Look at one number: Average score: 72. Clean, confident, and true — the arithmetic checks out. A teacher reads it and moves on, satisfied. Nothing about that single figure looks incomplete.
Now widen the frame. Behind the 72 sits a full spreadsheet: a handful of students near 95, most clustered close to 50, a few below 40. The average did not lie — it simply left almost everyone's actual score out of the picture.
That gap between the number and the spreadsheet behind it is where this chapter lives. Every statistical summary is a crop — a decision about what stays in the frame and what gets left out. Statistics for AI is the skill of telling an honest crop from one that hides the part that matters.
Suppose you had to build an AI system from nothing. What would you need to have figured out first? You would need a way to read patterns in data and trust which ones are real — that is statistics, this chapter's subject.
You would also need a way to turn a photograph, a sentence, or a sound clip into a grid of numbers. A computer can then actually compute with that grid — this is linear algebra. And you would need a way to say how confident a prediction is, which is probability, taken up in Chapter 9.
One branch remains: calculus, which improves a model step by step as it trains. Of the four, the one that pays off first is the one you can use the moment you read a claim. A student who can question a statistical claim already holds the most immediately useful branch.
Strip statistics down to what it actually does to data, and it is always four moves, in this order. Collect the numbers, then organise them so they can be compared. Analyse them for a pattern, then interpret the pattern into a conclusion you can act on.
Skip a step and the chain breaks. Collect without organising, and a pattern stays invisible in the mess. Analyse without interpreting, and you are left holding a number with no meaning attached to it. That is the exact gap this chapter keeps returning to.
AI already leans on statistics in places you would not first guess. In disaster management, flood-risk models predict which zones are dangerous. A single risk score can still hide how unevenly that risk actually falls across districts.
In sport, player analytics report batting averages and goals-per-game. Which single stat is the honest one depends entirely on the question being asked. In disease forecasting, a national "average incidence" figure can sit calmly over a regional emergency it never names.
Weather forecasting needs both halves at once: the centre and the range. A forecast that reports only the mean temperature fails the farmer who plans around it. Four domains, one same mistake — reporting one number without naming what it hides. The next three sections open exactly that crop wider, starting with the most familiar summary of all — the average.
Ask which "average" is the right one, and most people reach for the mean without thinking. There are three of them for a reason. They are not three versions of one tool — they are three different crops of the same data.
The mean keeps every point in the frame, weighing each one equally. That is exactly why one extreme value can pull it away from where most of the data actually sits. The median takes the middle of the ordered list instead, immune to extremes but blind to how far apart the values are.
The mode simply names the most common value, and says nothing about the rest of the distribution. Unlike the other two, it need not even be one number. A dataset can carry more than one tied mode, or none at all, when nothing repeats.
Each is honest for a different question, and misleading for the wrong one. Before you report a summary, name which crop you chose, and why — that choice is itself the claim.
Weaker. Dhruv adds up nine scores — 50, 52, 54, 55, 56, 58, 60, 64, 100 — and divides by nine: 61. "That's the class average," he says, reading off the one number. Every point sat inside that sum — including the single 100 doing most of the pulling.
Stronger. Nila orders the same nine scores and points at the middle one: 56. The single 100 drags the mean to its right — up to 61 — while the other eight, left on their own, average 56.1: almost exactly the median. For "what did a typical student score", 56 is the honest crop; 61 is dragged rightward by the one result that none of the other eight students came near.
"The mean is always the best average — just add and divide."
The mean is one crop, not the best one. Give it a class of five with one score of 99 and the rest at 50, and the mean is 59.8 — dragged well above where four of the five actually sit. Give it categorical data — favourite subject, say — and the mean has no meaning at all. Which summary is honest depends on the data and the question, never on habit.
Take a survey of car colours — a coded number for each: Red 1, White 2, Black 3, Silver 4, Blue 5. Add the codes, divide by the count, and you get a tidy decimal: 2.92. It looks exactly like an average should look.
Now swap the codes: call White 4 and Silver 2 instead. That same decimal changes, even though not one car has moved. A number that changes when you merely relabel the categories was never measuring anything real.
Categorical data — colours, yes/no answers, favourite subjects — has no meaningful distance between its values. There is nothing here for a mean to average. The mode, the single most common category, is the only measure of central tendency this kind of data actually supports.
