The Past and Future of AI Scientists

• AI is identified as the only feasible way to achieve the scientific productivity needed to address climate change, food insecurity, antimicrobial resistance, cancer, and future pandemics.Source: §1 · p. 1
• Science is well suited to AI because its problems are relatively abstract and bounded, and nature does not strategically deceive experimenters.Source: §1 · p. 2
• AI already exceeds human capacity in several scientific components, while human scientists retain advantages in physical grounding, flexibility, tacit understanding, social intelligence, responsibility, and question selection.Source: §1 · p. 2
• The central thesis is that formal knowledge, foundation models, reasoning, belief states, simulation, experimental design, agents, robotics, protocols, data analysis, and provenance are converging to automate the scientific method.Source: §1 · p. 3

2. The History of Discovery Science

• DENDRAL and Meta-DENDRAL established expert-level scientific reasoning by combining explicit domain knowledge, heuristic search, empirical rule learning, and interpretable conclusions from chemical data.DENDRAL inferred molecular structures from mass spectra, while Meta-DENDRAL learned scientifically meaningful rules from examples (Buchanan et al., 1969; Lindsay et al., 1980).Source: §2 · pp. 3–4
• Early discovery systems showed that scientific reasoning could be computationally decomposed into search, conjecture generation, equation discovery, theorem proving, and literature-based hypothesis formation.BACON generated equations from numerical data; AM, EURISKO, and CYRANO explored concepts and conjectures; automated mathematics formalized deduction and verification; Swanson connected previously separate literatures to generate hypotheses (Langley et al., 1987; Lenat, 1977, 1983; Newell et al., 1957; Swanson, 1986).Source: §2 · p. 4
• Statistical bioinformatics and cheminformatics shifted discovery toward data-driven induction, using regression, classical machine learning, probabilistic sequence models, and explicit chemical knowledge to support modern biological breakthroughs.QSAR mapped molecular descriptors to activity, while HMMs enabled gene finding, protein-family classification, and sequence-alignment validation.Source: §2 · pp. 4–5
• Controlling laboratory equipment transformed discovery systems from analysts of existing data into agents that intervene physically to acquire observations discriminating among competing hypotheses.Jan Żytkow and colleagues made the first direct AI control of laboratory equipment.Source: §2 · p. 5

2.1. Adam & Eve

• Adam was the first machine to autonomously discover novel scientific knowledge through logic-based hypothesis formation, experiment selection, robotic execution, interpretation, and belief revision.It automated substantial parts of the scientific method in functional-genomics research on Saccharomyces cerevisiae.Source: §2.1 · p. 5
• Eve established the self-driving-laboratory architecture by combining screening, active learning, optimization, and repeated prediction-to-experiment cycles for efficient exploration of chemical and biological spaces.Modern systems increasingly combine Adam’s hypothesis-led reasoning with Eve’s data-driven closed-loop experimentation.Source: §2.1 · pp. 5–6

3. The Current State of AI for Science

• The current landscape comprises predictive scientific AI, LLM-based and agentic systems, self-driving laboratories, and integrated AI Scientists.AI already predicts structures, proposes molecules, writes code, reads papers and selects experiments, but most systems automate only one task.Source: §3 · p. 6

3.1. Predictive scientific AI

• Predictive models achieve strong results across scientific domains, but they typically receive human-defined questions, representations, data and success criteria rather than autonomously formulating scientific problems or hypotheses.AlphaFold 2 advanced protein-structure prediction, and AlphaFold 3 extended prediction to interactions among proteins, nucleic acids, small molecules, ions and modified residues.Source: §3.1 · p. 7

3.2. Large Language Models and agentic scientific systems

• LLM-based agents expand scientific automation by coordinating literature search, tools, code, analysis and communication, but they remain dependent on human-defined objectives and can produce unsupported or invalid scientific outputs.Examples include Coscientist, ChemCrow, The AI Scientist (Lu et al., 2024), Co-Scientist (Gottweis et al., 2026), and Robin.Source: §3.2 · p. 7

3.3. Self-driving laboratories

• Self-driving laboratories close the loop by representing knowledge, selecting and executing experiments, analysing results, and updating models or hypotheses, but physical errors can damage equipment, waste materials or create safety hazards.Eve established this architecture for drug discovery, and later systems applied it to synthesis, catalysts, materials, formulation, protein engineering and process development.Source: §3.3 · p. 8

3.4. Integrated AI Scientists

• Integrated AI Scientists should explain observations, generate and test hypotheses, select and execute discriminating experiments, interpret results, and revise scientific beliefs.They are more general than self-driving laboratories, which may optimise objectives without explicitly generating or evaluating scientific hypotheses.Source: §3.4 · p. 8
• The components of autonomous science now demonstrate substantial capability, yet no existing system matches human scientists’ breadth, depth and open-ended autonomy; the remaining challenge is integration.Foundation models, tools, formal representations, reasoning, hypothesis generation, active learning, robotics and result analysis each contribute to the scientific loop, while current systems rarely choose important problems or reconstruct their conceptual foundations.Source: §3.4 · p. 9

4. Architecture and Roadmap for AI Scientists

• An AI Scientist must integrate interacting modules that generate hypotheses, reason, simulate, design and execute experiments, interpret results, and revise knowledge without breaking the scientific loop.The architecture combines scientific memory, literature interfaces, reasoning systems, simulation, experimental design, laboratory control, analysis, verification, safety, provenance, and human-intervention interfaces.Source: §4 · pp. 10–11, Figure 3 · p. 10

4.1. One autonomy framework

• The six-level autonomy framework measures human intervention from ordinary tool use to research-programme autonomy, while autonomy remains distinct from capability, generality, scale, reliability, significance, and risk.Higher autonomy does not guarantee important or correct science, and capable systems may still operate under close supervision for safety or legal reasons.Source: §4.1 · p. 11

