How does superintelligence work when an AI reads a table? A number inside a table rarely means much on its own. The number 15 might be a Mathematics score, an English score, a week number or the quantity of books returned. Its meaning depends on the row, column, headings and units surrounding it. If an AI loses even one heading, a perfectly copied number can become a false statement.
The practical mechanism is table-structure interpretation. An AI assistant must recognise which labels apply to each cell, keep rows aligned, distinguish subtotals from individual values and preserve missing or qualified information. The task is different from simply reading every number in order.
This easy guide follows an invented four-pupil quiz record with two subjects and two weekly assessments. We will calculate changes, test twenty statements and examine the mistakes caused by merged headings, lost units and copied rows. These records are fictional teaching examples rather than actual student results or a benchmark of any AI model. “How Superintelligence Works” is the name of an educational series, not evidence that present systems have attained artificial superintelligence.
1. A table is more than numbers arranged neatly
Consider the text “A 12 15 16 17”. Without labels, the reader cannot know what A stands for or what the four numbers measure. The same values could describe test scores, attendance or lengths of time. A table supplies structure so those numbers can be interpreted consistently.
In a good table, the heading above a column tells us what its values represent. The label at the start of a row identifies the entity being measured. A caption can explain the entire table’s purpose, date and units.
2. The relationship between headings and cells
If a cell lies in the row labelled Pupil A and the column labelled English, Week 2, its value is about that pupil’s Week 2 English result. The neighbouring Mathematics cell may show a different subject, even if its number happens to be identical.
An AI summariser that extracts values without their associated headings may produce a convincing answer about the wrong subject. The basic reliability test is whether the claimed number can be traced back to a unique row and column combination.
3. Grouped headings add another layer
A table may have English above two columns labelled Week 1 and Week 2, and Mathematics above another pair with the same week labels. In that layout, the word Week 2 appears twice. The subject heading determines which Week 2 cell is intended.
Flattening the table into a line such as “Week 1, Week 2, Week 1, Week 2” is insufficient unless the original group labels remain attached. Multi-level headers are one reason document extraction needs structural care.
4. Units are part of the meaning
Suppose the caption says every result is a score out of twenty marks. The number fifteen is then fifteen marks, not fifteen per cent. Reporting it as a percentage would require an explicit conversion: 15/20 = 75%.
An AI must not silently change units or invent a denominator. A quiz score, a duration in minutes and an attendance percentage may all contain the number fifteen while describing entirely different quantities.
5. The fictional Cedar Quiz Table
Imagine four fictional pupils identified only as A, B, C and D. Each completed one English quiz and one Mathematics quiz in Week 1, then another quiz in both subjects in Week 2. Every quiz is marked out of twenty. The following values are invented solely for teaching table interpretation.
The source does not describe the difficulty of the questions, the teaching method used between weeks or any pupil’s personal circumstances. Therefore a score change can be calculated, but a causal claim about the effectiveness of a lesson cannot be verified from this table.
6. The complete labelled record
- Column order: Pupil | English Week 1 | English Week 2 | Mathematics Week 1 | Mathematics Week 2.
- Pupil A: A | 12 | 15 | 16 | 17.
- Pupil B: B | 14 | 14 | 13 | 16.
- Pupil C: C | 10 | 13 | 11 | 13.
- Pupil D: D | 16 | 18 | 15 | 15.
- Unit for every score: marks out of 20.
- Important: the repeated Week 1 and Week 2 labels belong to different subject groups.
- Coverage: four pupils, four observed scores per pupil, sixteen numeric entries.
7. Read one cell with all its labels
A’s third numeric entry is sixteen. It represents the pupil’s Mathematics Week 1 score because of its position under the Mathematics group and first week subheading. It is not A’s second-week English mark, even though both are numbers in the same row.
To verify a statement, name all three parts: pupil, subject and week. “A scored sixteen” is incomplete; “Pupil A scored sixteen out of twenty in Mathematics Week 1” is specific enough to check.
8. Calculate English changes per pupil
English Week 2 minus English Week 1 gives the observed change. A: 15 − 12 = +3. B: 14 − 14 = 0. C: 13 − 10 = +3. D: 18 − 16 = +2.
