How does superintelligence work when an AI makes one tiny mistake at the beginning of a long calculation? The final answer may contain several wrong numbers even though every later arithmetic step is correct. If an assistant misreads eighteen pupils as twenty-eight, it can calculate the wrong total, print quantity, paper consumption and budget without making another multiplication error.
This mechanism is error propagation: one incorrect input or interpretation changes the results of dependent operations. A useful AI system must trace how claims depend on other claims, identify the point where the first incorrect value entered and recheck everything downstream when the source is corrected. A coherent explanation is not enough when its starting data are wrong.
This easy guide follows a fictional school-workbook printing plan through an original source record, an incorrect transcription, ten affected quantities and a completed correction audit. The invented counts are for teaching, not real school records or model measurements. The name “How Superintelligence Works” refers to an educational series on AI mechanisms; it does not claim current products are artificial superintelligence.
1. Why one error can produce many errors
Imagine a recipe whose first line says two cups of water but is accidentally copied as three. A cook who doubles the recipe using the incorrect figure can multiply accurately and still produce the wrong mixture. The initial mistake influences every later step that depends on that quantity.
The same pattern appears in spreadsheets, AI-generated summaries, financial models, school assessments and research reports. The visible error at the end may be a cost or total, but the actual cause can lie much earlier in source extraction.
2. An error is not always arithmetic
Some mistakes are numerical: 18 becomes 28, or a decimal point moves. Others are semantic: “number of pupils” is mistaken for “number of lessons”, or “pages” is confused with “sheets of paper”. Some result from using an outdated version of a source.
An AI that only checks its arithmetic may miss semantic and extraction errors completely. Reliable checking includes the meaning of the input, its origin and the operations performed on it.
3. The dependency-chain idea
A dependency chain represents how one result uses earlier values. Number of pupils determines the number of booklets needed; booklets determine printed pages; printed pages determine sheets; sheets determine an estimated printing cost.
If the pupil count changes, the downstream results should be recalculated. If the paper cost changes but enrolment is unchanged, the upstream counts should not necessarily change. Understanding the direction of dependency makes repairs more focused.
4. The fictional Cedar Workbook Project
Imagine Cedar School preparing revision booklets for three fictional, non-overlapping student groups. The English group has eighteen pupils, Mathematics has twelve and Science has ten. Each pupil receives one booklet, and the desk prepares five extra spare booklets.
Every booklet contains four printed pages. Printing is double-sided, so one physical sheet holds two pages in this simplified format. The invented paper-and-print charge is $0.08 per sheet. There are no covers, binding charges, taxes or additional wastage in the teaching calculation.
5. Freeze the correct source record
- English group: 18 pupils.
- Mathematics group: 12 pupils.
- Science group: 10 pupils.
- Groups are mutually exclusive in this example.
- One booklet is needed for each pupil.
- Extra spare booklets: 5.
- Printed pages in each booklet: 4.
- Double-sided printing: 2 pages per sheet.
- Cost in the fictional model: $0.08 per physical sheet.
- No additional costs or print waste are included.
6. Compute the correct student total
Add the three non-overlapping group counts: 18 + 12 + 10 = 40 pupils. The rule that groups are non-overlapping matters. If pupils could belong to two groups, adding all memberships would not necessarily give the number of unique people.
Here the source explicitly states that the groups do not overlap, so the arithmetic is a valid total. The AI should cite the three original figures rather than generating a convenient round number from memory.
7. Compute the correct number of booklets
Forty pupils need forty copies. Add five spares: 40 + 5 = 45 booklets. This intermediate result is a count of physical booklets, not a count of printed pages or paper sheets.
If the assistant instead reports 40 as the complete print run, it has omitted the spare requirement. Even a correct student total is not automatically the correct production total.
8. Convert booklets to printed pages
Each booklet contains four printed pages. Forty-five booklets therefore require 45 × 4 = 180 printed pages. The numerical result is larger because each physical booklet contains several page images.
Pages and paper sheets are different. An AI that reports 180 sheets at this stage ignores the double-sided requirement and overstates physical paper consumption.
9. Convert printed pages to sheets
With two printed pages per sheet, 180 pages require 180 ÷ 2 = 90 physical sheets in the toy model. Each booklet uses exactly two sheets because its four pages fit evenly across double-sided paper.
