How does superintelligence work when an AI sees two events happen together and has to decide whether one caused the other? Imagine that pupils who attend extra practice also tend to score higher. Did the practice help? Perhaps. But perhaps pupils who were already more motivated attended the sessions, or several changes occurred at once. A confident explanation can outrun the evidence long before anybody notices.
This guide explains correlation, causation and causal inference in easy language. A correlation is a relationship between measured variables. A causal claim says that changing one factor brings about a change in another under the relevant circumstances. Establishing the second takes more than spotting the first. Advanced AI systems can help analyse data, but they still need the right evidence, assumptions and checks.
We will study a fictional after-school revision club with twelve inspectable statements, separate findings from guesses, and show what stronger testing would require. All scores and pupil groups are invented illustrations, not results from a real tutor or an observed AI model. “How Superintelligence Works” names an educational series; it is not a claim that current AI has achieved artificial superintelligence.
1. Two things happening together is a beginning, not a conclusion
A school might notice that pupils who read more often have stronger vocabulary. That observation could arise because reading provides word exposure. It might also reflect home resources, earlier vocabulary knowledge, language interests, school differences and other influences.
Identifying a pattern is valuable: it points toward a question worth studying. Yet a description of which outcomes occur together is not itself an experiment. A careful AI should distinguish “associated with” from “caused by” and state what further evidence would separate the possibilities.
2. Why “after” is not the same as “because”
If a pupil’s test score increases after a study session, the score change happened later in time. But the pupil may have also slept better, received classroom feedback or encountered an easier assessment. The phrase “after the session” is a temporal statement; “because of the session” is a causal explanation.
An assistant that rewrites every sequence as a cause will sound persuasive and make unsupported claims. Time order is often necessary for a proposed cause to influence an effect, but time order alone is not sufficient evidence that the cause was responsible.
3. What correlation means in statistics
In a statistical context, correlation summarises aspects of how variables move together. A positive relationship means larger values of one variable tend to accompany larger values of another; a negative relationship means the opposite. A correlation coefficient can describe strength and direction for particular kinds of association.
But the numerical coefficient does not identify which variable caused the other, whether there is a hidden common cause or even whether a relationship is stable outside the sample. The Australian Bureau of Statistics explicitly distinguishes correlation from causation, including familiar examples of seasonal patterns.
4. The hidden third variable
Suppose sunscreen sales and ice-cream purchases both rise during hot weather. It would be a mistake to infer that buying sunscreen causes somebody to buy ice cream. Hot weather can influence both. The temperature and season are potential common causes.
In education, the analogous third variable might be prior motivation. A motivated pupil may choose extra revision and also spend time studying independently. If an AI ignores that background factor, it may attribute every advantage to the revision session itself.
5. The reverse-direction problem
Consider “Pupils who ask more questions have better understanding.” Do questions improve understanding, or do pupils who already understand more feel confident asking questions? Both processes might operate, and a third factor might influence both.
Correlation cannot generally tell us the direction of a causal arrow by itself. A model that sees two related variables must avoid automatically choosing the direction suggested by the order of the words in a sentence.
6. The fictional Willow Revision Club
Imagine twenty fictional pupils in a school. Ten chose to attend an optional revision club during a month, and ten did not. All twenty completed two practice assessments. The invented improvement indicator is simply whether the pupil scored higher on the later assessment than on the earlier one.
Eight attendees improved, leaving two attendees who did not record a higher score. Five non-attendees improved, leaving five who did not. These counts are complete for the fictional twenty-pupil exercise. The club attendance was voluntary, not randomly assigned, and we do not have verified information about prior motivation, lesson difficulty or other changes.
7. Build the two-by-two comparison
Among club attendees, eight of ten improved: 8/10 = 80%. Among non-attendees, five of ten improved: 5/10 = 50%. The observed rate difference is 80% minus 50%, or 30 percentage points. Across both groups, thirteen of twenty improved.
Those calculations are correct for this invented record. They do not reveal whether the club itself caused thirty additional percentage points of improvement. An observed rate difference is not automatically a causal treatment effect.
8. The fictional observation table
- Club attendance recorded: 10 pupils.
- Attendees with a higher later score: 8 pupils.
- Attendees without a higher later score: 2 pupils.
- No club attendance recorded: 10 pupils.
- Non-attendees with a higher later score: 5 pupils.
- Non-attendees without a higher later score: 5 pupils.
- Combined number with higher scores: 8 + 5 = 13.
- Combined number without higher scores: 2 + 5 = 7.
