AI and the Paradigm Shift in Science | From Applying Knowledge to Discovering the Unknown

Science is not simply a list of correct facts. It is a disciplined way of asking questions about the natural and social world, collecting evidence, testing explanations and revising conclusions. A theory is valuable because it explains observations and makes predictions that can be examined. A measurement is useful because its method, uncertainty and limitations can be understood. An experiment matters because others can inspect how it was designed and what happened.

Scientific knowledge is cumulative but revisable. Researchers build on established findings, challenge assumptions and improve methods. A result does not become reliable merely because it appears in a paper. It must be evaluated in relation to the quality of the evidence, the design, the analysis and subsequent work.

AI enters science as a set of tools that can support parts of this process. It may help explore complex datasets or identify patterns that humans would take longer to notice. The purpose is to strengthen the process of investigation, not to replace the requirements for evidence.

Contents · Next chapter

The centre includes established concepts, validated measurements, accepted methods, well-supported findings and specialist knowledge. Students and researchers need this foundation because they must understand what has already been tested before proposing something new. Without it, an apparent discovery may simply repeat a known result or misunderstand a basic principle.

The edge includes unresolved questions, anomalies, competing explanations, uncertain measurements and new combinations of established ideas. A researcher may notice that an experiment does not fit the expected result. A new dataset may challenge a model. A scientific field may lack suitable tools to measure a particular phenomenon. These situations create opportunities for investigation.

AI can help identify a possible edge by searching research literature, comparing datasets or generating alternative hypotheses. But the output must be treated as a proposal. The researcher needs to check whether the problem is genuinely open, whether the evidence is suitable and whether the proposed method can answer the question.

Contents · Previous chapter · Next chapter

AI may assist with identifying research questions, reviewing literature, organising references, exploring data, writing code, generating candidate structures, modelling relationships, planning experiments and drafting explanations. Its usefulness varies by field, task and the quality of the available data. Some activities are more mature than others. A language model’s ability to write a plausible abstract does not demonstrate that the research itself is sound.

The OECD Digital Education Outlook 2026 describes GenAI as a growing support for scientific research and institutional work. Its educational synthesis also separates successful task performance from learning. For a school investigation, those are different questions: did AI help produce an analysis, and did the student become better able to explain, check and revise that analysis? A policy synthesis can organise the questions, while evidence for a particular scientific capability needs to come from the relevant study.

This is a useful way to frame the present moment: AI is not equally capable at every stage of the scientific process. The researcher should select a tool for a specific task, test its output and retain a reliable record of what was done. Claims should reflect demonstrated capability rather than the most ambitious description of the tool.

One concrete example is Abramson and colleagues’ 2024 AlphaFold 3 research paper. The authors evaluate predictions of biomolecular structures against defined benchmarks and report improvements on several interaction tasks. They also describe limitations involving stereochemistry, hallucinated structures, molecular dynamics and difficult targets. The paper supports a specific claim about structure prediction under its evaluation conditions. It does not show that a predicted structure establishes all biological functions, that every proposed molecule works, or that experimental validation has become unnecessary.

Contents · Previous chapter · Next chapter

Scientific literature can be large, specialised and difficult to navigate. AI-assisted search may help identify relevant papers, explain unfamiliar terminology, compare research approaches and find areas where evidence appears inconsistent. This can make a field more approachable for students and researchers entering from another discipline.

But a literature review is not a pile of abstracts. Researchers need to assess study design, sample size, measurement, limitations, publication context and the relationship between a result and the conclusion claimed. A model may cite a non-existent paper, confuse two studies or present an old finding as current. Every important reference must be checked in the original publication.

An effective use of AI is to ask it to identify the assumptions behind a claim, suggest alternative search terms or compare the methods of two verified studies. The human researcher then reads the papers, evaluates their quality and decides what the evidence supports. AI can accelerate discovery of literature, but the work of scientific judgement remains.

Contents · Previous chapter · Next chapter

AI can propose many possible research questions, but researchers still need to decide which questions matter. A question may be technically interesting but irrelevant to the real phenomenon. It may be impossible to test with available instruments. It may already have been answered. It may carry ethical risks that outweigh the expected benefit. Human researchers contribute knowledge of the field, awareness of consequences and the judgement needed to choose a worthwhile direction.