There is no need for additional rounding in this case. A different page count—for example five pages per booklet—would require more careful per-booklet sheet calculation rather than simply dividing the total pages by two if booklets must be printed separately.
10. Compute the correct printing cost
At $0.08 per physical sheet, ninety sheets cost 90 × $0.08 = $7.20. The result answers the invented paper-and-print charge only. It is not a real price quote from a printer.
The complete correct chain is 40 pupils → 45 booklets → 180 pages → 90 sheets → $7.20. Every arrow depends on the earlier definitions. Keeping the chain visible makes the work easier to audit.
11. The first transcription mistake
Suppose an AI assistant reads the original enrolment table and mistakenly copies the English group as 28 rather than 18. Mathematics remains twelve, Science remains ten and all print specifications stay unchanged.
The mistake adds ten imaginary pupils. No actual enrolment change occurred. The corrupted value now enters an otherwise mathematically consistent spreadsheet or planning calculation.
12. The wrong total still looks reasonable
The assistant computes 28 + 12 + 10 = 50. This addition is arithmetically correct for the corrupted input. That is why arithmetic checking alone may not reveal the problem.
If a reviewer asks only whether the three displayed numbers add to fifty, the answer is yes. But the question that matters is whether the displayed English number matches the authorised source. It does not.
13. Follow the contaminated count into printing
The wrong fifty-pupil total produces 50 + 5 = 55 booklets. The model then calculates 55 × 4 = 220 printed pages, 220 ÷ 2 = 110 sheets, and 110 × $0.08 = $8.80.
Every operation after the copied count can be internally correct while the planned output remains wrong. The final budget difference is $8.80 − $7.20 = $1.60, caused by the single source error of ten extra pupils.
14. Correct versus corrupted result ledger
- English pupils: correct 18; corrupted 28; difference +10.
- Mathematics pupils: correct 12; corrupted 12; difference 0.
- Science pupils: correct 10; corrupted 10; difference 0.
- Total pupils: correct 40; corrupted 50; difference +10.
- Booklets including five spares: correct 45; corrupted 55; difference +10.
- Printed pages: correct 180; corrupted 220; difference +40.
- Physical sheets: correct 90; corrupted 110; difference +20.
- Printing charge: correct $7.20; corrupted $8.80; difference +$1.60.
- Source of the difference: the English pupil count was copied incorrectly.
- All later numerical conversions used the unchanged fictional print policy.
15. Why the cost error is not the root cause
A budget checker might identify the incorrect $8.80 total and ask for recalculation. But calculating 110 × $0.08 again still gives $8.80. The cost calculation is correct conditional on the wrong 110-sheet input.
Changing the final figure by hand to $7.20 could make one visible line accurate while leaving the worksheet count and paper order wrong. The repair must begin where the incorrect enrolment value entered.
16. Why an internally balanced total can be wrong
Forty-five booklets at four pages each yield 180 pages, which at two per sheet yield 90 sheets. The corrupted version is equally self-consistent: 55 × 4 = 220 pages, and 220 ÷ 2 = 110 sheets.
Both sets satisfy their internal arithmetic identities. Therefore a consistency check is necessary but insufficient. The numbers must also be validated against the correct input record. A self-consistent fantasy is still a fantasy.
17. Trace every result to its source
A trustworthy calculation can preserve a small provenance chain. The total pupils cell depends on three group-count source cells. Booklet count depends on the total and spare rule. Sheet count depends on booklet count and layout specifications. Cost depends on sheet count and rate.
With those links, a corrected English count can trigger precisely the dependent recalculations. Reviewers can also see that Mathematics and Science counts did not need to change.
18. Recalculation after the repair
Replace the corrupted 28 with the source-verified 18. Recalculate in order: 18 + 12 + 10 = 40; 40 + 5 = 45; 45 × 4 = 180; 180 ÷ 2 = 90; 90 × $0.08 = $7.20.
The corrected values agree with the original reference ledger. A final source check should confirm that the saved output contains the repaired numbers, rather than merely the assistant’s claim that recalculation occurred.
19. Twenty claims for an AI to classify
The Cedar desk asks a hypothetical classifier to use SUPPORTED, CONTRADICTED or NOT ESTABLISHED. Some statements concern the correct source, some describe the corrupted report as a report, and some claim things the record cannot tell us.
