How does superintelligence work when an AI answer contains a number that looks right but uses the wrong unit? An assistant can write “The journey takes 40 kilometres” or “The bottle contains 750 kilograms” in beautiful English, yet neither sentence answers a sensible time or volume question. Good numerical reasoning requires more than multiplying the digits. It must preserve what the quantity represents.
The practical mechanism is unit-aware reasoning: reading a measurement as a number with a dimension and a unit, translating between compatible units, checking the operation and reporting the result at appropriate precision. It matters in Mathematics, Science, engineering, travel, cooking, health information, school work and automated calculations.
This beginner-friendly guide develops a fictional school-garden project with twelve calculations that can be checked independently. The examples are invented, not measurements of real pupils or a benchmark of a named model. The series title “How Superintelligence Works” describes a learning programme, not an assertion that existing AI has achieved artificial superintelligence.
1. Numbers never tell the whole story
Compare “The path is 300” with “The path is 300 metres”. The first sentence gives a number with no interpretable measurement. The second tells us length. “300 seconds”, “300 millilitres” and “300 grams” are different quantities, even though the digits are identical.
A language model can learn to predict plausible unit words, but plausibility is not a substitute for explicit measurement data. A reliable system keeps the number, unit and measured property together from source to answer.
2. A unit and a dimension are not quite the same
Length is a physical dimension. Metre, centimetre and kilometre are units for length. Time is another dimension; seconds, minutes and hours can measure it. Mass is another dimension; kilogram and gram are related units. Two different units may represent the same dimension, making conversion possible.
By contrast, metres and kilograms do not measure the same kind of quantity. You cannot add three metres to two kilograms and obtain five of anything meaningful. The arithmetic 3 + 2 = 5 is valid as pure numbers; the physical expression is not. That is the first useful check an AI-assisted calculation should make.
3. SI is a reference system, not a magic calculator
The International System of Units, abbreviated SI, is the internationally agreed practical measurement system. The metre, kilogram and second are among its base units. Other familiar units, such as litres and minutes, are accepted for use with SI but are not themselves SI base units.
A system still needs to understand the task. Using a standard unit does not tell you whether a distance is a straight line or along a road, whether a mass was measured accurately, or whether the user means a date, a time interval or a duration. Standardisation prevents one class of confusion while leaving others to be solved.
4. The ordinary conversion chain
One kilometre is 1,000 metres. One metre is 100 centimetres. One kilogram is 1,000 grams. One litre is 1,000 millilitres. One hour is 60 minutes, and one minute is 60 seconds. These relationships are useful, but the conversion direction is easy to reverse when a student rushes.
When converting from a larger unit to a smaller unit, the numerical value usually becomes larger; a two-kilometre path is 2,000 metres. When converting from a smaller unit to a larger one, the numerical value becomes smaller; 750 centimetres is 7.5 metres. The physical quantity remains unchanged.
5. Dimensional consistency catches impossible answers
Suppose a pupil asks for the speed of a cyclist who travels 600 metres in 120 seconds. Speed is distance divided by time, so its derived unit is metres per second. The calculation is 600 ÷ 120 = 5 metres per second.
If the assistant answers “5 metres”, the number is right but the quantity is wrong. If it answers “5 seconds per metre”, it has inverted the unit. Unit structure lets a reviewer find an error even before checking a numerical result.
6. Establish the fictional garden project
Imagine the Cedar School Gardening Club preparing simple notes for a science display. Its records include path lengths, seed weights, water volumes, travel times and percentages. The club has twelve conversion questions, each using invented but deliberately consistent numbers.
The objective is to show why an AI should parse the measure, check the desired unit and calculate rather than improvise a convincing-looking sentence. No actual garden data, safety procedures or scientific findings are implied.
7. Twelve checked measurement cases
Each card below states the original quantity, the operation and the result. They are simple enough for readers to verify with an independent calculator. One card uses a percentage-point comparison rather than a physical unit; that example demonstrates how a missing reference scale can also change meaning.
The cards deliberately vary the operation: multiplication, division, addition after conversion, a derived unit, area conversion and percentages. That mixture is more educational than repeating twelve length conversions.
- A. Distance: 2 kilometres = 2 × 1,000 = 2,000 metres.
- B. Shorter length: 750 centimetres = 750 ÷ 100 = 7.5 metres.
- C. Water: 2.5 litres = 2.5 × 1,000 = 2,500 millilitres.