Three pupils recorded higher English scores in Week 2; one recorded the same score. No one recorded a lower English score in the invented figures. These are numerical comparisons, not verified explanations of why the scores changed.
9. Calculate Mathematics changes per pupil
Mathematics changes are A: 17 − 16 = +1, B: 16 − 13 = +3, C: 13 − 11 = +2 and D: 15 − 15 = 0.
Again, three pupils have higher second-week scores and one is unchanged. The identity of the unchanged pupil differs between subjects: B is unchanged in English, while D is unchanged in Mathematics. An AI that merges the two subject columns could lose that distinction.
10. Sum each of the four columns
English Week 1 totals 12 + 14 + 10 + 16 = 52 marks. English Week 2 totals 15 + 14 + 13 + 18 = 60 marks. Mathematics Week 1 totals 16 + 13 + 11 + 15 = 55 marks. Mathematics Week 2 totals 17 + 16 + 13 + 15 = 61 marks.
These totals are useful checks for copied data. English totals increase by eight marks across the four pupil results, while Mathematics totals increase by six. The aggregate differences must equal the sums of the individual differences.
11. Check the averages
Divide each subject-week total by four because all four fictional pupils have observed scores. English averages are 52/4 = 13 for Week 1 and 60/4 = 15 for Week 2. Mathematics averages are 55/4 = 13.75 and 61/4 = 15.25.
If a real table had missing scores, the denominator might differ between weeks or subjects. Using four automatically could then be wrong. In this complete toy record, four is the correct divisor for every column.
12. A calculated scorecard
- English Week 1 total: 52; mean: 13/20.
- English Week 2 total: 60; mean: 15/20.
- English total change across four pupils: +8.
- Mathematics Week 1 total: 55; mean: 13.75/20.
- Mathematics Week 2 total: 61; mean: 15.25/20.
- Mathematics total change across four pupils: +6.
- Pupils with higher English second-week scores: A, C, D — three.
- Pupils with higher Mathematics second-week scores: A, B, C — three.
- Unchanged English score: B.
- Unchanged Mathematics score: D.
- Number of observed numeric cells: 4 pupils × 4 score columns = 16.
13. Two totals can both be correct and mean different things
The total of sixty is the sum of Week 2 English marks across four pupils. It is not one student’s score out of twenty, nor a measured percentage. The average of fifteen is a different statistic derived from the same four values.
An assistant must name whether a number is an individual value, a sum, an average or a rate. Copying “60” correctly but describing it as the class average would still be a serious interpretation error.
14. Why a table caption matters
A caption such as “Quiz scores out of twenty marks, Weeks 1 and 2” tells the reader which unit and period apply to every score. Without it, even clearly arranged rows may require guessing about the marking scale.
Captions also assist accessibility. The W3C Web Accessibility Initiative recommends clear table structures so people using assistive technology can identify which headings apply to which data cells. The benefit extends to software agents that need to parse the same relationships.
15. What happens when a row shifts one column?
Suppose A’s first English value of twelve disappears during extraction but the remaining values slide one position to the left. The line might then look like “A 15 16 17”. A model could misinterpret fifteen as English Week 1 and sixteen as English Week 2.
The numbers would still look plausible. Detecting the shift requires checking column count and alignment, not simply asking whether every value lies between zero and twenty. Structural validation catches errors that ordinary arithmetic plausibility tests miss.
16. Why merged headings can be lost in conversion
A website might use one visible heading spanning two week columns. Converting it to plain text without the group association may leave only “Week 1” and “Week 2”, repeated for both subjects. The same problem can occur in PDF extraction, copied spreadsheets and OCR.
Good parsers preserve the header hierarchy or transform it into fully qualified column names such as “English — Week 1” and “Mathematics — Week 1”. That creates an auditable connection between values and their meaning.
17. Footnotes can change the value’s interpretation
A score marked with an asterisk might have been adjusted, provisional or based on a different assessment version. A table’s footnote may specify that a number excludes absentees or uses a changed denominator.
An AI should not discard such notes as formatting decoration. If the note qualifies a row or column, the resulting summary must preserve it. A mathematically correct mean can be misleading when it combines differently defined values.