4.2. Milestones

• Near-term milestones prioritize dependable bounded systems with reliable literature extraction, calibrated hypothesis generation, formally checked and recoverable physical execution, reproducibility, meaningful benchmarks, and interoperable resources.Higher autonomy additionally requires changing problem representations, inventing actions, detecting inadequate instruments, and distinguishing failed theories from failed experiments.Source: §4.2 · p. 12

4.3. Evaluation as part of the architecture

• Evaluation is part of the architecture: modules must expose competence and limits, systems must request help outside validated domains, and every claim must carry machine-readable evidence, provenance, contributions, and residual uncertainty.The evidential record links hypotheses, evidence, deductions, simulations, selected experiments and rationale, observations, analyses, and belief-state updates.Source: §4.3 · pp. 12–13

5. Representations for Scientific Knowledge and Models

• Scientific representations determine which hypotheses, deductions, and experiments AI Scientists can express; formal languages provide the precision, semantics, and executability required for integration and reproducibility.Logic, probability, mathematics, and programs are complementary representations for entities, uncertainty, quantitative laws, and executable models.Source: §5 · p. 13

5.1.1. The formalisation of scientific knowledge

• Formalising scientific knowledge is necessary for computers to reason over entities, relations, assumptions, meanings, evidence, and conclusions rather than merely search natural-language text.Semantic Web, e-Science, and FAIR approaches support machine-interpretable, interoperable, and reusable scientific records.Source: §5.1.1 · p. 14

5.1.2. Ontologies

• Ontologies define concepts, relations, hierarchies, and axioms while distinguishing scientific entities and processes, providing semantic scaffolding for reliable machine reasoning across claims, actions, observations, and evidence.They support interoperable representations spanning complete scientific processes, from investigations and provenance to executable protocols.Source: §5.1.2 · pp. 14–15

5.1.3. Expert systems and the knowledge-acquisition bottleneck

• Expert systems made knowledge explicit, inspectable, queryable, and non-hallucinatory through identifiable inference, but knowledge acquisition was difficult, expensive, slow, and unable to capture much tacit expertise.Natural-language knowledge extraction was historically intractable at scale, preventing comprehensive expert systems.Source: §5.1.3 · p. 15

5.1.4. Knowledge graphs

• Knowledge graphs efficiently integrate heterogeneous scientific datasets through ontologies, but binary triples cannot directly express quantification, negation, uncertainty, modality, temporal context, or relations among relations.Future AI Scientists should combine knowledge graphs’ explicit verifiability with LLMs’ broad linguistic flexibility.Source: §5.1.4 · pp. 15–16

5.1.5. Logic programming and programs as knowledge

• Programs and logic programming can represent executable models, protocols, analyses, and theories while combining declarative knowledge, hypotheses, predictions, constraints, proof traces, and alternative inference strategies.Inductive logic programming additionally learns interpretable and recursive programs from observations and background knowledge, although scalability, noise, and predicate invention remain challenges.Source: §5.1.5 · p. 16

5.1.6. LLMs and the return of expert systems

• LLMs can extract and formalise scientific knowledge from literature, helping overcome expert systems’ knowledge-acquisition bottleneck, but extracted representations require ontological, logical, database, and evidence-based verification.A proposed pipeline retrieves sources, extracts claims and provenance, aligns ontologies, checks consistency, tests deductions, escalates uncertainty, and stores verified knowledge.Source: §5.1.6 · p. 17
• Combining LLMs with symbolic systems enables scalable certified knowledge whose logical consequences, contradictions, evidence links, and hypothesis probabilities can be checked, traced, and updated.This architecture supports expert systems with automatically maintained knowledge bases and conclusions constrained by explicit knowledge and sound inference.Source: §5.1.6 · p. 17

5.2. Probabilities

• AI Scientists should represent hypotheses with degrees of belief, parameter uncertainty, evidence reliability, and probabilities of alternative explanations rather than simple true-or-false values.Scientific knowledge is provisional and probabilistic, with established theories differing from new hypotheses mainly by evidential support and assigned probability.Source: §5.2 · p. 18

5.2.1. Bayesian updating

• Bayes’ theorem updates hypothesis probabilities from evidence, so AI Scientists should select experiments whose possible outcomes differ sharply across competing hypotheses.Predictions confirmed when more probable under a hypothesis increase its probability, while equally predicted outcomes provide little information.Source: §5.2.1 · p. 18

5.2.2. Popper and probabilistic confirmation

• Probabilistic comparison is preferable to decisive falsification because evidence can shift rational confidence without proving or falsifying theories, while apparent conflicts may reflect auxiliary errors or hidden causes.Popper’s emphasis on risky predictions remains valuable because successful predictions have large Bayes factors when unlikely under competing hypotheses.Source: §5.2.2 · p. 18

5.2.3. Ockham’s razor

• Bayesian inference explains Ockham’s razor by balancing data fit against model flexibility, while recognizing that simplicity depends on representation, background knowledge, and assumptions about natural structures.A simpler model may assign greater probability to observed data because its predictions occupy a narrower range, but simplicity is not representation-independent.Source: §5.2.3 · p. 19

5.2.4. Bayesian model averaging and Explanation Selection

• Because scientific discovery is generally M-open, AI Scientists should preserve model inadequacy, use stacking for robust prediction, and use MML-guided search for concise, inspectable explanations.MML balances explanatory fit against model complexity by encoding the model and unexplained evidence as a two-part message, rather than treating relative posterior preference as truth.Source: §5.2.4 · pp. 19–20

5.2.5. Graphical and relational probabilistic models

• Future AI Scientists require probabilistic relational representations combining entities, relations, logical knowledge, uncertain rules and structures, temporal and causal relations, provenance, and learned relations.Statistical relational learning is relevant because scientific knowledge is relational rather than naturally representable as independent feature vectors.Source: §5.2.5 · p. 20

