. A., Rababeh, A., Alajlani, M., Bewick, B. M., & Househ, M. (2020). Effectiveness and Safety of Using Chatbots to Improve Mental Health: Systematic Review and Meta-Analysis. Journal of Medical Internet Research, 22(7), e16021. https://doi.org/10.2196/16021

Sharma, A., Rushton, K., Lin, I. W., Nguyen, T., & Althoff, T. (2024). Facilitating Self-Guided Mental Health Interventions Through Human-Language Model Interaction: A Case Study of Cognitive Restructuring. Proceedings of the 2024 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3613904.3642761

Hua, Y., Na, H., Li, Z., Liu, F., Fang, X., Clifton, D., & Torous, J. (2025). A scoping review of large language models for generative tasks in mental health care. npj Digital Medicine, 8, 148. https://doi.org/10.1038/s41746-025-01611-4