An AI system that codes categories and quietly averages them will hand you a confident, precise-looking number that means nothing at all. It is a category error dressed up as arithmetic.
Weaker. Dhruv counts twenty-five cars by colour: 1 red, 5 white, 15 black, 3 silver, 1 blue. He codes them Red 1 through Blue 5 and works out the mean: (1×1 + 5×2 + 15×3 + 3×4 + 1×5) ÷ 25 = 73 ÷ 25 = 2.92. "Average colour: 2.92," he says.
Stronger. Nila asks a different question: which colour actually appears most? Fifteen of the twenty-five cars are black — more than all the other colours combined. That is the mode, and it is the honest crop. Colours carry no numerical distance for a mean to average in the first place.
"You can calculate the mean for any type of data."
Only for numerical data, where the distance between values means something. Code a category and average the codes, and the arithmetic runs fine — but the result measures nothing, because the codes were arbitrary to begin with. For categorical data the mode is the only central-tendency measure that means what it claims to mean. An AI that averages category codes will output a precise number attached to nothing real.
Mean, median, and mode all report the centre of a dataset — where the values cluster. But the centre is only half the picture.
A class with a mean of 65 and scores from 40 to 95 is one teaching situation. A class with the same mean of 65 and scores bunched between 60 and 70 is a very different one. Same centre, two very different realities — and the centre alone cannot tell them apart.
Data spread is exactly the other half: how widely the values sit around that centre. In AI, ignoring spread is how a model's one reported accuracy number can look impressive. It can still quietly fail one segment of its users.
Two instruments measure that spread. Range is simply the maximum minus the minimum — the full width of the data, in one subtraction.
It is also fragile. Change one extreme value and the range jumps, because it is built from only two points and ignores everything in between. Variance asks a different question — square each value's distance from the mean, then average those squares. The arithmetic that squares and averages every distance comes in Class 10 — for now, hold the idea itself: every point contributes, not just the two extremes.
Because variance uses every point instead of two, it stays far more stable when an extreme value shows up. Range catches the extreme; variance reads the whole shape around it.
Weaker. Dhruv's weather log shows ten days at these temperatures (°C): 22, 28, 31, 29, 25, 40, 24, 26, 30, 27. He adds them and divides by ten: 28.2°C. "That's manageable," he says, looking only at the mean.
Stronger. Nila points at the 40 sitting in that same list. Range = 40 − 22 = 18°C. An eighteen-degree swing across ten days is not manageable for outdoor work, and the mean alone never showed it. Report only the mean, and a farmer plans for 28° and meets 40° more than once that week.
ai-statistics.range-variance
/50-system/standards/book-look/cell1-a4/adv-aryabhata.png
Aryabhata
hoopoe
Read the number honestly, including what it hides — a mean can be perfectly true and still bury the spread the person living under it will feel.
A national health report states one figure: average disease incidence, 2%. Read on its own, that sounds like a small, manageable number, spread evenly and thinly across the country.
It is not spread evenly at all. That same 2% average can sit over most districts running near 0.1%, and one district running at 25%. That one district is an active emergency the average never once names.
Two datasets can share an identical mean and hide completely different realities behind it. The harm an AI system needs to catch is usually sitting in the tail the mean was built to omit. It is not in the centre the mean reports.
An AI that reports only means will miss exactly that concentrated harm. Centre and spread have to travel together, or the report is a crop, not the whole picture.
"If I know the mean (average), I know enough about the data."
The mean gives you the centre and nothing else. Two datasets can share a mean and differ wildly in how their values spread around it. A disease averaging 2% nationally can sit near zero almost everywhere and at 25% in one district — the average never says so. Knowing enough means holding the centre and the spread together, not the centre alone.
Put the question to yourself the way an examiner would: what have these five instruments actually given you? Three ways to report the centre of a dataset — mean, median, mode. And two ways to measure how far it spreads: range and variance.
Together they answer this chapter's question: what was measured, and summarised how? Get that far, and you can already read a claim honestly instead of trusting whatever single number sits in the headline.
One question still waits, and it belongs to the next chapter. A number can be an honest crop of the data you have. It can still be drawn from a field that does not represent the situation you actually care about. Probability, in Chapter 9, asks exactly that: compared with whom, and is that comparison representative?