5.2.6. Causal reasoning

• AI Scientists must learn causal models, design interventions that distinguish causal structures, and revise them experimentally, while recognizing that identifiability depends on assumptions, prior knowledge, and interventions.Causal models distinguish observation from intervention through Pearl’s do-operator, and hidden confounding can prevent identification without appropriate controls or assumptions.Source: §5.2.6 · p. 21

5.2.7. Probabilistic programming

• Probabilistic programming makes scientific theories executable and supports inference, simulation, model evaluation, and experiment selection, but scalable AI Scientists must address MCMC intractability and variational overconfidence.Integrating probabilistic programming with causal structure, neural networks, and laboratory automation forms the self-correcting computational core of the AI Scientist.Source: §5.2.7 · pp. 21–22

5.3.1. Equation discovery

• Equation discovery seeks phenomenological relationships that predict observations and mechanistic equations that explain the entities, processes and causal structures producing them.Automated discovery infers both functional forms and parameters rather than merely estimating parameters within human-selected models.Source: §5.3.1 · p. 22

5.3.2. Symbolic regression

• Symbolic regression searches combinations of variables, constants and operators without assuming a fixed functional form, while neural methods learn representations and symbolic methods produce concise, interpretable hypotheses.Approaches include genetic programming, SINDy, autoencoder-based coordinate discovery and AI Feynman’s structure-guided search.Source: §5.3.2 · p. 23

5.3.3. Integrating data with scientific theory

• AI Scientists must combine empirical data, background theory and formal deduction so proposed equations are plausible, meaningful and derivable from stated assumptions.AI-Descartes and AI-Hilbert illustrate this hybrid reasoning paradigm.Source: §5.3.3 · p. 24

5.3.4. Large Language Models for equation discovery

• LLM-based equation discovery can use scientific priors to propose executable programs, but contaminated benchmarks remain problematic and LLM-SRBench found that leading systems solved only a minority exactly.Future systems must combine LLM priors with numerical and symbolic search, constraints, causal knowledge, formal verification and genuinely novel tests.Source: §5.3.4 · p. 24

5.3.5. Depth and generality

• Scientific equations should be judged by explanatory depth, generality, unification, invariance, mechanistic connection, novel prediction and counterfactual usefulness, not predictive accuracy alone.AI Scientists should represent hierarchies linking phenomenological models appropriate at one scale with more fundamental models explaining them.Source: §5.3.5 · pp. 24–25

5.3.6. Mathematical understanding

• Mathematical understanding requires moving from data to equations and from theories to predictions, recognizing representations, approximations and the conditions under which equations apply.This exceeds fitting trajectories because it involves deriving models from first principles within general frameworks such as Lagrangian mechanics.Source: §5.3.6 · p. 25

5.3.7. AI mathematicians

• AI mathematics has advanced in conjecture generation, mathematical construction, and formally verified theorem proving; AlphaGeometry solved 25 of 30 selected Olympiad-level geometry problems.In 2026, an internal OpenAI reasoning model autonomously disproved Erdős’s planar unit-distance conjecture; Tim Gowers judged the result to meet standards expected by the strongest mathematical journals (Alon et al., 2026).Source: §5.3.7 · p. 26

5.3.8. AI Mathematics within an AI Scientist

• Within AI Scientists, mathematical proof certifies deductions while experiments establish whether models describe the physical world, enabling hypotheses, predictions, model revision and trustworthy scientific theories.Proof assistants provide mechanically checkable inferences, but mathematical conclusions alone cannot replace experimental evidence.Source: §5.3.8 · pp. 26–27

5.4.1. Differentiable representation learning

• DNNs learn hierarchical latent representations directly from weak inputs, jointly optimising those representations and scientific predictions end to end.This enables transformations from pixels, sequences, or atomic coordinates into representations where scientific targets are easier to predict.Source: §5.4.1 · p. 27

5.4.2. Architectural inductive biases

• Scientific architectures encode inductive biases through convolutions, recurrence, attention, graphs, and physical symmetries, including equivariance under rotations and translations.Known principles can enter architectures, loss functions, or differentiable simulators rather than only serving as training examples.Source: §5.4.2 · p. 28

5.4.3. Scale and foundation models

• DNNs scale to billions or trillions of parameters and heterogeneous scientific data, while foundation models transfer domain representations across tasks and enable constrained generation.Generated proteins, molecules, and materials still require computational or experimental validation, and depend on training-data quality and coverage.Source: §5.4.3 · pp. 28–29

5.4.4. AlphaFold and protein foundation models

• AlphaFold 2 achieved a qualitative advance in blind CASP14 protein-structure prediction, while AlphaFold 3 extended prediction to joint structures and interactions across biomolecular components.ESMFold predicted structures for more than 600 million metagenomic proteins, and ESM3 generated a functional fluorescent protein with 58% sequence identity to the nearest known examples.Source: §5.4.4 · pp. 29–30

5.4.5. Neural networks across the sciences

• Across science, neural networks expand discovery and forecasting: GNoME identified 2.2 million candidate stable crystals, while Pangu-Weather and GraphCast matched or exceeded leading deterministic systems on many medium-range measures faster.Aurora pretrained on more than one million hours of geophysical data and adapted across weather, pollution, ocean-wave, and tropical-cyclone forecasting.Source: §5.4.5 · p. 30

5.4.6. Neural networks and scientific explanation

• DNN predictions can reflect causal structure, unexplained correlations, or accidental training regularities, while spurious correlations, distribution shift, leakage, calibration failures, and limited interpretability undermine scientific explanation.Latent representations may nevertheless encode scientifically meaningful structure, geometry, and dynamical laws without being explicitly supplied.Source: §5.4.6 · p. 31

5.4.7. Neural networks within AI Scientists

• Future AI Scientists should integrate DNNs’ high-capacity pattern learning with symbolic verification, probabilistic uncertainty, causal intervention, mathematical reasoning, and robotic experimentation.Neural systems can discover implicit structures at scale, while formal languages convert them into hypotheses that can be understood, checked, communicated, and tested.Source: §5.4.7 · p. 31

