How AI predicts your baby’s next nap

Every baby tracker now says “AI” somewhere on its store listing, and almost none of them say what it does. That vagueness is worth resisting, because the useful version of this feature is easy to explain — and knowing how it works tells you exactly when to trust it and when not to.

This is a description of how Pixy’s prediction works. Other apps do it differently, and some do not really do it at all.

What the prediction is actually doing

Every time you log a sleep, you also log a wake window without meaning to: the gap between the last wake-up and this sleep. After a few days you have a small dataset — your child’s actual awake stretches, not a book’s.

The prediction is three steps:

• Collect. Pull the recent wake windows out of the log, weighted towards the last few days rather than the last few months.
• Find the middle. Take the median of those windows.
• Project. Add it to the time your child last woke, and show the result as a range rather than a single minute.

That is it. There is no neural network deciding when your baby is sleepy, and there does not need to be — the interesting information is already in your log, it is just tedious to work out at 6am with a baby on your shoulder.

Why an age chart stops being enough

Published wake window charts are averages across a population. They are genuinely useful for the first days with a new baby, when you have no data of your own, and they get less useful every week after that.

The two are complements, not rivals. A good prediction uses the age band as a floor and moves onto your data as soon as there is enough of it — which is precisely what Pixy does. Until there is enough history, age-based guidance holds.

Median, not average

This sounds like a technicality and is the difference between a prediction that works and one that annoys you.

Suppose your baby’s last five wake windows were 2h00, 2h10, 1h55, 2h05 and 4h30 — because on that last one you were at a family lunch and the nap never happened.

• The average is 2h32, which is longer than four of the five actual windows. One unusual day has dragged every future prediction late.
• The median is 2h05, which is what your baby actually does.

Real parenting data is full of outliers — the car nap, the holiday, the day with the cousins. The median ignores them; the average lets each one distort the next week.

How much history it needs first

A few days of reasonably complete logging is usually the point where your own pattern starts beating the age default. There is no threshold at which it becomes finished — it keeps sharpening as long as you keep logging.

Three things temporarily degrade it, and all three are expected:

• A nap transition. The windows are actively lengthening, so recent history understates what your child now needs.
• A developmental leap. Windows shorten for a fortnight, then snap back.
• Gaps in the log. Missing sleeps make the surrounding windows look longer than they were.

The last one is worth saying plainly: a prediction built on partial logging is a prediction built on wrong numbers. If you have had a few chaotic days, expect it to be off until the data catches up.

How to read a predicted window

The mistake is treating it as an instruction. It is a prompt — the moment to stop what you are doing and start watching your baby.

• Start the wind-down before the window opens, not when it does. If the wind-down takes fifteen minutes, begin fifteen minutes early.
• Cues beat the app. If your baby shows early tired signs twenty minutes before the predicted window, go. The prediction describes the average day, not this one.
• A missed window is not a failed day. Shift everything later and carry on; the next prediction recalculates from the actual wake-up.

What it cannot do

Being specific about the limits is more useful than any feature list:

• It cannot see your child. Teething, a cold, a hot room, an exciting morning — none of that is in the log, and all of it changes the answer.
• It cannot diagnose anything. A tracking app is not a medical device. Snoring, breathing pauses, extreme sleepiness or a lost skill are questions for a paediatrician, and no app should imply otherwise.
• It cannot predict a first. The day your child drops a nap, the model is still describing the child they were last week.
• It cannot fix a sleep problem. Knowing when the window falls is not the same as your child being able to settle in it.

What “AI” usually means on an app listing

Since you will be comparing apps, it helps to know what the word is standing in for. In this category it is normally one of four quite different things:

1. Statistics on your own data

Medians, trends, projections. This is what sleep prediction is. It is genuinely useful and it is not what most people picture when they read “AI”.

2. A language model doing a task

Generating a recipe or a bedtime story, or answering a question. Real AI in the current sense, and the part that usually sits behind a subscription because each request costs the developer money.

3. Parsing what you typed

Turning “she slept two hours” into a structured log entry. Modest, and probably the feature that saves the most time in daily use.

4. Nothing in particular

A rules table with a marketing label on it. Not necessarily bad — a good rules table is fine — but you should know which one you are paying for.

A fair test when comparing: can the listing tell you which of these four it means? If it cannot, that is information too.

Your data and the AI features

These mechanisms have genuinely different privacy shapes, and they are worth separating rather than lumping under one word:

• The sleep prediction is a calculation over the logs in your own account.
• The AI assistants send what you ask them to a third-party model — Google Gemini, as set out in our privacy policy — because generating a story or a recipe requires one.

Whichever app you choose, that is the distinction worth asking about: which features stay with your own data, and which send something out. Pixy does not sell user data, and you can delete your account and everything in it from inside the app.

When to ignore the prediction

Regularly, and without guilt. The prediction is describing your child’s recent average. Ignore it when your baby is showing clear tired cues early, when they are ill or teething, during a nap transition, on a day that broke the routine, or when your own read of the situation disagrees with it.

A prediction that you override half the time is still doing its job — it has told you when to start paying attention. That is the whole ambition, and any app promising more than that is promising something the data cannot deliver.