A team wins 80% of its home matches. Before you believe that number, ask two questions: what got measured, and which matches were counted? Just as important — what sits outside that count?
Probability is the language built to answer that second question: it names the full field of outcomes — the sample space. Crop the same four wins down to five matches, and the number reads 80%. Zoom out to twenty matches, same four wins, and it reads 20%. Nothing about the wins changed. The frame did.
Name the full field before you count anything inside it. That discipline is what this whole chapter builds.
The Math for AI survey names four branches: statistics, linear algebra, probability, and calculus. Of these four, only two are actually worked out in Class 9 — statistics and probability. The other two appear by name only, not in your homework.
Probability earns its own chapter for one reason. Every AI prediction is a probability estimate — a number between 0 and 1, never a flat yes or no. This chapter's payoff is understanding that number — what it means, where it comes from, and when it is trustworthy enough to act on.
Probability is a number between 0 and 1, inclusive, that measures how likely an event is to occur. The formula: P(event) = favourable outcomes ÷ total equally likely outcomes.
Three values anchor the scale: P = 0 means impossible, P = 0.5 means equally likely either way, P = 1 means certain. Every calculation ahead runs through this formula — favourable outcomes on top, total on the bottom, divide.
Before calculating any probability, name the sample space — the complete set of all possible outcomes. A coin toss has sample space {H, T} — two outcomes. A standard die has {1, 2, 3, 4, 5, 6} — six outcomes. Toss two coins together and the sample space grows: {HH, HT, TH, TT} — four outcomes, one for each combination.
The sample space is the formula's denominator — the field the crop sits inside. A probability claim that does not name its sample space is incomplete: it has cropped the field without showing the frame.
Take the team that wins 80% of its home matches: four wins is the same four wins whether the sample space is five matches or twenty. Only the size of the named denominator changes the answer — not the wins.
One doubt worth flagging: does the sample space shift with the event you're asking about? It does not — section 4's die shows one fixed six-outcome sample space answering four separate questions in turn.
ai-probability.sample-space
/50-system/standards/book-look/cell1-a4/adv-dignaga.png
Dignaga
kingfisher
Before you trust any fraction, name its field exactly — the denominator is a definition, and a percentage means nothing until you have said what it was counted against.
A fair coin has probability 0.5 of landing heads. Toss it ten times and it feels like five heads should come up. Try it — you might get three. You might get eight.
Probability is a long-run frequency, not a guarantee about the next trial. Toss it hundreds of times, and the proportion of heads creeps toward 0.5. But any short run can sit far from it — and still be perfectly consistent with a fair coin.
The same rule holds for an AI model: predicted probability 0.7 is right roughly 70% of the time in the long run — any one prediction can still land in the wrong 30%. A model trained on a small dataset guesses from a narrow sample space, which gives an unreliable long-run estimate. A probability is only as trustworthy as the sample space behind it.
Every event sorts into one of four types — really, one question: how many outcomes qualify, against how many are possible?
An impossible event has probability 0 — nothing qualifies (rolling a 7). A certain event has probability 1 — every outcome qualifies (rolling 1 to 6). A simple event has exactly one qualifying outcome (rolling a 4, P = 1/6). A compound event groups more than one: rolling an even number, 2, 4, 6 qualify, P = 3/6 = 1/2.
Event type only tells you how to count the numerator — the denominator, the sample space, never moves.
"A compound event is always more likely — it covers more outcomes."
Compare the count, not the label: even numbers (2/4/6) give P = 1/2, a 4 alone gives P = 1/6 — compound wins only because three beats one. The group's size alone decides nothing — only the count against the fixed sample space does.
Calculate a probability in three fixed steps.
Two coin tosses: {HH, HT, TH, TT} — four outcomes. "At least one head" excludes TT: three outcomes qualify. P = 3/4 = 0.75.
That order keeps the denominator honest.
Four different questions, one die, one sample space: {1, 2, 3, 4, 5, 6}, six outcomes, unchanged throughout.
P(rolling a 3) = 1/6 — one outcome qualifies. P(rolling an odd number) = 3/6 = 1/2 — event set {1, 3, 5}. P(rolling a number greater than 6) = 0/6 = 0 — impossible, nothing qualifies. P(rolling a number from 1 to 6) = 6/6 = 1 — certain, every outcome qualifies.
The denominator is fixed by the experiment, not the question.
Teach a machine three guesses: win chance, rain chance, traffic-jam chance. Feed it a record of what actually happened.
In sports: win probability is wins ÷ total matches. For weather: fraction of similar days it rained. For traffic: fraction of similar hours it was congested.
Three domains, one structure: history supplies the sample space, and a bigger history gives a more reliable number.
A spam filter hands you P(spam) = 0.94. A medical AI hands you P(pneumonia) = 0.78 — numbers that close to 1 feel almost certain.
A high probability is not certainty. P = 0.94 means the model is right roughly 94% of the time in the long run. Any single prediction can still land in the remaining 6%. Only P = 1 means certain.
The model outputs the probability; a human or a rule turns it into an action. An honest AI report always states the probability, not just the decision.
"If an AI model gives a high probability — like 0.95 — the answer is definitely correct."
P = 0.95 means right ~95% of the time — not guaranteed correct. A 95%-accurate spam filter still misses 1 in 20 emails — the number that matters most. Only P = 1 is certain, and real AI systems rarely reach it.
Take a coin you suspect might be biased. Toss it 20 times: 13 heads, so P(heads) = 0.65 — far from 0.5. Biased coin, or a short, unlucky-looking run?
Toss 60: 0.53. Toss 120: 0.50 — the early 0.65 was illusion. A spurious pattern works the same way: if the effect is not real, a large enough sample converges its long-run frequency back to the true value, and the deviation disappears.
This is why more training data is the first cure for a spurious pattern. A pattern that survives a large sample is more likely genuine. One that shrinks as the sample grows is more likely noise.