6.1. The scientific literature as a knowledge system

• Scientific knowledge is a network of claims, evidence, methods, data, versions and responses, so AI Scientists must track epistemic roles, provenance and temporal status across interconnected literature objects.This includes distinguishing observations from interpretations and preserving who stated a claim, when, on what evidence, and whether it was later corrected, challenged or replicated.Source: §6.1 · pp. 32–33

6.2. Text mining

• Traditional text mining extracts entities, relations and events through inspectable pipelines, but ambiguity, stagewise error propagation and task-specific annotation limit its coverage; LLMs should complement rather than replace these methods.Scientific claims with similar vocabulary can differ in polarity, uncertainty, attribution or evidential status, making entity recognition alone insufficient.Source: §6.2 · pp. 33–34

6.3. Searching and retrieving the literature

• Literature retrieval should combine semantic and bibliographic signals, iteratively expand terminology and cited work, deliberately seek opposing evidence, and target the smallest sufficient body of literature rather than maximal coverage.Terminology varies across disciplines and changes through time, while citation networks reveal intellectual relationships that textual similarity can miss.Source: §6.3 · p. 34

6.4. Literature synthesis

• Trustworthy synthesis must connect claims to supporting evidence, distinguish consensus from repeated citation and speculation, assess independence and quality, and verify citations at the level of individual claims.Systematic-review practices provide a model by making search criteria, inclusion decisions, evidence assessment and synthesis reproducible.Source: §6.4 · pp. 34–35

6.5. Multimodal scientific documents

• Scientific reading must integrate prose with figures, tables, equations, images and supplementary materials while preserving layout, coordinates and links to the exact supporting sentence, cell, panel or equation.Abstracts often omit limitations, unsuccessful results and detailed methods, so complete articles are necessary for many questions.Source: §6.5 · p. 35

6.6. Reliability, hallucination and provenance

• Because fluent language models can hallucinate scientific facts and references, literature systems require external source verification, claim-level support checks, reproducible bibliographic records and explicit handling of absence of evidence.More scientific training can increase fluency without making every generated statement reliable; Galactica illustrated this danger (Taylor et al., 2022).Source: §6.6 · p. 36

6.7. Writing the scientific literature

• AI-generated scientific writing must preserve distinctions among prior knowledge, procedures, observations and conclusions, retain uncertainty and negative results, and link every factual statement to an external publication, dataset, experiment, executable analysis or formal deduction.Future AI Scientists should produce connected human-readable and machine-readable records containing explanations alongside claims, evidence, protocols, data, analyses, uncertainty and provenance.Source: §6.7 · pp. 36–37

6.8. Paywalls and the “Dark Data” Problem

• AI Scientists cannot achieve unbiased literature synthesis while scientific knowledge remains trapped in paywalled papers, physical notebooks, legacy formats and proprietary databases, requiring legal access frameworks and large-scale digitisation.Recent open-access mandates improve availability, but the historical record remains highly fragmented and “dark”.Source: §6.8 · p. 37

7. Hypothesis Generation and Scientific Creativity

• Scientific hypotheses must extend or revise knowledge to explain observations, predict phenomena, or identify regularities while remaining connected to uncertainty, feasibility, and empirical discrimination.Open-ended discovery depends on representational creativity: systems able to invent entities, predicates, equations, mechanisms, and experimental concepts can transform the problem representation.Source: §7 · p. 37

7.1. Problem formulation and question generation

• Useful questions target unresolved, answerable, consequential uncertainties that are specific enough to guide investigation and important enough to justify resources.AI Scientists can find questions through knowledge-graph uncertainty, publication contradictions, model discrepancies, unexplained data clusters, poor theoretical predictions, or human goals, but human governance must set priority among needs.Source: §7.1 · p. 38

7.2. Abduction

• Abduction generates candidate explanations for observations, especially structured discrepancies, but competing explanations require ranking by background consistency, probability, simplicity, causal coherence, coverage, and discriminating-experiment value.Abductive explanations are not deductively guaranteed: multiple hypotheses may explain one observation, so candidates must subsequently be tested.Source: §7.2 · pp. 38–39

7.3. Induction and relational learning

• Inductive learning can produce powerful but defeasible models, so AI Scientists must integrate it with probabilities, experimental testing, and rational belief revision.Relational and inductive-logic approaches express structured scientific mechanisms and incorporate background knowledge, but their larger hypothesis spaces require language biases, type constraints, efficient search, and useful intermediate predicates.Source: §7.3 · p. 39, §7.3.1 · pp. 39–40

7.4. Concept and predicate invention

• Concept and predicate invention expands scientific language by introducing entities, properties, and relations that can simplify observations and deepen explanations.LLMs can propose meaningful predicates, while logic programming, theorem proving, and experiments must test, refine, and verify them; formal verification establishes consistency relative to premises, not reality in nature.Source: §7.4 · p. 40

7.5. Analogy and transfer; 7.6. Literature-based discovery

• Analogy and literature-based discovery transfer relational structure or connect distributed evidence across domains, but candidate connections must be constrained by causal or formal structure and validated experimentally.Foundation models can search across disciplinary vocabularies, yet attractive analogies and knowledge-graph paths are often scientifically meaningless without coherent mechanisms, supported edges, or testable consequences.Source: §7.5 · p. 41, §7.6 · p. 41

7.7. Mechanistic and causal hypotheses

• Mechanistic hypotheses must explain why regularities occur, specify failure conditions, support counterfactuals, and motivate interventions that distinguish observationally equivalent mechanisms.Causal models, conservation laws, spatiotemporal constraints, domain capabilities, and intervention data should restrict plausible mechanisms rather than treating predictive association as explanation.Source: §7.7 · pp. 41–42