- Attendee improvement rate: 8/10 = 80%.
- Non-attendee improvement rate: 5/10 = 50%.
- Observed difference: 30 percentage points.
- Missing from the record: random assignment, the pupils’ motivations and a design that isolates the club’s causal effect.
9. Twelve statements for an AI to classify
The Willow desk asks the assistant to label each sentence SUPPORTED, CONTRADICTED or NOT ESTABLISHED from the source. If a statement makes a causal claim that the counts cannot prove, it belongs in the last category—not automatically false.
Notice how this structure tests both calculation and restraint. A cautious answer must report useful differences without pretending that every useful difference is already a proven effect.
- 01. “Eight out of ten club attendees improved.” — SUPPORTED.
- 02. “Five out of ten non-attendees improved.” — SUPPORTED.
- 03. “The observed improvement-rate difference was 30 percentage points.” — SUPPORTED.
- 04. “The revision club caused a 30-percentage-point improvement.” — NOT ESTABLISHED.
- 05. “Fewer attendees improved than non-attendees.” — CONTRADICTED; eight exceeds five.
- 06. “Both observed groups included ten pupils.” — SUPPORTED.
- 07. “Thirteen pupils in total recorded higher later scores.” — SUPPORTED.
- 08. “Every pupil improved.” — CONTRADICTED; seven did not show a higher later score.
- 09. “The pupils were allocated randomly to the club.” — CONTRADICTED: the setup explicitly says attendance was voluntary and not randomly allocated.
- 10. “Attendees were more motivated before the club began.” — NOT ESTABLISHED.
- 11. “Seven of the twenty did not record a higher later score.” — SUPPORTED.
- 12. “Three more attendees than non-attendees improved.” — SUPPORTED.
10. Check a methodological contradiction carefully
Statement 09 deserves a stricter reading than a numerical claim. The scenario explicitly describes voluntary attendance and says pupils were not randomly assigned. Therefore an assertion that they were randomly allocated is not just unsupported; it is CONTRADICTED by the setup. A classifier must read the methodological description, not only the two-by-two table.
The supported statements total seven (01, 02, 03, 06, 07, 11, 12). Contradicted statements total three (05, 08, 09). Not-established statements total two (04, 10). Seven plus three plus two equals twelve. This explicit reconciliation helps readers test whether each claim was categorised against the source rather than by a plausible guess.
11. Why “not established” is not “disproven”
Statement 04 says the club caused the observed difference. The invented data do not establish that cause. But lack of sufficient causal evidence is not a demonstration that the club had no effect. It might have helped substantially, helped a little, or had no effect; the observation alone cannot settle the size of the causal contribution.
Likewise, we have no measured motivation data. That does not prove all pupils had equal motivation. It means we should avoid claiming any specific motivation difference. This is one of the clearest habits separating scientific caution from empty scepticism.
12. What a confounder does
A confounder can help create or distort an observed association when it is related to both the exposure and outcome under the causal structure being studied. Here, prior motivation may influence the choice to attend and the amount of learning outside the club. That makes motivation a plausible concern, not a confirmed explanation.
Real researchers define confounders using subject-matter knowledge and assumptions, not merely by labelling every available background variable a confounder. Adjusting for the wrong variable can sometimes introduce bias. A good AI analysis should therefore distinguish proposed confounders from documented variables and validated adjustments.
13. Why voluntary participation matters
Pupils who choose optional activities may differ from those who do not in ways related to future scores. Even when the groups happen to have the same number of pupils, equal size does not mean equal readiness, support or prior performance.
A comparison of ten and ten is arithmetically balanced, but it may still be causally unfair. A model that says the study was controlled merely because both groups have ten participants is substituting symmetry of counts for an actual research design.
14. What random assignment changes
In an appropriately designed randomised experiment, participants are allocated to groups by a chance procedure rather than by preference or a researcher’s choice. This can reduce systematic differences between groups on measured and unmeasured baseline factors, although chance imbalance can remain and execution or attrition can create complications.
Random assignment is valuable because it gives researchers a stronger basis for comparing outcomes. It does not magically guarantee flawless evidence. Small samples, inconsistent participation, different assessments and missing results still require careful evaluation.
15. Random sampling and random assignment are different
Random sampling concerns how participants are selected from a wider population. Random assignment concerns how selected participants are allocated to conditions. An experiment may use random assignment among volunteers without those volunteers representing all pupils in Singapore.
Therefore even a well-conducted study of volunteer pupils might have stronger evidence about what happened to those volunteers than about what would happen to every schoolchild. A trustworthy AI report separates internal validity from wider generalisability.
