AI journaling works by passing what you have written to a language model, which recognises patterns in that text and produces a likely, relevant response — usually a short reflection and a next question. There is no understanding of you underneath it; there is very good pattern recognition over language.
This page explains the mechanics conceptually, without jargon and without dressing them up as something more mysterious than they are. For the wider picture of what these tools are for, see the pillar on AI journaling and AI reflection.
Step one: you write
Everything begins with your text. The tool has nothing else to work from — no reading of your face, your tone, your pauses, or your history beyond what has been written down. This is why entries that are a little fuller tend to produce more useful replies: there is simply more signal.
Step two: the text is sent to a language model
A language model is a system trained on very large amounts of text to predict what plausibly comes next. It is not a database of answers and it is not looking you up anywhere. Alongside your entry it receives an instruction — a set of rules about how to respond — which is what makes a reflective tool behave differently from a general chatbot even when the underlying model is similar.
Step three: pattern recognition, not comprehension
The model identifies structure in what you wrote: the situation described, the emotional register of the language, what is emphasised, what recurs, what sits in tension. It does this statistically, from the way such language usually behaves, not by knowing what your week was like.
This is the crux of the whole thing. Recognising that a passage carries grief is different from grieving with you. It is a genuine and useful capability, and it is narrower than it feels. We look at that distinction directly in can AI understand what you are feeling.
Step four: generating a next question
The reply is generated word by word, each choice conditioned on everything before it. A well-designed reflective tool constrains that generation heavily: stay in the person's own words, do not interpret as fact, offer one question rather than three, do not advise. The question you receive is a plausible next question given your text — not a chosen intervention, and not the product of a plan for you.
What it retains — and what it does not
A language model does not remember you between calls. Each request is answered from what is sent with it. Any sense of continuity comes from the surrounding product deciding what to include — typically the entry you are working on and the earlier turns of that same conversation.
Stated plainly for Echo Journey: a reflection is generated from the entry in front of it and the turns of that conversation. When Echo reflects on an entry, it does not pull your other, unrelated entries into that request; the only wider context is anonymised, community-level themes with no link to any individual. Weekly Insight is the deliberate exception: if you use it, it looks back across your own recent week of entries, check-ins and notes so it can summarise themes for you. It uses only your own data, and it runs only when you ask for it. Your entries stay private to your account unless you deliberately share a specific entry with a practitioner you are connected to.
Honest limitations
- Fluency is not accuracy. A reply can be well-written and still miss what is actually happening for you.
- It follows your framing. If you describe a situation one way, it will usually work inside that description rather than challenge it.
- It has no clinical judgement and cannot assess risk. It is not therapy and does not treat or diagnose anything.
- It can be repetitive. Pattern-based systems drift towards familiar shapes of question.
- It cannot tell how much something matters. Significance is yours to assign.
Why the plain version is worth knowing
Knowing the mechanics changes how you use the tool. If you know the reply is a plausible continuation rather than a considered judgement about you, you take the useful questions and leave the ones that miss — which is exactly the right relationship to have with it.
If you would rather try it than read about it, you can talk to Echo and see how a reflection actually reads.
This article is general information. It is not therapy, medical advice, or a substitute for professional support. If you are in crisis, contact emergency services on 999 in the UK or call the Samaritans on 116 123.