Prediction
Using what you know to guess at a value you don't
- A prediction is saying what a value you don't yet know might be, using the information you have right now. It isn't knowing the answer in advance and calling it.
- An AI's prediction never lands as one clean answer. It comes with how likely each option is.
- It isn't only about the future. Anything that's not yet confirmed counts as something to predict, even if it already happened.
- Whether a prediction is any good can only be judged after the real answer comes in and you check it against reality. Getting it right once doesn't make it a good prediction.
- Hand it a situation it's never seen and a prediction can swing wildly — and the model still hands over its answer with a perfectly confident face.
Contents
1The analogy
In the morning, the weather service says "70% chance of rain this afternoon." It didn't travel forward and peek at the afternoon. It looked at the record of pressure, wind, and clouds built up so far, and worked out how much it rained the other times the pattern looked like this. Prediction is filling in a spot that doesn't have an answer yet, using whatever information is on hand right now, exactly this way.
A forecast doesn't flatly say "it will rain." It comes out as a percentage. And if it doesn't rain, the forecast wasn't automatically wrong — the 30% side is just what happened.
Whether a forecast was good or bad only becomes clear once the day is over. You have to write down what the weather actually was, lay it next to the forecast, and compare the two over a long stretch before the verdict comes in.
2In detail
Prediction is building a case, not guessing right
People usually ask "does it guess right?" about a prediction first. But from the predictor's side, the order is backwards. Stating an answer when you already know it isn't prediction — that's just confirmation. Prediction is building the most plausible value you can from only what's in your hands right now, while the answer is still unknown.
So a prediction always carries two parts: what it was based on, and how solid that basis is. A forecast built from pressure readings and cloud photos and one built from nothing but yesterday's weather can both land on the same 70%, but they don't carry the same weight.
What an AI does has this same shape. It pulls out whatever rule was left behind in the mountain of cases it studied, lays that rule over a fresh input it's never seen, and produces a value. How well that rule was drawn out decides how good the prediction is.
Prediction comes out as probability
Inside a model, a prediction is mostly built as a score handed out across several candidates — rain tomorrow: 70, no rain: 30, say. The single word "rain" that shows up on your screen is just the highest-scoring candidate, picked out and displayed alone.
Look at the score alongside it and the judgment changes. A prediction split 70 and 30, and one split 51 and 49, both say "rain," but how much you should trust them is completely different. The second one really means the model is unsure too.
That's why, wherever a prediction gets used, it helps to keep the score alongside the chosen answer instead of taking the answer alone. A low score can be the trigger for a person to take a second look, or for the system to hold off on answering at all.
It isn't only the future being predicted
The word prediction pulls your mind toward the future, but it has nothing to do with time. Anything not yet confirmed is something to predict — what an object in a photo is, which department a document belonged to last month, what value used to sit in a blank cell. All of it counts.
In weather terms, filling in yesterday's temperature for a mountain village with no weather station, using readings from stations nearby, is a prediction too. Yesterday has already passed, but nobody ever measured that village's temperature.
Almost everything an AI service hands back is a prediction in this sense. A line of translated text, a recommended list, a spam flag — all of it is "a value that wasn't confirmed, filled in from data."
How good it is only shows up later
A prediction's grade can only be given after the answer is revealed. You have to write the forecast and the actual weather down side by side, and gather not just one day but many months of them before you can tell whether this forecast is any good. Being right once doesn't make it a good forecast, and being wrong once doesn't make it a bad one.
If you gather up every day the forecast said 70% chance of rain, and it actually rained on roughly seven days out of ten, that forecast is being honest with its numbers. Without this process of checking a prediction against reality, a prediction just stays a plausible-sounding statement.
It's weak against weather it hasn't lived through
Prediction leans on past records. So the moment something shows up with no shape in those past records, it loses its footing. That's why a forecast struggles hardest on the day the heaviest rain on record falls — there's no case to lean on, so it's forced to reach for whatever old case looks closest.
An AI model does the exact same thing. It performs astonishingly well inside the range it saw in training, then produces a strange value the moment an input falls outside that range. What's worse is that the score can still come back high even then. A confident voice and being correct are two separate things.
3More precisely
Prediction refers to the output you get from feeding a new input into a trained model. The process of obtaining that output is called inference, and the result is the prediction. In classification, out comes how likely each category is; in regression, out comes a single number on a continuous scale. Either way, the model hands over a score, not certainty, and a person decides whether to act on that score.
The analogy breaks down in places too. A weather service's forecast is calculated from the physical laws of the atmosphere, but most AI predictions have no law behind them at all. They follow only the relationships left in the data, without knowing why. That's why they can't tell the difference between two things moving together and one thing causing the other. A record showing more drowning accidents on days ice cream sold well can still produce a prediction, but selling less ice cream won't cut down on accidents.
There's a second gap too. A forecaster can point to the exact pressure system behind tomorrow's rain, but pulling the same kind of clear reason out of a model's prediction is often much harder, even for the people who built it. The prediction arrives; the story behind it has to be reconstructed separately, and sometimes it can't be.
4Try it yourself
5Common misconceptions
It's easy to think a prediction is about the future, but actually anything not yet confirmed counts, even something that already happened.
It's easy to think a confident-sounding model can be trusted, but actually an input it's never seen in training can produce a high score with no real grounding behind it.
It's easy to think getting it right once means it's a good prediction, but actually you can only tell if a prediction is any good by gathering many of them and checking against what actually happened.
7One-line summary
In shortPrediction is stating a value you don't yet know, together with a probability, using whatever information you have now — and whether it's any good only shows up once the real answer is revealed.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02