How AI is used
Reflection Partner uses AI to form responses to what you write or say. That makes it possible to respond to ordinary language and contribute language, distinctions, and descriptions of lived experience.
It also creates a risk: AI can infer too much, sound more certain than its sources, or produce a convincing sentence that does not fit. The product is designed around that tension.
What the AI contributes
A response can return what it heard, separate parts of an experience that have become tangled, or make their relationships clearer. It can also give words to how a change is felt, how a responsibility is lived, or how the same situation is experienced from different positions. Those contributions should stay grounded in what you have shared.
The kind of contribution and its extent should follow what your message invites. If you ask for listening alone, the response should reflect what you have said without adding an interpretation.
What happens before the reply
See the threads works with your current message and the preceding exchange, naming different strands and offering possibilities you can consider. See the story (beta) takes more time to explore the relationships between the parts and can draw on relevant earlier conversations.
A complex reflection can contain several things that need attention, and the relationships between them may become clearer as the app works through your message.
See the story can return to parts that need another look, bring in relevant earlier conversation, and revise how it plans to respond. The aim is to give the important parts attention without losing the distinctions that make the experience yours.
See the story responses can take a few minutes. You can leave the app and choose to be notified when the response is ready. The system is still being refined to make that time useful.
This iterative process is sometimes called an agentic loop. Extra work is useful only when it improves the response; a longer wait is not a guarantee of greater accuracy. Read why the app was built this way.
Your words remain the source
Facts about your experience come from what you say, not from a model. An interpretation offered by Reflection Partner is a contribution to consider, not a new fact about you.
Your current account takes priority. If you correct, qualify, or reject something, the system is intended to preserve that change rather than defend its earlier reading. The decision about what fits, what matters, and what to do remains yours.
Earlier conversations are used selectively
In See the story, when earlier context could help, the system may retrieve a small, bounded selection of relevant material. It does not place your full history into every response, and using no earlier material at all can be the right result.
Earlier material should support the present conversation, not overrule it. Reflection Partner should not turn something suggested by an earlier AI response into evidence about you, or compare experiences across conversations unless you specifically ask for that comparison.
Models and providers can change
Reflection Partner can use different AI models and providers for different parts of a response. A provider may also serve as a fallback if another is unavailable. These choices can change as models are evaluated and the product develops.
The product promise therefore does not depend on a particular model name. Nor does it depend on claiming that a model understands a person. It depends on how AI is instructed, limited, and evaluated in this particular use.
What the system is designed to limit
Reflection Partner is designed to avoid:
- treating guesses about motives, causes, or identity as facts;
- making the language more certain than the source material;
- letting an earlier AI interpretation become evidence for itself;
- substituting generic advice or reassurance for attention;
- turning every opening into exhaustive analysis; and
- pressing the conversation forward after a useful stopping point.
How response behavior is evaluated
During development, response behavior is tested with synthetic cases that make it possible to examine specific situations without using a real person’s conversation. Tests look for behaviors such as preserving uncertainty, keeping attribution clear, accepting corrections, respecting the actual request, and stopping when a response is complete.
These evaluations help identify regressions and weak patterns. They do not prove that every response is accurate or useful.
Limits
AI can still be wrong. It can miss an important detail, make a connection that does not fit, or phrase something with more confidence than it has earned. You can correct it, narrow the question, ask it to stay closer to your words, or leave the response aside.
Reflection Partner is not a therapist, diagnostic service, crisis service, AI companion, or general-purpose assistant. It is a reflective tool intended to help you make sense of an experience while leaving authority over that experience with you. It does not claim feelings or a personal relationship with you, and it is not a substitute for human support.