16. An observational study can still be valuable
Not every scientific question permits random assignment. Researchers cannot ethically randomise every exposure or educational experience. Observational studies can provide important evidence when their designs, comparisons, assumptions and independent checks are credible.
The correct lesson is not “observational data are useless”. It is that causal conclusions from observational data require deliberate methods and stronger justification than a raw association. The National Academies’ Reference Manual on Scientific Evidence and the CDC discuss this distinction in detail.
17. The counterfactual idea
A causal question asks, in effect: what would have happened to the same participant under the alternative condition? For a pupil who attended the club, we observe the score under attendance. We cannot simultaneously observe that very same pupil, at the same moment, having lived through the same month without the club.
This is the fundamental difficulty of causal inference. Comparisons, random assignment and statistical modelling help estimate outcomes under alternative conditions across groups, but the missing individual counterfactual is not directly observed.
18. Why the exact score change may matter
Our exercise uses a yes-or-no improvement indicator. It does not show whether a pupil improved by one mark or twenty marks. Two groups could have similar proportions of pupils with higher scores but very different average gains.
A richer study might analyse score differences, distribution of gains, baseline scores, variation across topics and long-term retention. The choice of outcome influences what “effective” means. A model must not invent effect sizes from a binary indicator.
19. Test difficulty is an alternative explanation
If the later assessment was easier, some pupils might score higher even without improved understanding. If the later test assessed different material, the difference could reflect content rather than learning progress. If scores were subject to marking changes, the interpretation becomes harder still.
A causal report should inspect whether assessments are comparable and whether improvement was defined in advance. The AI can calculate the recorded outcomes accurately and still be wrong to describe them as a measured learning gain if the underlying test was not comparable.
20. Regression to the mean
Pupils selected because of unusually low scores may obtain more typical results on later tests even without a powerful new intervention. This statistical tendency is called regression to the mean. It can make a programme appear successful if the comparison group and selection rule are ignored.
This does not mean genuine improvement never occurs. It means the design must distinguish improvement attributable to the programme from a rebound that might have happened anyway. A thoughtful AI explanation identifies the possibility while avoiding the unsupported claim that it definitely explains these fictional results.
21. Selection bias and missing outcomes
If only pupils who complete both assessments are included, the result may exclude those who stopped attending, missed the second paper or changed schools. The resulting observed group can differ systematically from the population the programme was meant to serve.
In our toy record, all twenty fictional pupils are included by construction. A real report must state how many were invited, enrolled, attended and completed assessments. Losing pupils between those stages changes denominators and can distort conclusions.
22. A plausible story is not evidence
“The club worked because its friendly tutors created confidence” is an attractive explanation. But friendliness, confidence and the route from one to the other were not measured in the case. The assistant should describe it as a possible mechanism, not as an established explanation.
Scientific communication benefits from separating observation, hypothesis and tested causal mechanism. One hypothesis can motivate a new study. It cannot retroactively turn an observational rate difference into evidence that the hypothesised chain occurred.
23. The fictional baseline model’s errors
Imagine a fictional assistant answering the twelve classification questions and getting eight right. It calls the observed thirty-point difference a proven causal effect, mislabels voluntary attendance as random assignment, guesses that motivation was higher among attendees, and says all pupils improved. This is a made-up teaching result of 8/12, or about 66.7%.
Those errors involve four different mistakes: causal overclaiming, study-design misreading, missing-variable invention and universal-claim failure. A repair should address each one explicitly rather than asking for more confident prose.
24. The revised fictional process
Require the assistant to identify the measured outcome, determine each group denominator, calculate the association, inspect group formation, list measured versus hypothetical alternative explanations and classify causal claims according to evidence strength. Suppose the revised fictional assistant answers all twelve correctly.
That is 12/12 only for a deliberately simple invented exercise. It does not establish reliable causal inference in complex real-world studies. Real evaluation would include varied designs, hidden confounding challenges, measurement errors, contradictory sources and independent human review.
25. A transfer exercise about reading
A school notices that pupils who borrow more library books score higher in comprehension. What can we say? We can say there is an association if the measurements and comparison are sound. We cannot conclude from that association alone that borrowing books caused the scores.
Possible explanations include additional reading practice, prior reading interest, home environment and school resources. The next useful question is not which explanation sounds nicest, but what study or additional evidence could help distinguish them.
