AI helps scientific researchers search literature, formulate hypotheses, plan controls and check calculations. It does not make a generated explanation true. Genuine scientific conclusions require authentic measurements, reproducible methods and an honest distinction between observation, inference and simulation.
In this chapter, the fictional Alto Light Lab tests a prediction about two invented transparent sheets. Readers calculate the mean of five values for each sheet, find that the data contradict the prediction and repeat the method with a changed reference reading. The exercise teaches why science is more than producing a plausible answer.
This is topic 092 in How Super Intelligence Works. The series title is an editorial umbrella for present-day AI, not a claim that current models possess artificial superintelligence.
Chapter contents and learning route
- 1. Question before answer
- 2. Literature search and source ownership
- 3. Testable predictions
- 4. The fictional Alto Light Lab
- 5. Variables and controls
- 6. Calculate P’s mean
- 7. Calculate Q’s mean
- 8. Normalisation and units
- 9. Repetition and independence
- 10. Confounders and causal claims
- 11. A simulation is not a measurement
- 12. Use statistics for the actual design
- 13. Reproducibility and audit trail
- 14. Scientific creativity and AI
- 15. Recent automated-hypothesis research
- 16. Ethical research and permitted tools
- 17. Changed-context practice
- 18. What the transfer result does not prove
- 19. A science-writing source check
- 20. Practical classroom exercise
- 21. Questions and limits
- 22. Sources and connected learning
1. Question before answer
AI can search, classify, model and propose scientific ideas, but none of those operations independently proves a hypothesis. A scientifically useful claim needs an identified phenomenon, transparent evidence and a method by which the proposed explanation could be contradicted.
A clear research question is narrower than ‘Can AI discover science?’ It asks what variable changes, what should be observed, and which alternative explanations must be separated. Human scientific responsibility begins before a model produces an attractive answer.
2. Literature search and source ownership
A literature search locates relevant studies. A model can rank passages and summarise methods, but a model-generated citation does not establish that a paper exists or says the cited thing. Check author, publication date, version, population and methods.
A review, primary experiment and editorial commentary have different evidential roles. A reputable study may support a specific finding while leaving broader generalisations unsettled. The source’s actual scope must remain attached to the resulting sentence.
3. Testable predictions
A hypothesis should make a prediction that some credible observation could contradict. ‘Sheet P will transmit more light than Q under a fixed lamp and sensor’ is testable within a defined measurement plan.
If the resulting measurements favour Q, the researcher must not retrospectively rewrite the original prediction and call the experiment a success. Retaining the hypothesis as written preserves a meaningful test of what was expected.
4. The fictional Alto Light Lab
A completely invented classroom laboratory uses two imaginary transparent sheets, P and Q, a constant lamp and a sensor. The no-sheet reference is 100 arbitrary brightness units. Sheet P yields five fictional values: 72, 70, 74, 71 and 73. Sheet Q yields 83, 81, 85, 82 and 84.
The predeclared prediction is that P produces the higher reading. Every material, value and observation is an original fictional teaching fixture. Nothing was measured with a real lamp or used to make optical product claims.
5. Variables and controls
The independent variable in the paper case is sheet identity; the dependent output is the sensor reading. A real study would need suitable controls for source intensity, sensor geometry, ambient illumination and mounting.
Changing more than one important condition between groups could confound the outcome. Listing a possible control in a model-produced report is not evidence that investigators actually implemented or verified it.
6. Calculate P’s mean
Add P’s five numbers: 72+70+74+71+73 = 360; divide by five for a mean of 72 units. Its readings range from 70 to 74. These are correct descriptions of the supplied packet.
Students can check the mean by subtracting 72 from each reading: the deviations 0, −2, +2, −1 and +1 sum to zero. Accurate arithmetic is not evidence of physical measurement.
7. Calculate Q’s mean
Q’s total is 83+81+85+82+84 = 415 and its mean is 83. The values range from 81 to 85. The difference in the fictional averages, Q minus P, is 83−72 = eleven units.