- D. Seed mass: 0.8 kilogram = 800 grams.
- E. Time: 90 minutes = 90 ÷ 60 = 1.5 hours.
- F. Speed: 72 kilometres per hour = 20 metres per second.
- G. Walking speed: 600 metres travelled in 120 seconds = 5 metres per second.
- H. A fraction of pupils: 30% of 40 fictional club members = 12 members.
- I. Percentage points: a rate changing from 20% to 30% increases by 10 percentage points.
- J. Rectangular bed: 3 metres × 4 metres = 12 square metres.
- K. Area units: 1 square metre = 10,000 square centimetres.
- L. Combined mass: 500 grams + 0.75 kilogram = 0.5 + 0.75 = 1.25 kilograms.
8. Check the speed conversion with units
The step in F is worth inspecting. Seventy-two kilometres per hour means seventy-two thousand metres per 3,600 seconds. Dividing 72,000 by 3,600 gives 20 metres per second. That is equivalent to dividing the numerical kilometres-per-hour value by 3.6 when converting to metres per second.
The result is not “20 kilometres per second”. A model could easily copy the first unit label after performing the division, producing a wildly different claim. The unit must change along with the number.
9. Why a percentage is not a headcount
Thirty per cent is a proportion: 30 out of every 100, or 0.30 of a whole. For forty members, 0.30 × 40 = 12 members. The final answer is a count of people, not twelve per cent. This matters in school reports because percentages sound meaningful even when the group size is omitted.
If the total population changed from forty to four hundred, thirty per cent would represent 120 members, not twelve. A useful AI answer shows the whole from which the percentage was calculated.
10. Percentage points are not relative percentage change
A fictional participation rate rises from 20% to 30%. The difference is ten percentage points. Relative to the original 20% rate, the increase is (30 − 20) / 20 = 0.5, or 50%. Those statements describe the same change from different baselines.
An assistant that reports “a 10% rise” without specifying percentage points or the reference denominator may mislead. In many practical contexts, a move from 20% to 30% is not a ten-per-cent relative increase. It is a ten-percentage-point increase and a fifty-per-cent relative increase.
11. The area trap: square the conversion factor
One metre equals one hundred centimetres, but one square metre contains 100 × 100 = 10,000 square centimetres. A square metre is an area, not a line segment. Using 100 as the conversion factor for an area silently treats two dimensions as one.
For volume, the factor is cubed. One cubic metre equals 1,000,000 cubic centimetres. The exponent on the unit tells you how many dimensions are multiplied. This is a strong example of why units should be treated as mathematical structures, not decorative suffixes.
12. Litres are volume, not mass
One litre measures volume. One kilogram measures mass. In particular physical circumstances, a relationship between a liquid’s volume and mass can be estimated from its density, but litres and kilograms are not generally interchangeable. The density must be known or appropriately specified.
An AI asked to convert two litres of an unknown liquid into kilograms cannot produce one exact result from unit definitions alone. The correct next step is to ask for the substance and conditions or to state a properly bounded assumption. Pretending that every litre weighs a kilogram would be a scientific error.
13. The unit in the question controls the unit in the answer
“How far did the pupil walk?” asks for a distance. “How long did the pupil walk?” asks for a duration. “How fast did the pupil walk?” asks for distance per time. Those questions are related but request different dimensions.
A model that extracts only the numbers from the paragraph may compute a correct operation for the wrong question. The first task is semantic: identify what the user wants. Then the mathematical operation can be selected and checked.
14. Why adding quantities requires a shared scale
Card L asks for 500 grams plus 0.75 kilogram. The values cannot be added as 500 + 0.75 without considering units. Convert 500 grams to 0.5 kilogram, then add 0.5 + 0.75 = 1.25 kilograms. Alternatively, convert 0.75 kilogram to 750 grams, then add 500 + 750 = 1,250 grams.
Both routes give the same mass. The agreement is a useful independent check. A robust AI workflow can test the answer by converting it back to the original unit to see whether the quantity has remained unchanged.
15. Why a calculator is useful but not sufficient
A calculator will correctly return 500.75 for 500 + 0.75. It cannot infer that the first number was in grams and the second in kilograms unless that information is explicitly encoded. The model or an upstream parser must supply the right expression.
This reveals a general system boundary. Language interpretation identifies the objects, units and requested operation. A deterministic calculation checks the arithmetic. The final explanation reports the value and unit. Errors can occur at any of those stages.