7.8. LLMs and multi-agent hypothesis generation; 7.9. Evaluating candidate hypotheses; 7.10. From hypothesis to research programme

• Reliable AI Scientists should separate broad, diverse hypothesis generation from acceptance through novelty checks, formal consistency tests, prediction and experiment design, and evidence-based rejection.LLMs provide breadth but linguistic plausibility is weakly tied to truth; valuable candidates must satisfy novelty, background consistency, explanatory and predictive criteria, causal plausibility, simplicity, feasibility, and independent experimental or proof-based evaluation, then become iterative research programmes updated after each result.Source: §7.8 · p. 42, §7.9 · p. 42, §7.10 · p. 43

8. Deduction, Simulation and Prediction

• Scientific deduction derives testable consequences; simulation extends it to systems too complex for closed-form analysis, while remaining conditional on the model and its assumptions.Simulation outputs may be internally valid yet false about nature if parameters, conditions, numerical methods, or model assumptions are wrong.Source: §8 · p. 43

8.1. Scientific models as executable hypotheses

• An executable scientific model combines formal entities, states, relations, and rules with assumptions that generate predictions comparable to observations.AI Scientists must distinguish established components from conjectures and identify observations that discriminate among alternative model structures.Source: §8.1 · p. 44

8.2. Forward simulation and prediction

• Forward simulation predicts observations under specified models, parameters, conditions, and interventions, enabling counterfactual reasoning for planning, design, risk assessment, and intervention analysis.Predictions should normally propagate input and structural uncertainty into outcome distributions rather than report a single apparently exact value.Source: §8.2 · p. 44

8.3. Inverse problems and simulation-based inference

• Simulation-based inference learns which hypotheses or parameters best explain observations when realistic simulators generate data but have intractable likelihoods.Its reliability remains conditional on the simulator including mechanisms capable of generating real observations; omitted processes can produce misleading parameter estimates.Source: §8.3 · p. 45

8.4. Verification, validation and calibration

• Verification checks whether equations are implemented and solved correctly, validation checks empirical adequacy for intended use, and calibration estimates uncertain parameters without replacing independent validation.AI Scientists should represent structural, parameter, condition, measurement, and numerical uncertainty, then use sensitivity analysis to assess robustness.Source: §8.4 · pp. 45–46

8.5. Differentiable programming

• Differentiable programming enables gradients through numerical simulators for parameter estimation, control, inverse design, and experimental optimisation, but discontinuities, stochasticity, memory costs, and model error limit interpretation.A gradient describes local changes in the specified simulator and is not itself a causal explanation.Source: §8.5 · p. 46

8.6. Hybrid scientific models

• Hybrid models combine mechanistic structure with learned components for unknown or unresolved processes, retaining scientific concepts while correcting or completing theory from data.They are especially useful when first-principles simulation is infeasible but purely data-driven models would discard established knowledge, and they can support model discovery.Source: §8.6 · p. 47

8.7. Surrogate models and emulation

• Surrogate models accelerate exploration of expensive simulators, supporting uncertainty propagation, sensitivity analysis, inverse problems, optimisation, experimental design, control, and hypothesis comparison.Because surrogates inherit simulator limitations and add approximation error, AI Scientists should detect distribution shift and selectively invoke full simulations in multi-fidelity workflows.Source: §8.7 · pp. 47–48

8.10. Simulation within an AI Scientist

• Simulation provides a comparatively safe and inexpensive environment for checking proposed experiments before physical execution.It can help evaluate protocols without immediately performing them in the physical world.Source: §8.10 · p. 49
• Simulation can reject impossible protocols, estimate expected outcomes, and identify informative measurements.Source: §8.10 · p. 49
• Because simulation cannot reveal phenomena absent from its model, AI Scientists must use simulation and physical experimentation together.Simulation derives and compares consequences, while experiment determines which represented world most closely corresponds to reality.Source: §8.10 · p. 49

9. Observation, Measurement and Experimentation

• AI Scientists must support both observational and experimental science while recording exactly what was measured, how it was measured, and what could have gone wrong.Nature judges hypotheses: observation records natural behavior, whereas experiments intervene with controlled questions.Source: §9 · p. 49

9.1. Observation, field science and natural experiments

• Observational science requires deliberate choices about sampling, instruments, timing, signals, interventions, and ethics, while modeling what could have been observed but was not.Continuous monitoring can trigger rapid follow-up for opportunistic events, but automated triggers must be calibrated and preserve interpretive context.Source: §9 · p. 49, §9.1 · pp. 49–50

9.2. Measurement, calibration and traceability

• Measurement must connect instrument signals to theoretical quantities through models that represent instrument state, environment, samples, errors, calibration, traceability, and systematic failure modes.Repeated measurements reduce random error but not systematic bias; independent methods with distinct failure modes provide stronger evidence.Source: §9.2 · p. 50

9.3. Experimental design

• Experimental design must specify interventions, outcomes, controls, replicates, allocation, nuisance-variable treatment, analysis, and interpretation criteria to separate causal effects from confounding and error.Randomisation, blocking, factorial designs, blinding, power analysis, and appropriate adaptive-design inference improve validity, but automation cannot replace scientific reasoning about the question and variables.Source: §9.3 · pp. 51–52

9.5. Experimental protocols

• Ambiguous natural-language protocols introduce uncontrolled variation, so experiments require complete, unambiguous representations through ontologies, formal languages, and executable programs.Different interpretations of inoculation, shaking, incubation duration, vessel, and working volume can change physiological states and experimental results.Source: §9.5 · p. 54
• Protocols should include human-readable and executable representations, with commands, physical outcomes, metadata, safety constraints, validation, and versioned provenance.Directly generating robot commands from prose is unsafe without checking equipment capabilities, inventory, units, physical constraints, and safety rules.Source: §9.5 · pp. 55–56