Under these stated readings Q, not P, has the higher average. The prediction favouring P has not been supported by this packet. The report should say so, instead of changing the direction of the original hypothesis.
8. Normalisation and units
Under a simplified linear comparison to the no-sheet baseline of 100, P’s mean represents 72% of the reference reading and Q’s 83%, an eleven-percentage-point difference.
These are ratios of arbitrary sensor values. They are not certified optical-transmission properties; real spectral and photometric interpretation depends on equipment, calibration and conditions. Preserve the boundary between a convenient mathematical ratio and validated physical meaning.
9. Repetition and independence
Five readings may share the same lamp drift, calibration error and sheet positioning. They are not automatically five independently prepared specimens. Overcounting correlated observations may create an inflated sense of certainty.
An appropriate follow-up design might alternate the sheet order and repeat setup sessions. The design should reflect the question and source of variation. A model can propose repeats but cannot claim they happened without evidence.
10. Confounders and causal claims
Suppose P was always measured first and Q later while the lamp output changed. Time and sheet identity would vary together. The difference could be caused partly by the source rather than the material.
A carefully designed experiment tries to distinguish those explanations. Even a corrected design has limits; no small sample justifies declaring the answer universally true for all lamps, sensors and transparent materials.
11. A simulation is not a measurement
An AI model can simulate plausible readings to help test analysis software. Those generated values must remain labelled simulated, not actual observations.
The companion World Models and Simulation article explores this boundary in a different system. Neither a detailed animation nor realistic-looking table independently proves that a laboratory test occurred.
12. Use statistics for the actual design
A p-value or uncertainty interval depends on a specified sampling model and assumptions about observations. The fictional reading table does not establish that each measurement is an independent replicate of a population.
Publishing a precise statistical significance claim without a justifiable design may confuse readers. The means and ranges are useful descriptive arithmetic; stronger inferential conclusions would require an appropriate, documented investigation.
13. Reproducibility and audit trail
A reproducible account identifies the actual materials, instrument conditions, units, source data, software versions and exclusion decisions. A colleague should know how to repeat the method or inspect its computation.
Do not allow a writing model to invent missing calibration dates, approval numbers, laboratory logs or sensor identities. A professional report may remain incomplete until genuine source records are provided.
14. Scientific creativity and AI
A 2025 npj Artificial Intelligence perspective discusses potential language-model roles across the scientific cycle, emphasising human goals and evaluation. That discussion is evidence about research directions, not proof that every agent-generated idea constitutes a discovery.
Scientific judgement includes choosing worthwhile questions, responding to anomalies and deciding which explanations deserve resources. A model that generates more hypotheses has not necessarily created more validated knowledge.
15. Recent automated-hypothesis research
Nature’s June 2026 discussion examines published studies of agent systems supporting biomedical hypothesis work. The articles address specific research tasks and evidence; they do not establish unlimited independent laboratory competence.
When describing the latest capability, retain publication dates and exact measured outcomes. Avoid promoting impressive demonstrations into universal claims about science or intelligence.
16. Ethical research and permitted tools
A model may suggest activities involving people, animals, sensitive records or hazardous materials. Their execution requires relevant qualified oversight and, where applicable, formal review and authorisation.
The invented Alto packet is entirely paper-based. Readers can practise arithmetic and source criticism without acting on operational instructions or inventing approval for a real experiment.
17. Changed-context practice
A second imaginary lamp has a no-sheet baseline of 200 units. Sheet R’s mean reading is 130 and sheet S’s is 150. R gives 130/200 = 65% of the baseline while S gives 150/200 = 75%.
S leads by twenty raw units and ten percentage points. Students must recompute with the new reference rather than reuse the first packet’s numbers. The calculation remains illustrative and does not establish real material properties.
18. What the transfer result does not prove
The R/S calculation shows that S has the higher ratio under the stated invented readings. It does not justify that S is generally safer, more efficient or superior across other wavelengths or equipment.