16. The fictional baseline with three unit mistakes
Suppose a made-up assistant gets nine of the twelve measurement cards correct but converts 1 square metre to 100 square centimetres, leaves 72 kilometres per hour as “20 kilometres per hour”, and adds 500 + 0.75 to produce “500.75 kilograms”. The illustrative score is 9/12 = 75%.
The mistakes are not identical. One is an area-scale error, one is a missing unit conversion in the final label, and one is an addition across incompatible scales. A vague instruction to “do the Maths carefully” does not isolate those causes. A unit-aware validator would inspect the dimensions and conversions explicitly.
17. The repaired fictional process
The revised process extracts each measurement as a tuple: quantity, unit, dimension, and source. It identifies the target dimension and unit, normalises compatible values before arithmetic, performs the operation, then checks that the output has the requested dimension. It reports any assumptions and verifies reversibility when practical.
Suppose the repaired fictional process gets all twelve cases correct. That is 12/12 in an invented teaching packet, not an observed model benchmark. A genuine evaluation would need fresh values, mixed units, adversarial wording, missing-unit cases and independent ground truth.
18. Units protect the reader from scale errors
A distance of 2,000 metres is the same as 2 kilometres, but the numeric values differ by a thousand. Mistaking metres for kilometres can make a proposed route a thousand times longer. In science and engineering, similar scale errors can be very consequential.
A good assistant should preserve explicit unit labels in intermediate steps, not just in the final sentence. A reader inspecting “72 km/h × (1,000 m / 1 km) × (1 h / 3,600 s)” can see how the conversions cancel. An unexplained 20 offers less opportunity to catch a mistaken factor.
19. Significant figures and rounding are separate from conversion
Suppose the original measurement is reported as 2.5 metres, not 2.500000 metres. Writing an answer with many extra decimal places can imply precision that was not measured. Some quantities, such as defined unit conversion factors, are exact; measured values carry uncertainty.
The appropriate rounding depends on the context. A school exercise may specify how many decimal places to use. A laboratory report may require attention to measurement uncertainty and significant figures. The assistant should not automatically interpret a long decimal output as more accurate.
20. Exact conversion and approximate real-world measurement
“One kilometre equals one thousand metres” is a defined conversion, not an estimate. “The walking path is about one kilometre long” is an approximate physical statement. Converting the latter to “exactly 1,000.000 metres” does not make the path’s measurement more certain.
Keep track of whether a number was exact, rounded or measured. If a result is approximate, say so. A unit-conversion algorithm can preserve numeric equivalence while remaining unable to improve the quality of the original observation.
21. When units are missing from the source
“Our class used 400 of water” gives too little information to infer a volume confidently. It might mean 400 millilitres, 400 litres or another quantity entirely. A model that invents “400 millilitres” because it is plausible has changed the record.
The best next step depends on the task. For a fictional example, the assistant can state a reasonable assumption. For a science report, it should request the missing unit or consult the original instrument record. Missing measurements are not repaired by fluent prose.
22. Unit symbols and words need care
The symbol “m” denotes metre; “min” denotes minute. “ms” commonly denotes millisecond, not metre-second. A capital letter can matter in scientific notation. In casual notes, people also use ambiguous abbreviations that require domain context.
The international SI reference helps standardise symbols and prefixes, but it does not guarantee that every informal message follows the standard. An AI should preserve the original notation, interpret it cautiously and clarify when multiple readings could change the answer.
23. Temperature conversion requires an offset
Not all unit conversions involve multiplication by a factor. Temperature scales provide a familiar example: a Celsius temperature and a Kelvin temperature differ by an offset. A value of 25 degrees Celsius corresponds to 298.15 kelvin. Multiplying 25 by an arbitrary constant would not work.
The distinction also affects temperature differences: a change of one degree Celsius has the same size as a change of one kelvin, but an absolute temperature reading is not numerically the same. A model must distinguish the kind of quantity from its measurement scale.
24. Rates combine dimensions
Speed is distance per time. Flow rate is volume per time. Density is mass per volume. Each derived quantity is built from dimensions, which helps us detect impossible calculations. A reported density in “metres per second” is suspicious even if the numerical value seems plausible.
Before calculating a rate, ask whether the numerator and denominator refer to the same observation period and object. Dividing one student’s total distance by another student’s time is arithmetically possible but answers no coherent speed question about either student.