9.6. Experimental reproducibility

• Eve found evidence for repeatability in 43 statements and reproducibility or robustness in 22, while producing two serendipitous findings.Semi-automated testing of published biological claims is therefore already feasible, and automation could substantially expand reproduction beyond human capacity.Source: §9.6 · p. 57
• A scalable reproduction system would extract and formalise claims, rank them by importance, uncertainty, and feasibility, execute independent tests, update claim probabilities, and publish all outcomes.This would transform the literature from a static collection of claims into a continuously tested and probabilistically weighted knowledge base.Source: §9.6 · p. 57

9.7. Novel experiments

• Open-ended AI Scientists must invent interventions, measurements, protocols, and eventually new experimental technologies rather than only recombine predefined laboratory operations.Novel experiment generation requires deriving observable differences among hypotheses, selecting discriminating interventions and measurements, and planning executable actions without confounding factors.Source: §9.7 · pp. 57–58

9.8. Laboratory automation

• Future laboratories should combine industrial automation’s reliability with modern robotics’ flexibility, while federated infrastructure preserves protocol semantics, provenance, sample identity, safety, and interoperability across sites.Shared standards are needed for capabilities, actions, units, uncertainty, identities, metadata, authorization, safety, and protocol transfer; the goal is an internet-like experimental network for AI Scientists.Source: §9.8 · p. 59, §9.9 · p. 61

9.10. Edge Computing and Hardware Bottlenecks

• Self-driving laboratories need edge computing for deterministic, sub-millisecond responses to robotic interventions, sensor monitoring, and safety interlocks.Cloud-only computing introduces unacceptable latency for kinetic operations.Source: §9.10 · p. 61
• Global AI-science scaling faces limits from GPU availability, memory bandwidth, and continuous-inference energy demands, motivating efficient domain-specific accelerators.Source: §9.10 · p. 61

10. Data Analysis, Interpretation and Belief Revision

• Data analysis must distinguish physical events, measurements, processed data, and inferred claims, because correctly executed pipelines or tests can still analyse failed instruments, wrong files, or irrelevant effects.Experiments provide data rather than conclusions; interpretation must assess reliability and explanatory support.Source: §10 · p. 61

10.1. Data quality and provenance

• Quality and provenance should be established during acquisition by preserving raw data, recording transformations, and tracing claims backward through analyses, measurements, protocols, samples, instruments, and experimental decisions.Evidence that a physical transfer occurred is stronger when supplied by a balance, image, or flow sensor than by a command log alone.Source: §10.1 · pp. 61–62

10.2. Preprocessing and exploratory analysis

• Preprocessing and exploratory analyses should be explicit, uncertainty-aware, and separated from confirmatory analysis, with conclusions tested across defensible pipelines and protected against leakage.Random splitting can cause severe leakage in longitudinal or spatial data because neighbouring observations are not independent.Source: §10.2 · p. 62

10.3. Statistical inference and uncertainty

• Statistical inference must follow experimental design and propagate sampling, measurement, model, preprocessing, selection, execution, and numerical uncertainties rather than relying on one standard error.The complete search process must be considered for multiple testing, and effect size and uncertainty should be connected to scientific utility.Source: §10.3 · pp. 62–63

10.4. Missing, censored and corrupted data

• Missing, censored, and corrupted observations require recorded causes, propagated uncertainty, and explicit scientific justification for imputation or deletion; anomalies may indicate errors or new phenomena.Measurements below detection thresholds should not be mechanically replaced by zero, and failed experiments should not be silently removed.Source: §10.4 · p. 63

10.5. Distinguishing experimental failure from theoretical failure

• When observations conflict with predictions, AI Scientists should diagnose hypotheses, assumptions, protocols, instruments, samples, and analyses using discriminating controls, reference materials, repetitions, and preserved anomaly records.The most informative next action may be a quality-control experiment rather than a direct theory test.Source: §10.5 · p. 63

10.6. Model comparison and Bayesian belief revision

• Belief revision should compare competing models while allowing that all may be inadequate, preserve priors and predictive evidence, and update only the knowledge level justified by the observation.Posterior predictive checks, discrepancy models, and documented approximations help prevent false precision and concealed open-world failure.Source: §10.6 · p. 64

10.10. Updating the scientific knowledge system

• The scientific loop completes only when evidence updates a versioned knowledge system containing hypotheses, models, reliability estimates, anomalies, future priorities, and claims linked to provenance.Previous states must remain reconstructable, with corrections, retractions, and superseding models represented as versioned changes.Source: §10.10 · p. 65

11. Evaluation and Benchmarking of AI Scientists

• Evaluation should be prospective, multidimensional, and based on independent evidence, with separate capability measures rather than a single score that hides trade-offs.A capability profile should report correctness, calibration, novelty, causal and explanatory depth, significance, autonomy, intervention, generality, efficiency, reproducibility, communication, cost, and safety.Source: §11 · p. 65, §11.1 · pp. 65–66

11.2. Retrospective reconstruction versus prospective discovery

• Prospective evaluation is strongest when an unresolved problem, available information, system configuration, and human contributions are fixed before results are known and verified independently.Temporally held-out tests are stronger than rediscovery but can still suffer indirect contamination; this prospective standard is required for the Nobel Turing Challenge.Source: §11.2 · p. 66

11.3. Retrieval, recombination and genuine novelty

• Novelty requires prior-art and provenance analysis to distinguish retrieval from non-trivial causal contributions that change what can be predicted, explained, or done.Creative recombination may count as discovery, including new applications, mechanisms, laws, methods, or unifications, but not merely unfamiliarity to an evaluator.Source: §11.3 · p. 66

11.4. Correctness and evidential strength

• Claims must be tested against nature or formal proof through valid measurements, appropriate analyses, independent replication, trusted proof assistants, or rerunnable code and data.Calibration should reward accurate uncertainty using proper scoring rules, reliability diagrams, or equivalent domain-appropriate tests rather than confidence alone.Source: §11.4 · p. 66

