How does superintelligence work when an AI says a warning system detects 80% of real mistakes? It is tempting to conclude that any warning must have an 80% chance of being correct. That is not generally true. If actual mistakes are uncommon, a large number of false alarms may appear among the system’s warnings even when its detection performance is useful.
This guide explains conditional probability and base rates in clear everyday language. We distinguish the chance of a flag when an answer is wrong from the chance that an answer is wrong when a flag appears. The order of those conditions matters. A reliable AI system should show the real denominator and avoid turning one conditional percentage into a different one.
We will examine a fictional school-answer checker across one thousand cases, calculate what its warnings mean, and compare the same checker in a different population. The numbers and machine behaviour are invented learning examples, not measurements of a live education service or real pupils. “How Superintelligence Works” is the educational title of this series, not a claim that current AI models have achieved artificial superintelligence.
1. The question hidden inside a percentage
Suppose someone says, “This checker finds eighty per cent of errors.” We must ask: eighty per cent of what? If the denominator is all genuinely incorrect responses, the figure describes sensitivity or recall for error detection. It does not directly tell us how many displayed warnings are correct.
Another question is: among all responses flagged by the checker, how many really contain an error? That has a different denominator: all flagged responses. A trustworthy explanation names both the event and the condition instead of repeating one attractive number.
2. What conditional probability means
Conditional probability describes the chance of one event given that another condition is known to hold. “Probability of a flag given the answer is wrong” and “probability the answer is wrong given a flag” reverse which event is assumed.
The quantities may be related through Bayes’ theorem, but they are not automatically equal. The difference is often a source of misleading medical, educational and technical claims. The mathematical habit is to spell out what follows the word given.
3. Sensitivity uses real errors as its denominator
If a checker detects eighty out of one hundred genuinely wrong answers, its sensitivity for detecting wrong answers is 80/100 = 80%. The twenty missed errors are false negatives.
This is useful because a checker with high sensitivity catches many real problems. But it says nothing by itself about the number of correct answers the checker might wrongly flag.
4. False-positive rate uses correct answers
If a checker incorrectly flags ninety of nine hundred genuinely correct responses, the false-positive rate is 90/900 = 10%. The remaining 810 correct responses are correctly passed.
That ten per cent may sound small. Yet in a large collection dominated by correct answers, the absolute number of false alarms can still be comparable to or exceed the number of useful warnings. The population’s composition matters.
5. The base rate is the starting prevalence
A base rate tells us how common the target condition is in the defined population before seeing the individual test result. For our fictional question checker, the target condition is a genuinely incorrect response.
If one hundred of one thousand responses are wrong, the base rate of errors is 100/1,000 = 10%. This background frequency will influence the interpretation of a positive warning.
6. The fictional Cedar Answer Checker
Imagine Cedar School has a complete reference set of one thousand fictional quiz responses. An independent answer key identifies one hundred as substantively wrong and nine hundred as correct. A made-up checker evaluates each response and either flags it for review or passes it.
By stipulation, it flags eighty of the one hundred real errors and misses twenty. It also falsely flags ninety of the nine hundred correct responses and correctly passes 810. This is a precisely defined toy dataset, not evidence about any actual AI.
7. The complete one-thousand-response record
- Total responses: 1,000.
- Actually wrong responses: 100.
- Actually correct responses: 900.
- True positives, wrong and flagged: 80.
- False negatives, wrong and not flagged: 20.
- False positives, correct but flagged: 90.
- True negatives, correct and not flagged: 810.
- All flags: 80 + 90 = 170.
- All not flagged: 20 + 810 = 830.
- Grand total: 80 + 20 + 90 + 810 = 1,000.
8. The 80% number is real but narrow
Among the one hundred wrong responses, the checker flags eighty. Therefore P(flag | wrong) = 80/100 = 80%. That matches the stated sensitivity.
Reading it as P(wrong | flag) = 80% would reverse the condition. We have not yet computed that second quantity. Many persuasive summaries make exactly this jump without showing the denominators.
9. Count how many warnings are real
The checker raises 170 warnings in total. Eighty point to genuinely incorrect responses; ninety are false alarms for correct responses.
The fraction of flags that correspond to actual errors is 80/170 = 8/17 ≈ 47.06%. In this invented collection, fewer than half of the warnings are correct even though the checker catches eighty per cent of real errors.