Careful wording matters. “The corrupted report says fifty” is true as a fact about the report. “Fifty pupils were enrolled” is contradicted by the original authoritative toy record.
- 01. “The source records eighteen English pupils.” — SUPPORTED.
- 02. “The source records twenty-eight English pupils.” — CONTRADICTED.
- 03. “The correct three-group total is forty.” — SUPPORTED.
- 04. “The corrupted spreadsheet’s displayed total is fifty.” — SUPPORTED.
- 05. “Fifty pupils actually enrolled in this toy record.” — CONTRADICTED.
- 06. “The correct print run needs forty-five booklets.” — SUPPORTED.
- 07. “The corrupted print run calculates fifty-five booklets.” — SUPPORTED.
- 08. “The correct page count is 180.” — SUPPORTED.
- 09. “The corrupted page count is 220.” — SUPPORTED.
- 10. “The correct sheet count is ninety.” — SUPPORTED.
- 11. “The corrupted sheet count is 110.” — SUPPORTED.
- 12. “The correct print charge is $7.20.” — SUPPORTED.
- 13. “The corrupted calculation produces an $8.80 charge.” — SUPPORTED.
- 14. “The two charges differ by $1.60.” — SUPPORTED.
- 15. “Rechecking multiplication alone proves the copied enrolment was correct.” — CONTRADICTED.
- 16. “The error affects only the final money total.” — CONTRADICTED.
- 17. “The original English count differs from the copied count by ten.” — SUPPORTED.
- 18. “The original Mathematics count was changed.” — CONTRADICTED.
- 19. “The data identify which specific ten pupils were added.” — NOT ESTABLISHED.
- 20. “The report confirms that the copier intentionally altered the count.” — NOT ESTABLISHED.
20. Reconcile the claim labels
Supported: 01, 03, 04, 06, 07, 08, 09, 10, 11, 12, 13, 14 and 17 — thirteen. Contradicted: 02, 05, 15, 16 and 18 — five. Not established: 19 and 20 — two. Thirteen plus five plus two equals twenty.
The last two statements are particularly important. Finding an error does not establish anybody’s intent or identify real individuals who were supposedly added. An AI should diagnose the data problem without inventing blame.
21. Error propagation can amplify more than money
If the corrupted report were published, it might lead to ordering extra materials, changing staff allocations, producing misleading attendance statistics and briefing parents using the wrong class size.
Those are possible consequences of using the report, not outcomes proved by the fictional ledger. The ledger establishes the calculations. It does not establish that any real order was placed or notification sent.
22. Rounding can also propagate
Suppose an early step rounds a measured value from 2.49 to 2 and every later multiplication uses the rounded number. The final error may become much larger than the original difference. Rounding should be delayed appropriately when precision matters.
Not every numerical model needs maximal precision. The important point is to understand when a value was rounded, how the rounding affects later steps and which final precision is justified by the original measurement.
23. Errors can travel through summaries
Imagine a source states “approximately forty participants”. A summary turns that into “exactly forty participants”, and a later article cites the summary as an exact count. A qualification has been lost before any arithmetic starts.
This is semantic error propagation. The model may preserve the number but strengthen the claim. Source citations should attach to the original wording, including approximation and time boundaries, not merely to the nearest supporting topic.
24. How feedback can help repair the right layer
A reviewer can say, “The English count should be eighteen, not twenty-eight.” That feedback identifies the earliest incorrect field. The system can then repair the field and recalculate its dependent outputs.
If the reviewer simply says “the budget looks wrong”, the assistant must diagnose which input or formula failed. A useful correction workflow returns to the source rather than forcing the final number to match the reviewer’s expectation.
25. A regression can appear during repair
Changing the English count could accidentally overwrite the Science count or break a formula reference. That would create a new mistake while resolving the original one. The correct repair includes regression checks for unchanged fields and totals.
Therefore the reviewer should confirm not only that the reported cost is now $7.20, but that the source values, booklet count, page count and sheet count still reconcile.
26. What is an invariant in this scenario?
An invariant is a condition expected to remain true through valid transformations. Here, total pupils equal the sum of the three non-overlapping group counts. Number of booklets equals pupils plus five spares. Pages equal four per booklet. Sheets equal two per booklet.
