Go deeper · Advanced track

Can AI know why, not just what usually happens?

Partly. Most machine learning finds what usually goes together. Knowing why takes extra tools: facts stored as explicit links, beliefs updated with evidence, and experiments that change one thing on purpose. These work today. Whether big neural models learn causes on their own is an open research question.

Grades 9+25 minutesFractions; Bayes’ rule optionalOptional advanced page

What you will be able to do

Follow fact triples to answer a question, update a belief from prior to posterior with counts, and explain why “doing” (a randomized experiment) can reveal a cause that “seeing” cannot.

Before you start

Read first: How sure should an AI be? (conditional probability, base rates) and What did we teach it by accident? (correlation vs causation). Optional: How does a computer plan without learning? This page is an optional extra, not needed for the main route.

1 · You have already seen this

Umbrellas don’t make it rain

Most people can tell when two things happen together without one causing the other.

On days when lots of people carry umbrellas, the streets are wet. Nobody thinks the umbrellas wet the streets. Rain causes both, and you know that because you have watched it rain.

A program that only reads records has no such background. It can learn that two things go together and predict one from the other, yet it cannot say which, if either, is the cause. This page is about knowledge and causality: tools that help AI get from what usually happens to why. It shows how they work, lets you try one, and shows where they go wrong.

2 · What happens, step by step

From what goes together to why

Getting from a pattern to a cause in four steps. The three sections after this one go deeper into steps 1, 2 and 3, and the lab lets you try step 4.

How AI can reason about causes

Four steps

  1. 1 · See what goes togetherRecords show which things rise and fall together, and a model can learn to predict one from the other. This first level of questions is called seeing, and most machine learning stays there.
  2. 2 · Look up known factsA knowledge graph stores facts as triples, such as (Paris, capital of, France), so a program can follow the links to an answer and show which facts it used.
  3. 3 · Update your beliefWith Bayesian reasoning, you start from how likely something was before a clue and work out how far the clue should move you. Each new belief becomes the starting point for the next clue.
  4. 4 · Change one thing on purposeTo test a cause, let a coin decide where the change happens. Hidden causes then spread across both groups, so a gap that remains comes from the change or from chance.

Going deeper · step 1

The kiosk owner’s question

A beach kiosk owner notices that on days she sells lots of ice cream, the lifeguards report more sunburns.

A model trained on her records would learn to predict sunburns from sales. But it cannot answer her real question: if I sell more ice cream, will more people get burned? The records show what happened together, not what causes what. (The answer is no: hot, sunny weather drives both.)

The computer scientist Judea Pearl describes three levels of questions. Seeing: what usually goes with what? Doing: what happens if I change something? Imagining: what would have happened otherwise? Most machine learning lives on the first level. Four ideas help AI reach higher.

Going deeper · step 2

Facts as links you can follow

Facts stored explicitly, so a program can look them up and combine them.

A knowledge graph stores each fact as a triple: a subject, a relation and an object, such as (Paris, capital of, France). Things become dots and relations become labeled arrows. To answer “Which river flows through the capital of France?”, a program follows two links: find the thing that is capital of France (Paris), then the thing that flows through Paris (the Seine). Nothing is learned; it searches step by step.

Search engines use them for fact boxes; Google launched its Knowledge Graph in 2012 with more than 3.5 billion facts. Every answer can be traced to stored facts. But the graph knows only what someone put in, facts go out of date, and most links say what is connected, not why.

A tiny knowledge graph Five things connected by labeled arrows: Seine flows through Paris. Paris is the capital of France. France is part of Europe. France uses the euro as currency. Loire flows through France. flows throughcapital of part ofcurrency flows through Seine Paris France Europe euro Loire
Five real facts; each arrow is one triple.

Predict first

The graph has no triple that says (Seine, flows through, France). Can a program still answer “Does the Seine flow through France?”

Show the answer

Only with a rule. Add “a river that flows through a city also flows through that city’s country,” and the program can follow Seine → Paris → France. Rules like this are the “symbolic” side of AI: explicit and checkable, but a wrong rule gives confident wrong answers.

Going deeper · step 3

Changing your mind by the right amount

Bayesian reasoning is a recipe for updating a belief when new evidence arrives.

You start with a prior: how likely something is before the new clue. The likelihood says how often you would see that clue if the idea were true, compared with if it were false. Combining them gives the posterior: your updated belief, which becomes the prior for the next clue. The probability page counted the same way for a spam filter.