10. Why this is not a contradiction
The checker has 80% sensitivity to real mistakes and about 47.1% precision among its warnings. Both percentages can be correct because they answer different conditional questions.
One denominator contains one hundred genuinely wrong answers. The other contains 170 flagged answers. The difference is caused partly by the much larger population of correct responses, which provides many opportunities for false alarms.
11. What a negative result means
The checker passes 830 responses. Among them, 810 are genuinely correct and twenty are wrong but missed. Therefore P(correct | no flag) = 810/830 ≈ 97.59%.
That is a reassuring observed negative predictive proportion within this particular fictional population. It is not proof that any individual unflagged answer must be correct. Twenty genuinely wrong responses remain in the passed group.
12. The probability of an error among unflagged responses
Among the 830 cases without warnings, twenty are actually wrong. The observed fraction is 20/830 ≈ 2.41%.
This is complementary to the 97.59% correct-without-flag fraction. It must not be confused with the checker’s false-negative rate of 20/100 = 20%, which uses all genuinely wrong responses as its denominator.
13. Accuracy answers a third question
The checker makes 80 correct flags and 810 correct passes, totalling 890 correct classifications. Overall accuracy is therefore 890/1,000 = 89%.
That figure can be reported as part of an evaluation, but it hides which mistakes occurred. In a collection with many correct responses, a system may achieve high accuracy by passing almost everything while missing important errors.
14. Bayes’ theorem makes the reversal explicit
Write E for a genuinely wrong answer and F for a checker flag. The desired probability after a flag is P(E | F). Bayes’ theorem combines the error base rate with the checker’s sensitivity and false-positive rate.
For this toy record, P(E | F) = (0.80 × 0.10) / [(0.80 × 0.10) + (0.10 × 0.90)] = 0.08/0.17 ≈ 47.06%. The same result follows directly from counting eighty genuine errors among 170 warnings.
15. Natural frequencies can be easier than formulas
Instead of starting with symbols, say: imagine one thousand answers. One hundred are wrong. Eighty of the wrong ones are flagged. Nine hundred are correct, but ninety of these are also flagged. Now ask what percentage of the 170 flagged answers are wrong.
The answer is eighty out of 170. This concrete explanation is often easier for beginners because every number has a visible meaning and the final denominator is no longer hidden.
16. Change the base rate without changing the checker
Now imagine a different fictional set of one thousand responses where five hundred are wrong and five hundred are correct. Keep the same sensitivity of 80% and false-positive rate of 10%. This produces four hundred true warnings and fifty false warnings.
The total flagged is 450. Among those warnings, four hundred are genuine errors, so the precision is 400/450 ≈ 88.89%. The checker has not changed; the population has.
17. Two populations, two different warning meanings
In the first population, errors have a 10% base rate and flagged-case precision is about 47.1%. In the second, errors have a 50% base rate and flagged-case precision is about 88.9%.
This does not mean a system should deliberately increase the number of wrong answers to improve a metric. It demonstrates that a test’s predictive meaning depends on how common the target condition is among the cases being tested.
18. The comparison scorecard
- Population A: 1,000 responses; 100 wrong, 900 correct.
- Population A: 80 true flags, 90 false flags.
- Population A: 170 flags, of which 80/170 ≈ 47.06% are genuine.
- Population B: 1,000 responses; 500 wrong, 500 correct.
- Population B: 400 true flags and 50 false flags.
- Population B: 450 flags, of which 400/450 ≈ 88.89% are genuine.
- Both populations use 80% sensitivity and 10% false-positive rate.
- Base rate of actual errors: 10% in A and 50% in B.
- The predictive meaning of a flag changes with the target condition’s prevalence.
19. Why “an 80% reliable warning” is ambiguous
A notice advertising “80% accurate detection” may refer to recall, overall accuracy, precision or something else. Those are different statistical measures. A conscientious AI assistant asks which measure was actually reported.
If the report means sensitivity, translating it to “80% of warnings are right” is unsupported. The latter is precision and requires knowledge of false alarms and the tested population.
20. The test’s target condition must stay fixed
In our example POSITIVE means the answer is wrong. If another report defines positive as correct, the labels and conditional interpretations change. A warning can be called a positive prediction only under the chosen convention.
An AI should state the target condition rather than assume that positive means desirable or negative means bad. Statistical terminology follows the test design, not ordinary emotional meanings of the words.