A strong research brief states what is known, what is not known, why the gap matters and what evidence could reduce uncertainty. AI can help identify assumptions and propose alternative ways to frame the question. The researcher then checks the literature and discusses the plan with colleagues or relevant experts. This is a better use of AI than asking it to invent a breakthrough without context.

Scientific novelty also needs perspective. A result may be new to one research group but already established elsewhere. AI-assisted literature review can reduce this risk, but only if its references are checked and the search is broad enough to include other languages, disciplines and research traditions.

Contents · Previous chapter · Next chapter

A hypothesis is a proposed explanation that can be tested. AI can help generate candidate hypotheses by combining concepts, identifying patterns or proposing mechanisms. This is potentially useful because researchers may overlook a connection outside their immediate specialisation. It can also help a student move from a vague curiosity to a question with measurable variables.

The danger is that a generated hypothesis may be obvious, already tested, physically implausible or impossible to investigate with available methods. A plausible sentence is not a scientific contribution. Researchers must check prior work, identify what observation would support or reject the hypothesis and decide whether the question is worth the resources required.

AI can therefore be most useful as a hypothesis challenger. Ask it to propose reasons the hypothesis may be wrong, identify confounding variables, outline an alternative mechanism and suggest what evidence would discriminate between explanations. This encourages the researcher to test the idea rather than become attached to the first generated proposal.

Contents · Previous chapter · Next chapter

AI can help analyse large datasets, identify clusters, detect anomalies and generate code for statistical or computational work. These capabilities can reveal patterns that deserve further investigation. However, large datasets contain noise, missing values, measurement errors and biases. A pattern can appear significant because of the way data were selected or transformed.

Researchers should preserve the raw data, document preprocessing, record software versions and describe analytical choices. They need suitable baselines, tests and validation data. If a model identifies an anomaly, the researcher should ask whether it is a real phenomenon, an instrument issue, a data-entry problem or an artefact of the analysis.

Interpretability matters. A system may make a strong prediction without providing a clear explanation of why. In some fields this may be acceptable for exploration, but important scientific claims need methods that allow other researchers to assess the result. A black-box prediction can point towards an experiment; it should not automatically replace the experiment.

Contents · Previous chapter · Next chapter

Science eventually meets the physical world. A computer simulation can explore possibilities, but a material must still be measured, a chemical reaction observed, a biological system studied under appropriate controls or a physical prediction compared with actual observations. AI may assist with experiment planning or control some laboratory equipment, but the experiment still needs a defensible method and appropriate safety oversight.

Researchers must define variables, controls, calibration procedures, sample selection and stopping criteria. They should record unexpected results rather than omit them because they do not fit the generated hypothesis. Equipment failure, contamination and environmental conditions may matter. A machine can help plan a sequence, but it does not remove the need to understand the apparatus or the consequences of an action.

Laboratory AI should be governed by the risks of the domain. Chemical, biological, clinical, radiological and engineering experiments require trained supervision, approved procedures and relevant institutional controls. A student should never treat a generated protocol as authority to conduct an unsafe activity.

Contents · Previous chapter · Next chapter

A scientific result becomes more useful when other researchers can understand how it was obtained and test whether it holds. Reproducibility may require access to data, methods, code, instruments, materials and sufficient detail about the analysis. The appropriate standard varies by field, especially where privacy, safety or commercial restrictions apply, but the principle of transparent methods remains important.

AI adds new information that may need to be recorded. Which model or service was used? What version? What input data were supplied? Did the model generate code, classify samples, transform data or draft text? Which outputs were accepted, checked or modified? If these details materially affect the method, they belong in the research record.

An AI-generated analysis can be difficult to reproduce if the service changes or if the prompt and data context are not retained. Researchers should preserve code and outputs where appropriate, document important settings and avoid relying on undocumented interactions for critical results. The objective is not to archive every casual conversation. It is to preserve enough of the method for another person to understand the work.

Contents · Previous chapter · Next chapter

Researchers remain responsible for the accuracy of their claims, the integrity of their data, proper attribution and the treatment of people or animals involved in research. A language model cannot accept professional accountability. Institutions and publishers have their own policies, and researchers should follow those policies.