18. Twenty statements to classify
The Cedar desk asks a fictional assistant to label statements as SUPPORTED, CONTRADICTED or NOT ESTABLISHED based solely on the complete toy table. A supported claim follows from the scores. A contradicted claim conflicts with them. A not-established claim requires evidence not provided.
This test helps readers see the boundary between extracting a number correctly and making a stronger claim about learning or participants’ background.
- 01. “A scored 12 in English Week 1.” — SUPPORTED.
- 02. “A scored 17 in Mathematics Week 2.” — SUPPORTED.
- 03. “B’s English scores were equal in both weeks.” — SUPPORTED.
- 04. “C’s English result increased by three marks.” — SUPPORTED.
- 05. “D’s Mathematics result increased.” — CONTRADICTED: 15 became 15.
- 06. “Most pupils scored higher in English Week 2.” — SUPPORTED: three out of four.
- 07. “All four pupils improved in Mathematics.” — CONTRADICTED: D was unchanged.
- 08. “B scored 16 in Mathematics Week 2.” — SUPPORTED.
- 09. “A scored 17 in Mathematics Week 1.” — CONTRADICTED: it was 16.
- 10. “The English Week 2 total was 60 marks.” — SUPPORTED.
- 11. “The Mathematics Week 2 total was 61 marks.” — SUPPORTED.
- 12. “The English Week 1 mean was 13.” — SUPPORTED.
- 13. “The Mathematics Week 2 mean was 15.25.” — SUPPORTED.
- 14. “C’s Mathematics Week 2 result exceeded C’s English Week 2 result.” — CONTRADICTED: both were 13.
- 15. “D scored 18 in English Week 2.” — SUPPORTED.
- 16. “The four English scores increased by eight marks in total.” — SUPPORTED.
- 17. “The higher second-week English mean proves a particular tutoring method caused learning.” — NOT ESTABLISHED.
- 18. “All four pupils attended the same class together.” — NOT ESTABLISHED.
- 19. “The English Week 2 mean was 15.” — SUPPORTED.
- 20. “B scored 13 in Mathematics Week 1.” — SUPPORTED.
19. Reconcile the twenty outcomes
Supported statements are 01, 02, 03, 04, 06, 08, 10, 11, 12, 13, 15, 16, 19 and 20: fourteen. Contradicted statements are 05, 07, 09 and 14: four. Not-established statements are 17 and 18: two. Fourteen plus four plus two equals twenty.
The final two test whether the model knows what the table does not contain. It records scores, not intervention design or classroom membership. A model should not turn a score improvement into proof of a specific cause.
20. How to inspect a table manually
First read the caption. Second identify each row label. Third trace the complete heading above each value. Fourth identify units and denominators. Fifth check for notes and missing cells. Sixth perform any calculation with explicit row and column references.
A good reading habit is to say the full meaning of one cell aloud. If you cannot explain which pupil, subject, week and scale the cell represents, do not calculate from it yet.
21. A table can contain more than one denominator
Imagine the Week 1 Mathematics score came from a twenty-mark quiz but the Week 2 assessment was out of thirty. A direct comparison of raw marks would then be misleading. Fifteen out of thirty and fifteen out of twenty are not the same level of performance.
In our invented table, every quiz is explicitly out of twenty. That common denominator makes raw score changes comparable as numbers, although assessment difficulty and content still limit any broader learning interpretation.
22. Why an absent cell is not zero
Suppose C has no Week 2 English score because the quiz was not taken. Replacing the blank with zero would turn a missing result into an observed failure, artificially lowering the total and mean.
A robust AI keeps missing, not applicable and measured zero distinct. If a score is absent, it should report the observed denominator or seek the real value, not silently invent one.
23. Row labels can be ambiguous too
Some spreadsheets contain a row called “Total” alongside actual pupil rows. If an AI includes that total row when calculating another total, it double-counts the values. Similarly, subtotal rows should not be treated as new people.
Before aggregation, classify each row’s role: ordinary observation, subtotal, grand total, heading or note. That distinction is as important as reading column headings correctly.
24. Charts inherit errors from tables
A graph built from a misread table can look professional while displaying the wrong series. If English and Mathematics week labels are swapped, the chart may attribute an improvement to the wrong subject.
The repair must start in the source table and transformation process. Making the chart more attractive does not validate its numbers. Every plotted point should trace back to its correct labelled cell.