A more comprehensive claim would need authentic observations, suitable controls and relevant uncertainty analysis. The correct reasoning method transfers while the substantive answer remains bounded to this fixture.
19. A science-writing source check
A model can write an elegant abstract even when it confuses source data, calculated results and speculation. Before publication, map each consequential statement to a measurement, calculation, verified citation or explicit hypothesis.
Charts deserve the same audit. Check that every plotted value matches the table, units are labelled and missing observations are not silently replaced by plausible numbers. A smooth graph is not a certification of truthful data.
20. Practical classroom exercise
Ask a learner to write the predicted direction, calculate the two means unaided, check the original hypothesis and identify a confounder. Then show an AI explanation and ask which parts are genuinely supported by the packet.
A second round changes the reference reading and the sheet values. This checks independent scientific thinking rather than a pupil’s ability to repeat one generated answer. The goal is learning that transfers beyond a familiar prompt.
21. Questions and limits
Can AI propose hypotheses? Yes, but proposals need meaningful tests. Can it perform analysis? It can assist with permitted calculations and tools, while inputs and interpretation still need verification.
Is an AI-generated experiment a completed experiment? No. Can an unexpected result be valuable? Yes, if the measurement and design are sound. A defensible report identifies what happened and what remains uncertain.
Deeper science laboratory: controls, source provenance and changed conditions
Why the units must stay visible
The expression ‘P averaged 72’ is incomplete unless the measurement quantity, source and reference are understood. The Alto packet defines arbitrary brightness units and a no-sheet reading of 100. An accurate description says P averaged 72 of these stipulated sensor units under the fictional setup. It does not quietly call the readings lux, watts or absolute transmittance.
When pupils change instruments, the numerical scales can differ. A value on one meter may not be comparable to a value on another without calibration and a relevant physical interpretation. Naming the unit or its limitations prevents a mathematically correct answer from becoming an invalid scientific claim.
The full descriptive table
The P row contains five observations, a sum of 360, mean of 72 and observed range of 70–74. The Q row contains five observations, sum of 415, mean of 83 and range of 81–85. The complete dataset contains ten readings; each group contains five. The difference Q−P is eleven units.
A reviewer can reconstruct every number from the original values. If an AI-generated report states P has six trials or Q’s sum is 425, the inconsistency is detectable without needing a more powerful model. Preserve the raw observations with the summary rather than presenting an unexplained chart.
How to check a result with a second method
One check computes the means by addition. Another notices that P’s readings are symmetric around 72: 70 and 74 balance, 71 and 73 balance, and 72 is central. Q similarly balances around 83. These checks are independent ways to verify the packet’s arithmetic.
A second method should not simply ask the same model to repeat its answer in more confident language. The point is to expose arithmetic structure or use a separate trusted calculation, not to manufacture agreement between correlated guesses.
The mean does not explain the mechanism
Suppose Q has a larger mean transmitted-light reading. That says what the supplied numbers show, but not why the difference exists. It could reflect a genuine property of the sheet under controlled conditions or a systematic variation in how the measurements were taken.
A causal explanation requires additional evidence about materials, source intensity, sensor behaviour and possible confounders. The researcher should resist an AI assistant’s tendency to complete a neat narrative by inventing a physical mechanism not tested in the packet.
What a genuine laboratory log would need
A real research log would identify the operator, date and conditions, instrument and calibration information, observations and any deviations from protocol. It would distinguish a recorded value from a later transcription or corrected data entry. Every exclusion decision should be traceable.
An AI writing helper could convert these genuine notes into a readable methods section. But when a note is absent, it must mark the missing field or ask an authorised researcher for it. Generating a plausible serial number is not a legitimate way to complete the record.
Control with a blank measurement
The 100-unit no-sheet reference gives a baseline for one specified lamp and sensor. A researcher may also need to determine what reading occurs with the source off, to separate background offset from the transmitted signal. The Alto packet does not supply such a dark reading.