Worked example

Will it rain this afternoon? Two clues, counted out of 1,000

  1. PriorIn this made-up town, 30% of summer afternoons are rainy: out of 1,000, 300 rainy and 700 dry.
  2. Clue 1: dark clouds at noonLikelihoods: noon clouds come before 80% of rainy afternoons and 20% of dry ones. Rainy with clouds: 80% of 300 = 240. Dry with clouds: 20% of 700 = 140. Cloudy in total: 380.
  3. Posterior after clue 1Only the 380 cloudy afternoons are still possible. Rainy: 240 ÷ 380 ≈ 63%, up from 30%.
  4. Clue 2: a falling barometerThe posterior is the new prior: 240 rainy, 140 dry. A falling barometer comes with 70% of rainy afternoons and 25% of dry ones. Rainy: 70% of 240 = 168. Dry: 25% of 140 = 35. Total: 203.
  5. Posterior after clue 2168 ÷ 203 ≈ 83%. Fairly sure, not certain: 35 of those afternoons stay dry.
  6. Optional: the formulaBayes’ rule: P(rain | clouds) = P(clouds | rain) × P(rain) ÷ P(clouds) = 0.8 × 0.3 ÷ (0.8 × 0.3 + 0.2 × 0.7) = 0.24 ÷ 0.38 ≈ 0.63.

All numbers are made up. Step 4 assumes the two clues are independent once you know the weather; a real forecaster would have to check that.

Bayes tells you how far a clue should move you, not why clue and rain go together. For causes, you step from seeing to doing.

3 · Try it

Change one thing on purpose

A hidden common cause can make two things move together. An experiment breaks that link.

A hidden cause that pushes two things up together is a confounder: here, the weather. You can adjust for it by comparing only days with the same weather, but only if you know and measure every confounder.

The stronger test is an intervention: change one thing yourself. In a randomized experiment, a coin decides which days (or people, or fields) get the change. The coin knows nothing about the weather, so hot and cool days spread across both groups, and any remaining gap comes from the change or from chance.

Predict first

Busy ice-cream days have about three times as many sunburns as quiet days. The kiosk gives away 100 free ice creams on days chosen by a coin flip. Will sunburns go up on free days?

Show the answer

No, apart from chance. The coin picks hot and cool days alike, so free days get the same mix of weather. Try four coin-flip schedules in the lab.

Lab

Seeing vs doing at the beach kiosk

Start by seeing: compare busy and quiet days. Reveal the hidden cause. Then do: let a coin choose the days for free ice cream, or for free sunscreen.

Simulated in your browser with 40 made-up days. Nothing is stored.

    Forty made-up summer days at a beach kiosk. Weather (cool, warm or hot) drives both ice cream sales and sunburns. Here are the average sunburns per day for each comparison the lab makes.

    Seeing vs doing at a made-up beach kiosk
    ComparisonGroup 1: sunburns a dayGroup 2: sunburns a dayGap
    Seeing: busy vs quiet days20 busiest days: 5.020 quietest days: 1.63.4 more on busy days
    Same kind of day: cool daysbusier half: 0.7quieter half: 1.6-0.9
    Same kind of day: warm daysbusier half: 2.4quieter half: 2.9-0.4
    Same kind of day: hot daysbusier half: 6.8quieter half: 6.70.2
    Doing: free ice cream on coin-flip days23 free days: 3.417 other days: 3.20.2 (chance)
    Doing: free sunscreen on coin-flip days23 free days: 1.517 other days: 3.2-1.8 (a real effect)

    Seeing shows a big gap. Comparing days of the same kind, or letting a coin choose the free-ice-cream days, makes the gap shrink to chance. Free sunscreen on coin-flip days still cuts sunburns, because in this made-up world sunscreen really does prevent them.

    Neuro-symbolic AI

    Patterns plus rules

    Each side is good at what the other finds hard.

    A neural network learns fuzzy patterns from data, so it copes with messy input. But what it knows is spread across millions of numbers; you cannot point to the fact it used. A symbolic system, like the knowledge graph above, is the opposite: checkable facts and rules that break on messy input.

    Neuro-symbolic AI combines the two: a network turns messy input into symbols, and a reasoning step applies rules to them. A loose version already works: a chatbot that looks facts up (retrieval) or calls a calculator through a tool call instead of guessing. How best to join the halves is still debated.

    Evidence vs speculation

    What works, and what is open

    Where the line runs, as of October 2026.

    Evidence vs open questions (October 2026)
    IdeaWorks today (evidence)Still an open question
    Knowledge graphsSearch engines show fact panels built from them (Google announced its Knowledge Graph in 2012); Wikidata is an open one anyone can edit.Keeping huge graphs complete, correct and up to date, especially when facts are pulled from text automatically.
    Bayesian reasoningSpam filters, forecasting and much of modern statistics update beliefs with evidence this way.Exact updating gets too slow for very large models, so they rely on approximations.
    Experiments (doing)Randomized experiments are the standard test of cause in medicine, farming and app design (A/B tests).Many questions cannot be tested this way because the experiment would be unsafe, unfair or impossible.
    Causal inference toolsCausal diagrams and adjusting for known confounders let researchers estimate effects from observational data.Every estimate rests on assumptions about hidden causes that the data alone cannot check.
    Large neural modelsThey often answer textbook cause-and-effect questions correctly in text.Whether they learn causal models of the world or mostly reuse causal statements from their training text; researchers disagree.
    Neuro-symbolic AIModels that call a calculator, a database or a search tool are a loose, working version of the idea.How best to combine learned patterns with explicit rules inside one system.