AI can create or amplify research misconduct if it is used to fabricate references, generate images that misrepresent experimental results, conceal negative findings or produce text that the author does not understand. It can also help detect anomalies and support routine quality checks. The same technology can assist integrity or undermine it depending on the workflow and governance.

Researchers should disclose material AI assistance where required, verify every citation, preserve original data and ensure that figures and descriptions represent the actual work. Human responsibility is not a final signature after automation. It is present throughout the design, execution, analysis and communication of research.

Contents · Previous chapter · Next chapter

Many important questions cross disciplinary boundaries. Climate research involves physics, chemistry, biology, geography, economics and public policy. Medical research can involve molecular biology, engineering, statistics, behavioural science and ethics. Materials research may connect chemistry, computation, manufacturing and environmental analysis.

AI can help researchers navigate unfamiliar fields, translate specialist terminology and identify possible connections across publications. This may lower the initial barrier to interdisciplinary work. But superficial similarity is not enough. A concept may have different meanings in two fields, and a method suitable in one domain may be invalid in another.

Human collaborators remain essential because they understand the assumptions and standards of their disciplines. AI can suggest that two findings may be related; researchers must decide whether the connection is scientifically meaningful and design a test that can distinguish a genuine relationship from coincidence.

Contents · Previous chapter · Next chapter

Science is usually collaborative. A research team may include experimental specialists, statisticians, engineers, technicians, domain experts, data stewards and students. AI can support parts of their work, but the team still needs to coordinate responsibilities. Someone must understand the research question, someone must validate the data, someone must supervise the method and someone must take responsibility for the conclusion.

AI may help bridge terminology between specialists or make a first comparison across disciplines. But a team should not allow a fluent system to become the unexamined authority on a field it does not understand. Domain experts must be able to challenge outputs and explain where a suggested method is inappropriate.

The new scientific team may include AI tools as routine collaborators in analysis and workflow. Its strength will depend on whether human members can direct the system, understand the limitations and evaluate the result. The objective is a better research process, not the appearance that an automated system has replaced the need for scientific expertise.

Contents · Previous chapter · Next chapter

Students need to learn scientific foundations: observation, measurement, variables, evidence, models, uncertainty and the difference between a claim and a result. They also need direct experience of experiments and explanations. AI does not remove these requirements. It gives teachers another way to help students explore alternatives and investigate questions beyond a worksheet.

A student studying ecosystems might ask AI to suggest several explanations for a change in a local observation. The class can then examine which variables are measurable, which public data are suitable and what a safe investigation could establish. Students should identify what AI suggested, what evidence they checked and which conclusion remains uncertain.

The eduKate educational direction begins with AI and the Paradigm Shift in Education. AI-Powered Assignments provides a practical design for AI-assisted inquiry, while From the Centre to the Edge shows how teachers can retain the subject foundation and open a route towards the unknown.

Contents · Previous chapter · Next chapter

Proposed lesson with invented data: a teacher gives students the temperature increases recorded for black and white cards in three imaginary classroom trials. The values below are constructed for checking reasoning. They are not real experimental observations, evidence of a material property, or a recommended laboratory protocol. Students begin by deciding what a mean can show and what it cannot show.

The first calculation is straightforward: the mean increase is 4°C for the black cards and 7/3°C for the white cards. The difference between the means is 5/3°C, approximately 1.67°C. Students verify the arithmetic themselves before asking AI to analyse the table. The aim is to have enough mathematical control to recognise when the tool changes units, confuses an increase with a final temperature or calculates a percentage with an unsuitable denominator.

The teacher then presents an intentionally flawed interpretation: “Black objects are always hotter, and colour alone caused the result.” Students mark two distinct problems. Three invented trials cannot establish an unrestricted claim about every object and condition. The table also does not tell us whether material, size, starting temperature, exposure or measurement position differed. A correct mean cannot repair an incomplete design.

AI can help by proposing alternative explanations and questions about the method. Learners must classify each suggestion as relevant, irrelevant or requiring more information. For example, a difference in measurement position could matter; an invented story about the manufacturer of the cards does not become evidence. Students should explain why a proposed control addresses a specific competing explanation rather than copying a generic list of “fair test” conditions.

A second version changes the third black-card value from 5°C to 1°C. The black mean becomes 8/3°C, approximately 2.67°C, and the difference from the white mean becomes 1/3°C, approximately 0.33°C. This sensitivity check shows how strongly one observation affects a small dataset. It does not prove that the changed value should be removed. A learner needs a documented reason to treat an observation as a measurement error rather than discarding it because it weakens the preferred result.