21. Twenty-four claims for an AI to classify
The fictional desk asks the model to label statements SUPPORTED, CONTRADICTED or NOT ESTABLISHED, using only the two stipulated populations. Claims about real school outcomes, pupil ability or live model accuracy remain outside the evidence.
The test focuses on the direction of conditional probabilities and the influence of base rates. A correct fraction must have the correct denominator.
- 01. “Population A has one thousand responses.” — SUPPORTED.
- 02. “Exactly one hundred responses in A are wrong.” — SUPPORTED.
- 03. “A’s error base rate is ten per cent.” — SUPPORTED.
- 04. “The checker flags eighty of one hundred wrong responses in A.” — SUPPORTED.
- 05. “Sensitivity for wrong-answer detection in A is 80%.” — SUPPORTED.
- 06. “The checker flags ninety genuinely correct responses in A.” — SUPPORTED.
- 07. “False-positive rate in A is 10%.” — SUPPORTED.
- 08. “All warnings in A are genuine errors.” — CONTRADICTED.
- 09. “There are 170 flags in A.” — SUPPORTED.
- 10. “Precision among A’s flags is about 47.1%.” — SUPPORTED.
- 11. “Eighty per cent of A’s flags are right.” — CONTRADICTED.
- 12. “Exactly twenty wrong answers receive no flag in A.” — SUPPORTED.
- 13. “There are 830 unflagged cases in A.” — SUPPORTED.
- 14. “Among unflagged A cases, 810 are genuinely correct.” — SUPPORTED.
- 15. “All unflagged answers in A are correct.” — CONTRADICTED.
- 16. “Overall classification accuracy in A is 89%.” — SUPPORTED.
- 17. “The same sensitivity and false-positive rate are used in population B.” — SUPPORTED.
- 18. “Population B contains five hundred wrong answers.” — SUPPORTED.
- 19. “B produces 400 genuine flags and 50 false flags.” — SUPPORTED.
- 20. “B’s precision among flags is approximately 88.9%.” — SUPPORTED.
- 21. “The meaning of a flag is identical in A and B.” — CONTRADICTED.
- 22. “These counts prove the checker is safe for every real exam.” — NOT ESTABLISHED.
- 23. “These counts identify which individual pupils will make errors tomorrow.” — NOT ESTABLISHED.
- 24. “Changing only the base rate can change precision.” — SUPPORTED.
22. Reconcile the twenty-four labels
Supported are 01, 02, 03, 04, 05, 06, 07, 09, 10, 12, 13, 14, 16, 17, 18, 19, 20 and 24: eighteen. Contradicted are 08, 11, 15 and 21: four. Not established are 22 and 23: two. Eighteen plus four plus two equals twenty-four.
The key lesson is that 80% sensitivity and 47.1% flagged-case precision are not competing answers to one question. They are correct descriptions of different conditional groups.
23. Why a positive prediction is not a verdict
A flagged answer is a request for review based on a predicted error. It is not itself a confirmed mistake. A teacher should check the response against the answer key and the marking rules before treating the prediction as a fact.
That distinction protects pupils from being told their work is wrong based solely on an algorithmic warning. Automated checking can prioritise attention while independent verification establishes the actual result.
24. The inverse conditional mistake
In natural language, people often hear “most errors cause a warning” and mentally substitute “most warnings mean errors”. This is a reversal of the conditioning event. The first sentence begins with actual errors; the second begins with warnings.
One quick repair is to rewrite both statements as explicit questions: “Out of all errors, how many get flagged?” and “Out of all flags, how many are errors?” Then fill in the correct counts.
25. What calibration does and does not add
A calibrated probability estimate may communicate uncertainty about an outcome under conditions similar to its evaluation set. It does not eliminate base-rate dependence, guarantee any individual result or grant permission for a high-consequence action.
Calibration and the confusion matrix measure different properties. A system can have useful sensitivity yet poor positive predictive value in a low-prevalence population. A trustworthy report includes relevant base rates and does not rely on one flattering score.
26. The effect of a stricter warning threshold
An AI checker may flag fewer cases when its decision threshold is raised. This might reduce false alarms but also miss more real errors, depending on the score distribution. The new sensitivity and false-positive rate must be measured rather than assumed.
We cannot compute the results of changing the threshold from our two-population counts alone. The fixture supplies one operating point, not the full set of possible score cutoffs.