Without that additional information, our normalised ratios remain teaching calculations under the stipulated simplification. Do not insert a zero dark reading just because it would make the arithmetic convenient. A missing control belongs in the experimental limitations.
How ordering can mislead
Suppose a source becomes gradually brighter while a researcher measures P five times and then Q five times. The apparent difference may correlate with sheet type simply because the order was fixed. A temporal trend is a plausible competing explanation for the invented pattern.
Alternating the measurements or randomising order could help distinguish that effect, subject to the actual instrument and safety conditions. A model can suggest this repair, but the final report must identify whether the improved procedure was used or remains only a proposal.
Why a follow-up must change one question at a time
After noticing Q’s higher reading, a researcher might want to test whether sheet thickness matters. Changing thickness, lamp colour and detector placement at the same time makes attribution difficult. A follow-up should define the factor of interest and appropriately control other conditions.
This is not a rule that every experiment changes only one variable; factorial designs can study several factors deliberately. The underlying requirement is an interpretable design in which the relevant comparisons and interactions are specified before the results are assessed.
Evidence quality is not proportional to prose length
A two-sentence result with correct numbers, units and a limited claim may be stronger than an eight-page AI-generated explanation filled with unsupported assumptions. Readers should look for sample identity, methods, controlled comparisons and evidence for each inference.
An AI system can still provide real value by organising details clearly. The condition is that its expanded explanation does not silently add results, observations or literature claims that the research team never supplied.
A source could report a different population
A paper measuring one type of optical material at a particular wavelength cannot automatically validate claims about every transparent sheet. When an AI assistant imports a source, it should preserve the study’s intended material, apparatus, measurement settings and limitations.
Scientific transfer requires a reasoned bridge from the original evidence to the new conditions. Sometimes that bridge can be tested. Sometimes the honest answer is that the research cited is related but not sufficient to establish the requested conclusion.
A model can bias hypothesis selection
A text system trained on abundant familiar literature may rank conventional explanations more highly than less-documented anomalies. That may be useful for generating a sensible first list of possibilities, but it can also leave surprising observations underexplored.
Human investigators should remain willing to examine unusual but testable alternatives. The goal is not to prefer novelty blindly; it is to avoid making source familiarity synonymous with scientific truth. Research questions deserve evaluation on their plausibility, importance and feasibility.
Computational experiments are still model-dependent
A computational model of light passing through a material may help predict how conditions should change. But its outputs depend on equations, assumptions, parameter choices and numerical implementation. Repeating a simulation a thousand times does not convert incorrect physics into real evidence.
Compare simulations with appropriately obtained empirical observations before using them to assert physical properties. When the model and the world disagree, the disagreement can be a productive research problem rather than a reason to conceal the measurements.
When calculations are automated
A reproducible spreadsheet or script can calculate means and normalised ratios quickly. However, a program might select the wrong rows, use a wrong denominator or mix instrument settings. Code execution verifies the chosen computation, not necessarily the meaning of its inputs.
The accompanying Code Execution chapter explores independent arithmetic checks. For Alto, the simplest verification remains adding five numbers per sheet, dividing by five and retaining the no-sheet reference in the ratio.
A transfer with unequal reference readings
Imagine a third fictional measurement session with a new source baseline of 250 units. The first sheet reads 150; the second reads 175. Raw values differ by twenty-five, but normalised ratios are 150/250 = 60% and 175/250 = 70%. The relative result is a ten-percentage-point difference.
This case forces the learner to reconstruct the rule instead of reusing the earlier denominators. The numerical answer is complete for the given packet, while the scientific conclusion remains restricted to the invented measurement conditions.
A changed denominator trap
Suppose someone compares the new 175-unit reading directly with Alto’s 83-unit mean and concludes the new sheet transmits more than twice as much light. The readings came from different baselines, so that raw comparison is not enough to support the inference.
Normalising under a meaningful common scale may help, but physical comparability still depends on instrument calibration and experimental settings. A mathematical transformation cannot make fundamentally incompatible experiments equivalent by itself.