The final product is a compact research note containing the checked calculation, a bounded descriptive claim, two competing explanations and a plan for what additional evidence would help. Any actual experiment would use the teacher’s approved apparatus and procedures. The teacher can ask AI to critique the note, but the student decides whether the critique is valid and records one accepted and one rejected suggestion with reasons.

A separate assessment presents unfamiliar synthetic data, perhaps involving different materials with overlapping readings, and asks for an explanation without AI. A later lesson revisits the same distinction between a pattern and a causal claim. These checks assess different things: assisted analysis, independent understanding and retention. They do not by themselves prove that the whole teaching approach is effective; a school would need a suitable comparison and consistent assessment to make that stronger claim.

This is the classroom connection to scientific discovery. AI can expand the number of explanations considered and make a first analysis easier to produce. The learner becomes more capable only if they can inspect the calculation, recognise what the design leaves unknown, and change their conclusion when evidence changes. The same habits matter when the dataset is far too large for a student to inspect row by row: clear definitions, independent checks, suitable comparison data and a record of how the conclusion was reached.

Contents · Previous chapter · Next chapter

Scientific knowledge reaches the public through schools, museums, universities, news organisations, public agencies and professional institutions. AI can help translate findings, explain technical vocabulary and make datasets easier to explore. It can also make preliminary results sound settled or make an unsupported claim look like a scientific consensus.

Public-facing explanations should distinguish established findings, emerging evidence, model predictions, hypotheses and areas of disagreement. The original study and its limitations should be accessible where possible. A generated diagram should not be presented as an observed image unless it is one. A simulation should be labelled as a simulation.

Science communication should invite questions without pretending that every question has a simple answer. A public that understands uncertainty is better prepared to distinguish a careful scientific conclusion from a confident promotional claim.

Contents · Previous chapter · Next chapter

Scientific capability matters beyond universities and laboratories. Public agencies use science to understand environmental risks, infrastructure, health and resources. Companies use research to develop products and processes. Communities use evidence to understand local conditions. Schools help the next generation learn how to ask testable questions and interpret observations.

AI may help more people access scientific explanations and explore datasets, but public participation requires understandable sources and honest communication about uncertainty. A chart without context can mislead. A prediction without assumptions can create false certainty. A generated explanation should lead to the underlying study or official data when the claim is important.

The paradigm shift is to use AI to expand the reach of disciplined inquiry while protecting the standards that make scientific knowledge reliable. Human curiosity sets direction; methods test the possibilities; evidence constrains conclusions; and shared records allow others to check and extend the result.

Contents · Previous chapter · Next chapter

Researchers and institutions should assess whether AI actually improves the quality of science. Did it help identify a useful question? Did it reduce a repetitive workload without introducing unreviewed errors? Did it improve the design or analysis of an experiment? Could other researchers reproduce the result? Did the system create new dependencies or obscure the method?

Counting AI-assisted papers or generated hypotheses is not enough. A large number of outputs may include low-quality work. More meaningful measures include reproducibility, data quality, validation, useful discoveries, time saved in well-defined tasks and the ability of other teams to build on the results.

Institutions should also be willing to stop using a tool when its performance is unreliable or when its risks outweigh its benefits. Scientific progress depends on evidence-based decisions about methods, including methods that involve AI.

Contents · Previous chapter · Next chapter

AI can increase the number of candidate molecules, equations, simulations, draft papers or possible explanations that a research team can produce. But the number of outputs is not a measure of scientific progress. A useful discovery must survive tests that are appropriate to the field. It must be distinguishable from a coincidence, an artefact of the dataset or a result that was already known.

Research institutions should therefore evaluate whether AI-assisted work produces results that other researchers can verify and build upon. The relevant questions include whether methods are documented, data quality is adequate, uncertainty is reported and independent validation supports the claim. More generated hypotheses may increase the search space, but they can also increase the burden of checking.

Scientists need the freedom to explore unusual ideas, but they also need a disciplined method for deciding which ideas deserve resources. AI can help rank or compare candidates, yet its ranking criteria may reflect assumptions that researchers should inspect. An efficient search process is valuable only when it helps find results that matter.

