Clarity seekers5 min read

What AI can and can't tell you about your own patterns

An honest account of where machine reading of a journal is genuinely useful, where it is confidently wrong, and how to tell which one you are looking at.

We make a product that reads people's journals and tells them what it found. So take the following with the appropriate pinch of salt — but we would rather set expectations correctly than have you discover the limits on your own and conclude the whole category is nonsense.

Here is what this technology is actually good at, and what it is not.

What it genuinely does well

Counting things you cannot count. You cannot hold four hundred entries in your head. A machine can read all of them and report that one subject appears in thirty-one of them, clustered in two periods eight months apart. That is a real fact about your life that was previously unavailable to you, and it required no interpretation to produce.

Finding the thread when you changed vocabulary. This is the strongest case. In 2021 you wrote about "the situation at work". In 2024 you wrote about "whether I actually want this". You will never find both with a keyword search, because they share no words. Matching on meaning rather than wording surfaces connections that are genuinely hard to find by hand.

Noticing what is absent. Subjects that stop appearing are nearly invisible to memory — you do not notice an absence. A machine comparing two periods will say that something which dominated spring has not been mentioned since June, and that is frequently more interesting than anything that is present.

Summarising a period you have forgotten. Reading a month back and getting a paragraph of what was in it is a genuine saving, and the paragraph is drawn from your own sentences rather than invented.

Notice what these have in common: they are all retrieval and aggregation. The machine is reading more than you can and reporting what is there. It is not interpreting.

What it does badly

Causation. The model will tell you two things co-occurred. It has no idea whether one caused the other, and neither do you, and the sentence will sound equally confident either way. Your sleep and your mood move together — that is in the data. Which direction, and whether some third thing drives both, is not.

Knowing what it has not been told. A journal is a biased sample. It over-represents the days you felt like writing, which for most people means the hard days and the unusually good ones. Any reading of that archive inherits the bias. A system that says you were preoccupied with one thing all year may be accurately describing your journal while badly describing your year.

Resisting the pull toward a tidy story. These models are built to produce coherent, plausible text, and coherence is exactly the thing to be suspicious of here. Human lives are substantially noise. A reading that resolves fifteen years into three clean themes has almost certainly discarded the parts that did not fit. Anthropic publishes a good deal about how these systems behave and where they fail, and the short version relevant here is that fluency is not accuracy.

Telling you what to do. It does not know your constraints, the people involved, what you have already tried, or what you can afford to lose. An apparently insightful instruction from something that has read only your journal should be treated as a prompt for your own thinking, not as advice.

Weighting importance. A machine sees frequency. Frequency and significance are different. The thing you wrote about once, in four lines, at 3am, may be the most important entry in the archive, and it will be statistically invisible.

How to tell a real finding from a plausible one

A practical test: can you check it?

A finding like "this subject appears in thirty-one entries, here they are" is checkable. Open them. If they are about what it claims, it is right. This is the category to trust.

A finding like "you tend to withdraw when work gets difficult" is an interpretation. It may be true. It may also be a sentence that would sound equally plausible applied to half the population. Treat it as a hypothesis, go and read the actual entries, and decide for yourself.

The rule of thumb: trust the pointing, question the explaining. Anything that directs you to specific entries you can read is doing the useful job. Anything that summarises your character is doing the job you should do yourself.

This is why the useful question to ask of any tool in this category is not "what does it conclude" but "can I see what it read". Ember's themes open into the underlying entries for exactly this reason — the finding is only as good as your ability to check it, and seeing the patterns means seeing the material they came from.

The thing it cannot do at all

It does not know what any of it meant to you.

It can tell you that your father appears in forty entries across six years, that the tone shifted after a particular spring, and that you stopped writing about him for nine months. Every one of those is true and none of them is the thing. The thing is not in the archive, and no amount of reading will extract it.

What the machine does is get you to the material faster. The reading is still yours.

What this means for how you use it

  • Use it to find, not to conclude. Its job is to put the right forty entries in front of you.
  • Always open the entries. A finding you have not verified is a rumour about yourself.
  • Be sceptical of anything flattering or tidy. Both are signs of a model producing plausible text.
  • Keep the parts that do not fit. The anomalies are usually where the information is, and they are exactly what summarising destroys.
  • Do not outsource the judgement. Write your own entry about what you make of it. That entry will be worth more than the machine's paragraph, and in two years it will be the one you reread.

Why bother, then

Because the alternative is an archive nobody reads.

The realistic comparison is not between a machine reading your journal and you reading it carefully. It is between a machine reading it and nobody reading it, which is the status quo for almost everyone with more than a year of entries. Most people never go back, because doing it by hand is genuinely hard work with no obvious entry point.

Something that reads four hundred entries and says "here are the thirty-one about this, in order" is not insight. It is a table of contents for your own life, and that turns out to be the part that was missing.

If you have an archive sitting unread, bringing it somewhere that will read it is a more useful move than writing more into the same silence.

A journal that reads itself back

Ember reads across everything you write and surfaces what keeps returning. Bring an old archive with you, or start tonight.

Keep reading

  • Clarity seekers

    Journal prompts for self-discovery

    Most prompt lists produce pleasant, useless entries. These are built to get at something you did not already know, with notes on why each one works.

    4 min read