Peer review is not a truth stamp
Independent scientific review may identify methodological weaknesses, unsupported conclusions or unclear presentation. Publication can indicate that a work has passed a particular editorial process, but that process does not prove every claim correct or permanently settled.
An AI-assisted review can suggest concerns, but reviewer comments may themselves be mistaken or biased. Authors and reviewers need access to relevant evidence and must distinguish an observed flaw from a speculative objection. The overall process values traceability and correction.
Write a responsible result paragraph
A faithful Alto report could say: ‘In this invented five-reading-per-sheet packet, P averaged 72 units and Q averaged 83, so the supplied Q mean was eleven units higher. This contradicts the original prediction favouring P. The figures are fictional and cannot establish physical properties of any actual sheet.’
That paragraph is useful because it supplies the sample description, calculation, direction and scope. It does not conceal the contradicted hypothesis, claim to have performed an experiment or attach a citation unrelated to the values.
What independent mastery would look like
A pupil can reproduce the P and Q averages without viewing the AI explanation, then correctly compute the ratios when the reference changes from 100 to 200 or 250. The pupil should also be able to name one plausible confounder and explain why a simulated result is not an observed one.
Those changed-context skills are stronger evidence of scientific literacy than copying an elegant paragraph about the importance of research. The teacher can use the child’s work to identify which rule is stable and which part needs renewed explanation.
A compact source-audit checklist
Before treating an AI science answer as a reliable statement, check the research question, study type, publication date, population or material, measurement units, controls, number of observations and actual reported result. Ask which of these facts came from a primary source and which were assumed.
Then inspect the conclusion for overreach: does an association become a cause, an average become a universal law, or a simulation become a measurement? Correct the claim where evidence ends, and identify an appropriate next test when an uncertainty matters.
Connect the science workflow to other chapters
Science combines the same mechanisms treated elsewhere in eduKate’s curriculum: retrieval for relevant evidence, language models for explanations, code for checking computations, simulations for proposed outcomes and human authorisation for real research. Each component can work while the complete study remains invalid.
The Gold Standard of Education stresses usable learning that transfers into life. A pupil who can test a new hypothesis, check arithmetic and recognise the limits of evidence has developed a method rather than simply absorbed an AI-generated answer.
The final evidence register: proposed, fictional, calculated and measured
An evidence register for the Alto exercise can label each element by status. The hypothesis favouring P is proposed. The five P and five Q readings are invented fixtures. The means of 72 and 83, the eleven-unit gap, and the percentages are calculations from those fixtures. Any assertion about the physical properties of a genuine sheet would be a new empirical claim for which the packet supplies no measurements. This four-way separation prevents a convincing AI explanation from borrowing authority it has not earned.
To practise without relying on the model, a student can receive the numbers in a different order and recompute each total. The tutor can then change only the source baseline, leaving the sheet readings unchanged. This reveals whether the pupil understands why normalisation depends on a reference rather than on a memorised numerical answer. Next, change the values themselves. The final result should include an explanation of what changed and what cannot be inferred.
A scientific AI assistant should also distinguish a suggestion to collect more data from having done so. If the user asks whether the new condition has been tested, the honest answer is that only the supplied fictional calculation was completed. No real measurement has been added. Reporting an unresolved evidence gap can be the right scientific outcome and the most useful next step for a researcher.
Continue through the four applied Super Intelligence chapters: Healthcare and Qualified Oversight · Software Engineering and Tests · Creativity, Media and Authorship. Each examines a different path from AI-generated output to evidence, qualified responsibility, verified work and human judgement. Explore the complete 100-topic reading map for the surrounding mechanisms.
22. Sources and connected learning
Read Exploring the role of large language models in the scientific method (2025) and AI systems devise hypotheses and ways to test them (2026). Neither validates this invented lab.
Continue through Research Agents, Grounding, Evaluations, How Super Intelligence Works, How X Works and the Gold Standard of Education. Alto data are entirely fictional; no experiment was conducted. Sources reviewed 22 October 2026.