27. Real populations may differ
Our two populations are deliberately simple and have exact known counts. Real school responses vary by topic, year level, marking criteria and selected group. Estimates from one class may not transfer to another.
Before using a published performance percentage, check what sample produced it and whether your actual tasks resemble that sample. A good AI explanation preserves those limits instead of promising universal reliability.
28. A rare-event problem
Suppose real mistakes become very rare. Even a modest false-positive rate can produce many flags among the large population of correct items. The meaning of a warning can then be less impressive than the detection percentage alone suggests.
This is not an argument against automated screening. It is an argument for context-aware interpretation, good thresholds, independent review and honest reporting of false alarms.
29. An illustration beyond classrooms
Imagine a fictional library scanner that flags potentially duplicated records. If genuine duplicates are uncommon, false alarms among otherwise unique entries may make up much of the flagged list.
The same base-rate principle applies. However, the consequences and appropriate checks vary by domain; one should not copy an educational threshold into health, security or financial settings without domain-specific validation.
30. Why overall accuracy can hide problems
In population A, 900 of one thousand answers are correct. A naive checker that passes everything would be correct 900/1,000 = 90% of the time, slightly higher than the invented checker’s 89% classification accuracy. But it would detect zero genuine errors.
This comparison illustrates that the best measure depends on the task’s purpose. If the point is to find erroneous answers for review, detecting real errors matters, as do false-alarm costs and appropriate user safeguards.
31. What a complete model report should include
Name the positive class, the population, the base rate, the number of true and false positives and negatives, and the exact measure being reported. Distinguish performance on an evaluation set from forecasts about new inputs.
Those details let the reader independently reconstruct precision, recall, specificity and accuracy. A single unexplained 80% figure does not.
32. The fictional baseline error
Imagine an assistant saying, “Because the checker detects eighty per cent of mistakes, every flagged answer has an eighty per cent chance of being wrong.” This is a classic conditional-probability reversal.
A repaired answer begins with the complete two-by-two table, counts all flags, and reports eighty genuine errors out of 170 flags. The correction is not a change in tone: it is a different denominator.
33. A practice question for pupils
Suppose another invented tool finds eighteen out of twenty faulty items, while it also raises twelve false alarms on correct items. What fraction of its thirty flags are genuine faults? Eighteen divided by thirty equals 60%.
The sensitivity is 18/20 = 90%, which is not the same as the flagged-case precision of 60%. This miniature exercise repeats the key distinction with simpler counts.
34. A simple way to teach the concept
Draw a large square containing one thousand dots, colour one hundred as wrong and nine hundred as correct, then mark eighty wrong dots and ninety correct dots with a warning. Ask the pupil to count dots inside the warning group.
This physical grouping makes the conditional direction visible. The answer changes depending on which set forms the denominator, so the student learns to identify the group before applying division.
Frequently asked questions
Does 80% detection mean 80% of flags are correct? No. The first is sensitivity; the second is positive predictive value or precision.
What is a base rate? The prevalence of the target condition in the relevant population.
Can the same checker have different precision in two populations? Yes, even if sensitivity and false-positive rate stay constant.
What does a false positive mean here? A correct answer was flagged as though it were wrong.
Does a flag prove a pupil made a mistake? No. The work needs suitable verification.
Why does the denominator change? Conditioning on wrong answers and conditioning on flags select different sets of cases.
Does a worked example prove real AI reliability? No. The numbers here are fictional and designed for learning.
The central lesson
Conditional probability is about which event has already been assumed. The chance of a warning when an answer is wrong is not the chance an answer is wrong when a warning appears. Base rates connect the two and can change the meaning of a prediction dramatically.
Before accepting an AI percentage, ask “Out of which group?” Then build the two-by-two table and calculate from the correct denominator. That one habit prevents a common and consequential misunderstanding.
Connected learning and primary references
- How Super Intelligence Works — complete learning curriculum.
- False Positives and False Negatives — four outcome types.
- Probability and Confidence — why a percentage is not a promise.
- Sampling Bias — which population was tested.
- OpenStax: Principles of Data Science, Probability Theory — conditional probability and Bayes’ theorem.
- Scikit-learn: Model Evaluation — precision, recall, confusion matrices and the effects of prevalence.
- NIST AI Risk Management Framework — evaluating systems for their actual intended context.