25. Why accessibility helps machine reading
W3C guidance recommends markup that clearly associates table headings with data cells. For complex tables, headers may need explicit relationships using scope or header IDs. These structures help screen-reader users and can also support more reliable software extraction.
A plain-text screenshot or badly flattened copy may lose that structure. AI applications should prefer structured source data when available rather than inventing table boundaries from a confusing text stream.
26. A fictional baseline’s mistakes
Imagine a made-up AI assistant correctly classifies sixteen of the twenty claims but confuses A’s two Mathematics week columns, treats D’s unchanged result as an increase, loses the meaning of “most” and asserts that the improved scores prove a tutoring intervention caused the change.
That is an illustrative 16/20 = 80%, not a real model measurement. The errors arise from different causes: column alignment, comparison logic, quantifier interpretation and unsupported causal inference. They need different checks.
27. The revised fictional method
Require the assistant to reconstruct fully qualified column labels, verify row lengths, identify units and distinguish descriptive statistics from causal claims. It must return a source cell for each number and mark unsupported interpretations as unknown.
Suppose that revised process classifies all twenty fictional statements correctly. A local 20/20 result would illustrate the value of the workflow, not prove that any deployed AI handles every complex table. Independent evaluation remains necessary.
28. A transfer exercise on a timetable
A timetable contains columns Monday, Tuesday and Wednesday, with rows for different classes. The value “Science” in Tuesday’s row for Class B does not mean every class has Science on Tuesday. The row label restricts the statement.
Now imagine the timetable includes a merged heading “Weeks 1–4”. That date range is part of the scope. A system must not apply the timetable to Week 5 without separate evidence.
29. A transfer exercise on an expenses table
Suppose a table shows price per item and number of items. A line with $8 and quantity 3 describes a line total of $24, not $11 or $8 for all three together. The headings determine which arithmetic operation is appropriate.
An AI may know multiplication but still misread the table. The important skill is connecting numbers to the correct roles before calculating.
30. What to include in a verification packet
Keep the source file or page location, table caption, heading hierarchy, row IDs, units, any footnotes, extracted cell values, formulas and final reported numbers. When the source changes, check dependent calculations again.
This level of traceability lets reviewers distinguish a data-extraction error from a calculation error. Without it, a wrong result may be blamed vaguely on “AI reasoning” even when the original table was flattened incorrectly.
31. What a responsible summary sounds like
A grounded summary of the toy data says: “Across four fictional pupils, average English scores rose from 13 to 15 out of twenty, and average Mathematics scores rose from 13.75 to 15.25.” It also notes that the table does not establish why those changes occurred.
That statement preserves subjects, periods, population and units. It is more useful than saying “All scores went up” or “The tuition worked”, neither of which follows from the data.
Frequently asked questions
Why can an AI copy a table number and still be wrong? It may attach the number to the wrong row, heading, unit or statistic.
What is a grouped header? A higher-level heading that applies to several subordinate columns, such as English above Week 1 and Week 2.
Should an empty cell count as zero? Not unless the source explicitly defines it as an observed zero.
Can a higher average prove a teaching method caused improvement? No. Causal claims require additional study design and evidence.
Why are captions and header markup useful? They preserve context and help people and software interpret cells correctly.
Can a model infer every merged-cell structure from text? Not reliably; structured source data and validation are preferable.
The central lesson
A table’s intelligence is in its relationships: one number belongs to a row, a column, a unit and a source. Strip away those relationships and the number becomes ambiguous even if the digits survive perfectly.
Before trusting an AI’s numerical summary, insist on the full label, inspect the original cell and reconcile the calculation. That simple discipline turns table reading into a precise, checkable form of reasoning.
Connected reading and authoritative references
- How Super Intelligence Works — educational curriculum.
- Quantifiers — how words such as most and all affect interpretation.
- Unknown, None and Zero — preserving empty and unmeasured cells.
- Measurements and Units — interpreting scales and numerical operations.
- W3C Web Accessibility Initiative: Tables Tutorial — table headings, captions and data relationships.
- W3C Tables with Two Headers — explicit row and column associations.
- MDN: Table Accessibility — source structure and the scope attribute.