26. A transfer exercise about school attendance
Suppose pupils attending more lessons tend to have better test results. Does this prove lessons have no effect? No. Does it prove that attendance causes every improvement? No. Attendance, health, preparation, access, prior ability and other factors may all matter.
A rigorous explanation can acknowledge that attendance plausibly influences learning while remaining clear about what a particular dataset has shown. This avoids the opposite errors of credulity and blanket dismissal.
27. Causal language in the media
Watch for verbs such as causes, prevents, doubles, leads to and boosts. Ask whether the underlying source measured an association, conducted an intervention or relied on a theoretical explanation. A headline can be much stronger than the evidence in a study.
An AI assistant summarising news should preserve the study design and uncertainty. “Researchers observed a higher rate among participants who chose the activity” is not interchangeable with “the activity increased the rate”.
28. How an AI system should cite the claim
An auditable summary identifies the source, period, study population, treatment or exposure, outcome, comparison group and uncertainty. If it reports a percentage, it shows the numerator and denominator. If it reports a causal effect, it states what design and assumptions support the inference.
Source links alone are not enough if they point to articles that do not contain the stated numbers or describe a different population. Grounding requires the source passage to support the actual claim. A model must not turn an adjacent paper into a citation for an invented result.
29. Statistical significance is not a causal certificate
A result can be statistically significant yet fail to establish causation when confounding, selection or measurement bias remain. The CDC’s Field Epidemiology Manual specifically distinguishes significance from a cause-and-effect association.
Likewise, an imprecise result in a small study is not proof of no effect. The language of certainty should match the design and data. An AI should avoid using the word “proven” merely because a p-value, confidence interval or large percentage appears in a source.
30. When an experiment is not available
Researchers may use natural experiments, quasi-experimental methods, longitudinal observations, instrumental variables or other carefully justified designs when randomised trials are impractical. Each method has assumptions and failure modes; naming one does not automatically make a causal claim valid.
For a broad educational explanation, the key is to understand what the comparison is supposed to identify. A credible study should explain why the comparison isolates the factor of interest and what other causes might remain.
31. The questions a parent can ask
If a tuition programme advertises an improvement rate, ask how improvement was measured, whether all enrolled pupils were counted, whether the same examination standard was used, how many were compared and whether other support was present.
A genuine improvement in reported scores can still be meaningful to a family. It just should not be inflated into universal proof of causation without appropriate evidence. Clear reporting is more useful than a dramatic claim that nobody can verify.
32. A practical test for AI builders
Create paired prompts that differ only in the evidence available. In the first, present observational counts. In the second, add a credible random-assignment protocol and appropriate outcome data. A reliable model should become more confident about causal interpretation only when the design actually justifies it.
Add cases with reverse causality, a common cause, unequal baseline readiness and conflicting data. Ask the system to label descriptive statistics separately from causal estimates. Measure unsupported causal statements, not just whether the generated summary sounds scientific.
Frequently asked questions
Is correlation always meaningless? No. It describes patterns that can be useful for prediction and generating hypotheses.
Does correlation prove causation? Not by itself. Causal inference needs a suitable research design and justified assumptions.
Does a stronger correlation necessarily mean a stronger cause? No. A large association can arise from confounding, bias or other processes.
If something happens first, did it cause what followed? Not necessarily. Sequence alone is insufficient.
Do randomised experiments guarantee the correct answer? No. They strengthen certain inferences but can still suffer from small samples, poor measurement, noncompliance and other issues.
Why not call the club ineffective? Our invented observation does not identify the causal effect. Not established is different from disproven.
The central lesson
An observed difference is a measurement. A causal explanation is an additional claim about why the difference occurred. An AI system becomes more reliable when it keeps those two layers separate, checks the method that generated the comparison and refuses to treat a compelling story as proof.
For students, parents, teachers and builders, the useful habit is simple: when someone says “because”, ask what evidence establishes the causal link, what alternative explanations remain and what comparison could change the answer. That question turns a pattern into the beginning of inquiry rather than the end of it.
Connected lessons and authoritative sources
- How Super Intelligence Works — full curriculum — the broader series on reasoning and verification.
- Comparisons and Baselines — what a fair comparison must establish.
- Quantifiers — rates, denominators and what most really means.
- Australian Bureau of Statistics: Correlation and causation — official definitions and common-cause examples.
- CDC Field Epidemiology Manual: Analysing and Interpreting Data — significance, confounding and causal caution.
- National Academies: Reference Manual on Scientific Evidence — observational versus experimental causal evidence.
- NIH/NICHD: Using Research and Reason in Education — experimental design and third-variable explanations.