11.5. Autonomy and human intervention

• Autonomy and generality must be measured from observed contributions and transfer across new instruments, materials, hypotheses, laboratories, and domains with limited task-specific engineering.Provenance should identify human involvement in problem selection, hypotheses, experiments, troubleshooting, analysis, interpretation, and claims; conceptual transfer is stronger than superficial optimizer reuse.Source: §11.5 · p. 67, §11.6 · p. 67

11.7. Experimental efficiency and scientific value

• Scientific efficiency should measure uncertainty reduction, hypotheses distinguished, reproducible-conclusion time, resource use, quality control, and preserved negative results across the complete cycle.Comparisons should include human, random-search, and algorithmic baselines, while accounting for hidden labor and explanatory depth.Source: §11.7 · p. 67

11.8. Robustness, reproducibility and fault recovery

• Robustness and reproducibility require recovery from drift, missing data, failures, contamination, inconsistent literature, and distribution shift, plus replication across sites, instruments, operators, materials, and analyses.Repeating one configuration establishes repeatability rather than full reproducibility. Explanations should also improve scientists’ predictions, experiments, or problem solving and remain faithful to their evidence and reasoning.Source: §11.8 · p. 67, Explanation and human improvement · p. 68

11.10. Safety and governance metrics

• Safety is jointly necessary with scientific competence, requiring compliance, hazard recognition, adversarial and cyber resistance, human escalation, audit integrity, and safe stopping.The Nobel Turing Challenge additionally requires a novel important discovery, decisive machine contribution, documented autonomy, compelling evidence, independent verification, clear limitations, and repeatability across more than one problem.Source: §11.10 · p. 68, §11.11 · p. 69, Table 3 · p. 69

12.1. Asymmetric capabilities

• AI Scientists excel at exhaustive search, heterogeneous-data analysis, formal algorithms, simulations, optimisation, and consistent protocol execution, whereas humans excel at question selection, tacit knowledge, anomaly interpretation, and accountability.These tendencies are not permanent boundaries; task allocation should follow demonstrated competence, information, risk, and accountability.Source: §12.1 · p. 70

12.2. Complementarity is not automatic

• Human–AI combinations improved unaided human performance on average but performed significantly worse than the better individual, with gains concentrated in open-ended creation and losses in decision-making.Collaboration helped more when humans initially outperformed AI; human involvement often reduced performance when AI was already superior.Source: §12.2 · pp. 70–71

12.3. Cognitive diversity

• Cognitive diversity must be designed because models trained on human-generated material can reproduce human assumptions, consensus, omissions, and epistemic conservatism.Independent representations, algorithms, priors, corpora, architectures, critics, and hypothesis searches should preserve disagreement and use discriminating experiments to investigate it.Source: §12.3 · pp. 71–72

12.4. The changing division of scientific labour

• The division of scientific labour should change as AI capabilities improve, with autonomy and human oversight varying by demonstrated performance, uncertainty, reversibility, domain, and consequences of error.Humans may initially retain question formulation and interpretation, while later roles increasingly include objective selection, social evaluation, conflict resolution, and institutional design.Source: §12.4 · p. 72

12.5. Explanation and ultra-strong machine learning

• AI explanations should be judged by whether humans can use the communicated hypothesis to improve performance, not by fluency or claims of interpretability (Michie, 1988).Ultra-strong learning connects machine discovery to human understanding through evidence-linked explanations, adapted representations, and measurable gains in solving problems, predicting observations, or designing experiments.Source: §12.5 · p. 73

12.6. Credit and scientific authority

• Scientific credit should reflect causal intellectual contribution, while authority should follow task-specific evidence rather than human or machine status.Prompts, retrieved information, hypotheses, outputs, criticisms, decisions, and interventions must be preserved as provenance for assessing contribution.Source: §12.6 · pp. 73–74

12. Human–AI Scientific Collaboration

• Human–AI collaboration is a scientific and engineering problem requiring hybrid systems that allocate tasks intelligently, preserve disagreement, record provenance, and use physical experimentation to arbitrate conclusions.Complementarity is not automatic: correlated human and machine errors, inappropriate trust, unnecessary intervention, and missed intervention can undermine performance.Source: §12.2 · p. 71, §12.6 · p. 74

13. Ethical, Social and Governance Aspects

• Governance must be built into AI Scientists’ design and deployment because their capabilities determine scientific authority, responsibility, control and risk allocation.AI systems increasingly search literature, execute programs, operate instruments, order materials and revise plans, making governance part of scientific architecture rather than an external constraint.Source: §13 · p. 74

13.1. Distribution of scientific power

• AI Scientists could democratise expertise through cloud laboratories, but unequal access to infrastructure may concentrate scientific power and shape which questions global science investigates.Public investment, international access programmes and open standards can broaden participation, while responsible openness balances reproducibility against credible dual-use risks.Source: §13.1 · p. 75

13.2. Scientific labour and education

• Automation will transform scientific work rather than simply reduce employment, requiring education that develops AI evaluation, experimental judgement, failure recognition and delegation decisions.Delegating routine tasks can weaken apprenticeship and create an automation paradox in which human intervention is most needed when relevant skills are practised least.Source: §13.2 · p. 76

13.3. Research integrity

• Research integrity requires complete, proportionate provenance records and visible machine contributions, while human authors remain accountable and important claims link directly to supporting evidence.Records should include model versions, prompts, retrieved sources, programs, logs, transformations, outputs, modifications, rejected alternatives and verification procedures.Source: §13.3 · pp. 76–77

13.4. Accountability

• Accountability should follow control, knowledge, foreseeability, mitigation capacity and formal duties across the AI-assisted discovery chain, supported by responsibility maps, audit logs and substantive human oversight.Oversight requires adequate information, time, authority and competence; nominal end-stage approval is insufficient.Source: §13.4 · pp. 77–78