25. Percentages and averages need explicit populations
A percentage attached to “students improved” requires a group size and a definition of improved. An average attached to “revision time” requires the observations to be comparable. Mixing minutes and hours in the same column without conversion can distort the mean.
Unit-aware reasoning is therefore also data-aware reasoning. If one table row says 45 minutes and another says 1.5 hours, convert them to 45 and 90 minutes before averaging. The mean is 67.5 minutes, or 1.125 hours, for that two-value example.
26. Separate a model’s estimate from a measured value
An AI may estimate that a walk of 2 kilometres takes roughly 25 minutes based on an assumed speed. That can be useful as an illustrative estimate if the assumed walking speed is stated. It is not the same as an observed travel time for a particular route, person or traffic condition.
When the user needs an actual itinerary or route duration, a mapping or transport source is the appropriate evidence. A conversion from distance to time needs speed, and speed may vary. No equation removes the uncertainty in the missing input.
27. A transfer exercise with new numbers
Try these without looking at the previous cards. A: 350 centimetres is how many metres? B: A pupil cycles 1,200 metres in 240 seconds; what is the average speed in metres per second? C: A rectangle measures 50 centimetres by 80 centimetres; what is its area in square centimetres?
Answers: A is 3.5 metres. B is 5 metres per second. C is 4,000 square centimetres, or 0.4 square metre. The last result follows because 50 cm = 0.5 m and 80 cm = 0.8 m, so 0.5 × 0.8 = 0.4 square metre.
28. What to check before sharing an AI calculation
Confirm the original numbers, unit symbols and measurement dimension. Identify the requested result. Make sure every addition uses compatible units. Check conversion factors, exponents and numerator/denominator order. Check whether the source was approximate and whether rounding is appropriate.
Finally, inspect the final sentence. Sometimes the intermediate arithmetic is correct but the written label is wrong. A system that explains the steps should still verify that the answer says metres, metres per second or square metres as appropriate.
29. Designing a fair benchmark
Include conversions that require multiplication and division, mixed-unit addition, area and volume, rates, percentages, units with offsets, missing-unit requests and changes in stated precision. Create ground-truth calculations before seeing the model’s responses. Record the formula, source measurement and accepted rounding tolerance.
Measure the arithmetic result and dimensional correctness separately. A model that returns “5 kilograms per second” where the expected answer is “5 metres per second” should not receive full credit just because the number five matches. In serious applications, the units are part of the answer.
30. Research, standards and the wider system
The BIPM’s SI Brochure provides the authoritative unit definitions and accepted usage notes; its ninth edition was updated in 2026. That reference is different from a particular application’s numerical evaluation. Models can make arithmetic or unit mistakes, and a system that checks dimensions and uses deterministic computation can often expose errors.
The broader lesson is architectural: language models are useful at interpreting flexible human descriptions, while explicit data types, calculators, unit libraries and source records provide specialised checks. A coherent system combines them instead of relying on plausibility alone.
Frequently asked questions
Can you add metres and kilograms? No. They describe different dimensions. No meaningful sum exists without a new, explicitly defined model.
Is one litre always one kilogram? No. Volume and mass are distinct. Converting between them requires density under the relevant conditions.
Why is one square metre not 100 square centimetres? Both length dimensions change by 100, making the area factor 100 × 100 = 10,000.
What is the difference between a percentage and a percentage point? A percentage point is the arithmetic difference between two percentages; a relative percentage change uses the earlier value as denominator.
Do AI models always calculate correctly? No. A reliable application checks arithmetic and dimensions with suitable tools and human review when needed.
Does a correct conversion prove the original measurement was accurate? No. Conversion preserves a value’s stated meaning; it does not verify the source measurement.
The central lesson
The number is only half of a measurement. The unit tells you what kind of quantity you have, and the operation determines what kind of quantity you can obtain. A strong AI answer preserves all three: original measure, valid conversion and correctly labelled result.
This is a practical foundation for students and for advanced AI systems. When the units make sense, the calculation is easier to inspect; when the units do not make sense, fluent language should not be allowed to hide the error.
Connected articles and source references
- How Super Intelligence Works — complete curriculum.
- Quantifiers — Some, Every, Most and Exactly — precision about counts and denominators.
- Negation and Exceptions — interpreting conditions and thresholds.
- BIPM, The International System of Units (SI), ninth edition, updated 2026 — authoritative unit definitions.
- BIPM, The SI — base quantities, unit definitions and standard practice.
- NIST Generative AI Profile — reliability and verification practices.