    When someone says an AI “understands cause and effect,” ask what was tested. Answering “Does sunlight help plants grow?” can come from text that says so. A stronger test asks about a made-up situation the model never read about. Such tests are still being built, so treat confident claims either way with care.

    4 · Common mix-ups

    What people get wrong about causes

    Four ideas that trip people up, and what to think instead.

    • Mix-up: A model that predicts well has found the cause.

      Better: Prediction only needs things that go together. Sales predict sunburns well, yet changing sales changes nothing.

    • Mix-up: A randomized experiment always gives exactly the right answer.

      Better: Chance still plays a part, as the four coin flips show. Bigger experiments shrink the chance gaps; repeating them builds confidence.

    • Mix-up: Knowledge graphs and neural networks are rival camps.

      Better: Today they are often used together: a model reads the question, and a graph or database supplies checkable facts.

    • Mix-up: If a chatbot explains a cause correctly, it understands causation.

      Better: It may be repeating explanations from its training text. Whether large models learn causal structure is an open question.

    5 · Check yourself

    Three quick questions

    Answer in your head or on paper before you open each one.

    1. A knowledge graph contains (Ada, works at, City Library) and (City Library, located in, Riverton). Which two links answer “In which town does Ada work?”, and what does the program need besides the two triples?

      Check your answer

      Follow Ada → works at → City Library, then City Library → located in → Riverton. The program also needs a rule or a query that says those two relations can be chained: someone works in the town where their workplace is located. (Ada and the library are made up.)

    2. In a made-up plant nursery, 40% of seedlings will grow blue flowers. A leaf test turns purple for 90% of blue-flower seedlings and 10% of the others. Out of 1,000 seedlings, what is the chance a purple-test seedling will flower blue?

      Check your answer

      Prior: 400 blue-flower, 600 other. Purple: 90% of 400 = 360 and 10% of 600 = 60, so 420 purple in total. Posterior: 360 ÷ 420 ≈ 86%, up from the 40% prior.

    3. A made-up school notices that students who eat breakfast at school get better grades. Name a possible confounder and describe an experiment that could test the breakfast itself.

      Check your answer

      Possible confounders: arriving early, family routines, or how far away students live. A fair test offers free breakfast to classes or days chosen at random and compares the results, so those other factors end up spread across both groups.

    For grades 6–8

    The coin-flip kiosk

    The seeing vs doing idea works with paper cards and a coin; skip Bayes’ rule.

    20 minutes · groups of four

    Did the coin fool us?

    1. Print the lab’s fallback table or write 40 day cards with weather, ice creams sold and sunburns.
    2. Sort the cards into busy and quiet days. Count the sunburns in each pile.
    3. Flip a coin for each card: heads means “free ice cream day.” Count sunburns on heads and tails cards.
    4. Compare the two gaps.

    Talk about it

    Why did the gap disappear?

    Good answers notice that the coin gave each pile a mix of hot and cool days, so the weather could no longer make one pile look worse. Younger classes can do the same with plants and a watering can.

    Honest limits

    What this page leaves out

    What this page leaves out

    Bayesian networks (many linked beliefs at once), causal diagrams and the rules for reading them, counterfactual (“imagining”) questions, natural experiments, logic programming and how knowledge graphs are built from text all get a sentence or less. The kiosk data is made up and much tidier than real data.

    Sources

    Where this comes from

    1. Judea Pearl and Dana Mackenzie (2018), The Book of Why: The New Science of Cause and Effect, Basic Books: the ladder of seeing, doing and imagining, confounders, interventions and randomized experiments.
    2. Stuart Russell and Peter Norvig, Artificial Intelligence: A Modern Approach, 4th edition, table of contents: chapters on knowledge representation, quantifying uncertainty and probabilistic reasoning (including causal networks).
    3. Aidan Hogan et al. (2020; ACM Computing Surveys, 2021), “Knowledge Graphs”: how graphs of triples are built, queried and used.
    4. Amit Singhal, Google (May 16, 2012), “Introducing the Knowledge Graph: things, not strings”: 500 million objects and 3.5 billion facts at launch.
    5. Bernhard Schölkopf et al. (2021), “Towards Causal Representation Learning”: why most machine learning captures associations, and open problems in learning causal structure.
    6. Artur d’Avila Garcez and Luis C. Lamb (2020), “Neurosymbolic AI: The 3rd Wave”: a research perspective on combining neural learning with symbolic reasoning.

    AI words on this page

    Knowledge graph, Bayesian reasoning, Causal inference, Neuro-symbolic AI, Correlation vs causation, Conditional probability, Base rate, Symbolic AI, Search and planning, Retrieval-augmented generation, Tool call

    All definitions live in one place: AI words in plain language →

    Review status

    Written by ARQIV. Last checked October 2026. Independent subject review and classroom pilots are still to come, so treat the activities as drafts and tell us about anything that is wrong or unclear.