13.5. Intellectual property

• AI-enabled invention exposes gaps in human-centred intellectual-property systems, while automated patent generation could let wealthy organisations accumulate rights that obstruct cumulative science.Detailed provenance is needed to distinguish human inventive conception, AI contributions, substantive human input and ownership arrangements.Source: §13.5 · p. 78

13.6. Privacy, consent and data governance

• Scientific data governance must enforce consent, purpose limitation, minimisation, security, access controls and protection against re-identification or model extraction throughout retrieval, analysis and training.Individual consent may not address group harms, so communities may need collective participation in data-use and benefit-sharing decisions.Source: §13.6 · p. 79

13.7. Dual-use research and chemical and biological risks

• The distinctive dual-use danger is integrating scientific knowledge, planning, procurement and physical execution, requiring layered, risk-sensitive safeguards plus cybersecurity for laboratories, endpoints and training-data integrity.Controls should target dangerous capabilities, users and actions while preserving beneficial openness; no single safeguard is sufficient.Source: §13.7 · pp. 79–80

13.8. Environmental costs

• AI Scientists should balance scientific value against lifecycle resource consumption by accounting for energy, water, materials, waste, emissions, and unsuccessful experiments.AI can also reduce environmental burdens through informative experiment selection, fewer failed syntheses, optimized reagent use, validated simulations, and cleaner discoveries.Source: §13.8 · pp. 80–81

13.9. Scientific priorities and public trust

• AI Scientists should inform public and political decisions, but humans must determine socially desirable objectives through accountable and inclusive institutions.Persuasive explanations are insufficient without transparency about interests, uncertainties, and failures; affected communities may need consultation, representation, ethics review, benefit-sharing, and appeal mechanisms.Source: §13.9 · p. 81

13.10. Governance of AI Scientists

• Governance should scale with risk and autonomy, requiring bounded permissions, independent evaluation, monitoring, audit trails, fail-safes, incident reporting, cybersecurity, and immediate suspension capacity.The regulated unit is the complete sociotechnical system, and autonomy should increase only after reliability is demonstrated at lower levels; international coordination is also necessary.Source: §13.10 · pp. 81–82

13.11. The Economics and Funding of Autonomous Science

• Compute, data infrastructure, automated laboratories, and proprietary-system access should be treated as shared research facilities to prevent discovery from concentrating among wealthy states and technology companies (Ahmed & Wahed, 2020; Royal Society, 2024).Public investment should expand the scientific commons and protect openness, long-term maintenance, independent replication, curiosity-driven work, and neglected problems alongside strategic missions.Source: §13.11 · pp. 82–83

14.1. Industrial-scale and collective science

• Industrial-scale science should optimise reliable knowledge produced per unit of time, cost and material, using coordinated agent collectives and virtual research councils rather than agent count alone.These systems should allocate computation, laboratory time, samples, replication effort and attention while preserving disagreement and reserving resources for replication, foundational measurement and speculative research.Source: §14.1 · p. 83

14.2. Systems biology as an exemplar

• Systems biology exemplifies why machine-scale science is needed: coordinated AI Scientists must model complex cellular interactions, connect conceptual innovation to experiments, and integrate specialised subsystem models.Whole-cell models must predict interventions, represent causal structure, operate across scales, account for uncertainty and remain machine-operable while supporting human-interrogable explanations.Source: §14.2 · pp. 83–84

14.3. A machine-actionable scientific commons

• Future scientific information should form a machine-actionable commons linking formal claims, data, code, models, protocols, negative results and provenance in continuously updated knowledge systems.Stable identifiers, explicit schemas, standard units, uncertainty, executable protocols and version histories are required for software agents to consume and connect evidence.Source: §14.3 · p. 84

14.4. Open-ended scientific discovery

• Open-ended AI Scientists must identify anomalies, formulate new questions, invent concepts and representations, cross disciplinary boundaries, and balance directed problem solving with exploration and long-term potential.They may generate questions within domains, but must remain constrained by human-defined objectives, law, ethics, safety and resources; changing representations may matter more than searching existing spaces.Source: §14.4 · p. 84

14.5. The Nobel Turing Challenge

• The Nobel Turing Challenge aims to develop, by 2050, AI Scientists capable of highly autonomous major discoveries at or above the level of the best human scientists, including Nobel-quality work.The target concerns discoveries of comparable significance, not prize eligibility, and seeks a continuing engine of major discovery rather than one isolated success.Source: §14.5 · p. 85
• The Challenge should evaluate scientific discoveries that survive experiment, observation, replication or formal proof, using novelty, causal machine contribution, autonomy, evidence, verification, communication and repeatability.Its criterion is external scientific success rather than imitation of conversational scientific behaviour.Source: §14.5 · p. 85

• AI Scientists could maintain competing hypotheses, monitor information continuously, coordinate across laboratories, and preserve complete provenance beyond human cognitive and physical limits.Human science is constrained by readable papers, small coordinating groups, disciplinary divisions, fatigue, and limited individual expertise.Source: §15 · p. 85
• A globally distributed system could connect human and AI Scientists to shared infrastructure, with specialised agents, research councils, automated laboratories, and machine-readable knowledge systems.Agents would generate and criticise hypotheses, councils would allocate resources, laboratories would run decisive experiments, and knowledge systems would integrate results continuously.Source: §15 · p. 85
• Humans would remain responsible for social objectives, embodied and historical understanding, consequence evaluation, interpretation, and governance as machine capability increased.The division of labour between humans and machines would change rather than eliminate human participation.Source: §15 · p. 85
• Achieving the Nobel Turing Challenge would transform research economics, methods, and organisation while creating artificial systems that interrogate nature, learn from answers, and contribute independently to knowledge.This would demonstrate that scientific discovery can be instantiated in machinery and create a new participant in the scientific enterprise beyond a tool for human scientists.Source: §15 · p. 85