These relations can reveal mismatched formulas or inconsistent updates. They cannot independently prove that the first English number was copied correctly. Invariants and source verification solve different parts of the problem.
27. A dependency-aware test plan
After changing the English count, test the group total, print run, page calculation, sheet calculation and cost. Also test the unchanged Maths and Science values. Check that user-facing explanatory text uses the same numbers.
A test that checks only the corrected field can miss stale values elsewhere in a generated report. The larger the content network, the more important it becomes to know which outputs depend on each input.
28. A transfer exercise with test averages
Four pupils score 12, 16, 18 and 14 marks. The total is 60 and the mean is 15. Suppose 18 is mistakenly copied as 28. The total becomes 70 and the mean becomes 17.5. The average calculation is internally correct for the wrong copied data.
The fix is to check the source score of eighteen, restore it and recompute. Changing only the mean back to fifteen would leave an inconsistent total and an incorrect individual score.
29. A transfer exercise with a recipe
A recipe calls for 250 millilitres of water per batch. For three batches, that is 750 millilitres. If an AI misreads the source as 350 millilitres per batch, it calculates 1,050 millilitres correctly from the corrupted input.
The difference of 300 millilitres is a propagated source mistake. It is not evidence that the multiplication algorithm itself is wrong. Always separate input interpretation from computation.
30. How to teach the dependency chain
Write the source record on one card and each derived result on its own card. Draw arrows showing which earlier result each card needs. Change one source number and ask pupils to identify which downstream cards must be recalculated.
Then change only the cost per sheet. The pupil total and number of sheets should remain unchanged, while the budget changes. This exercise teaches targeted repair rather than indiscriminate rewriting.
31. Error severity is not just error size
A ten-pupil transcription difference may be more consequential than a small currency difference if it changes staffing or access decisions. Conversely, a single wrong recipient address could be important even though its numerical difference is not measurable.
Severity depends on what the value controls. A good system records the nature of the error, its dependent actions and whether external effects occurred. It should not prioritise errors solely by the size of the final number.
32. What an AI should say when checking failed
If the authorised enrolment source cannot be accessed, a model should not state that the copied figures were independently verified. It can explain the arithmetic conditional on the supplied numbers and mark the input check as incomplete.
An honest interim answer is more useful than a false claim that a complete audit occurred. Source verification is an action with evidence, not a decorative sentence.
33. A realistic repair protocol
First freeze the reported error and source version. Second compare the derived field with the authoritative original. Third locate the earliest incorrect value or relation. Fourth repair that point. Fifth recompute the dependency chain. Sixth run independent checks. Seventh confirm that any saved or published report actually contains the repaired values.
That sequence connects plain-language reasoning with data processing and responsible AI operation. Every stage has a distinguishable output, so a reviewer can see whether the repair is complete or only proposed.
Frequently asked questions
Can later arithmetic be correct when the final answer is wrong? Yes, if an earlier input or interpretation is wrong.
What is the root cause? The earliest relevant mistake that explains the downstream errors, such as the copied English count in this example.
Is an internally consistent table always accurate? No. Incorrect inputs can produce a consistent but false report.
How do you prevent propagation? Verify source values, keep dependency links and recompute affected outputs after corrections.
What is a regression? A newly introduced mistake caused by a change intended to repair something else.
Does error propagation prove bad intent? No. A transcription error alone does not establish motive.
Can AI check its own outputs? It can assist, but independent source and deterministic checks are valuable because the model may share the same mistaken assumptions.
The central lesson
One wrong input can create a whole family of wrong outputs. The final budget is often just the visible symptom. Reliable AI reasoning traces conclusions backward to their sources, repairs the earliest error and verifies every result that depended on it.
Good mathematical working is therefore not only about doing the operation correctly. It is also about asking whether the numbers represented the correct things in the first place, and whether the final answer still connects to the original evidence.
Related lessons and sources
- How Super Intelligence Works — main curriculum.
- Multi-Step Word Problems — understanding a story before calculating.
- Tables, Rows and Columns — source labels and extraction errors.
- Feedback Loops — repairs and regressions.
- NIST AI Risk Management Framework — validation and trustworthy use.
- W3C PROV-O — modelling information provenance.
- Microsoft Excel Support — spreadsheet references and recalculation.