Contents · Previous chapter · Next chapter

Scientific progress does not consist only of successful experiments. A failed replication, an unexpected measurement or a hypothesis that does not survive testing can be valuable. It may reveal a limitation in an instrument, an incorrect assumption or a boundary condition that was previously overlooked. These outcomes help researchers understand where a model works and where it does not.

AI systems may encourage users to produce a neat explanation even when the evidence is uncertain. Researchers should resist that pressure. An incomplete result should be described as incomplete. An anomaly should be investigated before being celebrated as a breakthrough. A model prediction should remain a prediction until suitable observations support it.

Scientific humility is not a lack of ambition. It is the willingness to let evidence change a preferred explanation. AI should increase the number of possibilities people can explore while making it easier to test, reject and revise them. That is a stronger form of intelligence than simply generating answers that sound convincing.

Contents · Previous chapter · Next chapter

AI could widen participation in science by helping smaller research teams search literature, write code, translate technical material and explore datasets. It could also widen inequality if advanced tools, computing resources, high-quality data and expert support are available only to well-funded institutions. The ability to generate an analysis is not the same as the ability to validate it, and institutions with limited resources may lack the expertise needed to check complex outputs.

Scientific collaboration should therefore include access to infrastructure, open methods where appropriate, training and shared standards. Data cannot always be made public, especially where privacy, security, cultural rights or commercial obligations apply. But researchers should explain the reason for restrictions and preserve enough methodological detail for responsible scrutiny.

Students from different backgrounds should also have opportunities to experience investigation. A discovery-centred curriculum does not require every school to own advanced laboratory equipment. Students can work with public datasets, simple observations, simulations and supervised experiments. What matters is learning how to ask a question, select evidence, test an explanation and recognise a limitation.

Contents · Previous chapter · Next chapter

A research institution should be able to learn from its own methods. It can record which AI uses improved a workflow, which introduced errors, which datasets were inadequate and which review procedures caught problems. This evidence should inform procurement, training and future research design. A tool that worked for one task should not automatically be approved for another field with different risks.

Research leaders should encourage staff to report model failures and unexpected outcomes without treating every failure as misconduct. The purpose is to detect problems early and improve the process. At the same time, institutions must address deliberate fabrication, undisclosed manipulation and careless publication. A learning culture needs both openness about mistakes and clear standards of integrity.

AI can help science become more exploratory, connected and productive. But a research system becomes stronger only when its methods remain inspectable and its conclusions remain answerable to evidence. The edge is where the next question appears; the scientific method is how the answer earns trust.

Contents · Previous chapter · Next chapter

A scientific result should be strong enough to invite scrutiny. Researchers should identify what would disprove their explanation, what alternative mechanisms remain possible and which measurements are most important. AI can help generate counter-hypotheses and test cases, but researchers must decide which are scientifically plausible and how to evaluate them. A system that only confirms a preferred result is not supporting discovery; it is reinforcing confirmation bias.

The scientific edge is therefore not just the place where new ideas appear. It is where ideas are exposed to disciplined challenge. AI can increase the number of challenges researchers can consider, while experimental methods and independent review determine which conclusions survive.

Researchers should also distinguish exploration from confirmation. A model may generate a useful lead during an exploratory analysis, but the result should be tested with suitable data or an independent method before it is treated as confirmed. This distinction helps prevent a promising computational pattern from becoming an overstated scientific conclusion.

Contents · Previous chapter · Next chapter

The centre of science is necessary because it carries the knowledge and methods that earlier generations established. The edge is necessary because the world still contains questions that existing explanations do not fully answer. AI can help researchers move between the two by expanding search, analysis and modelling. It cannot guarantee that a generated pattern is real or that a hypothesis is important.

The paradigm shift is not from scientists to machines. It is from a research process limited by some of the practical costs of searching, comparing and exploring towards a process in which more possibilities can be investigated. The responsibility for evidence, method, ethics and interpretation becomes more important as the volume of possible output increases.

Scientific capability grows when people can ask better questions, test more carefully, share methods and correct mistakes. AI can contribute to that process when it is used as an instrument of disciplined inquiry. The goal is not simply to produce science faster. It is to discover more reliably and understand more deeply.

Contents · Previous chapter · Continue the series