AI for Science

• Where do scientific paradigms live?
• What role does human understanding and agency play?
• What’s an emerging playbook for AI-for-science?

• Paradigms: shift in science or in underpinning information infrastructure?
• Understanding: what do we still demand (and from whom)?

• Questions we care about are often qualitative (meaning, values, lived outcomes).
• Questions we can test are often quantitative (metrics, averages, effect sizes).
• AI may help bridge this gap by turning language and practice into analysable data.

• Go back go ICML 2015 Deep Learning Workshop
• Try to imagine a technology with mass impact on coding and creative work.

• Struggle whole of science across fields and institutions.
• LLM interfaces lower cost of moving across disciplines
• Increase the risk of “plausible” error.
• Main skill: skepticism

• Task capabilities: what the model can do (predict, reason, generate, control) (Bommasani et al., 2022; Krenn et al., 2022).
• Workflow needs: where value appears (hypothesis, design, analysis, writing, coding) (Arranz et al., 2023; Berman et al., 2024; European Research Council, 2023).
• Context constraints: data, compute, latency, validation culture, uncertainty tolerance (Duarte and others, 2018).
• See Cranmer et al. (2026) (in review).

• ML is strong at pattern extraction and function approximation.
• Science needs mechanism: why, when, and under what interventions.
• So many deployments are still guess-and-verify pipelines.
• Key question: when is prediction enough, and when do we require explanation? (Bender et al., 2021; Pearl and Mackenzie, 2018).

• Causality: distinguish correlation from interventionally robust structure (Pearl and Mackenzie, 2018; Schölkopf, 2022).
• Abstraction: discover useful intermediate scales and effective theories (Anderson, 1972; Jaynes, 1957).
• Simulation: hybrid mechanistic + learned systems with explicit validity regimes (Cranmer et al., 2020; Oreskes et al., 1994).

• A model family \(f_\theta\) that maps inputs to outputs.
• A learning objective that scores how well \(f_\theta\) matches data (and priors/constraints).
• Optimisation + scale: we fit \(\theta\) with large compute and large datasets.

• AlphaFold
• GraphCast/Aardvark

• If the training data are rigorous equations (PDE/ODE solvers), we have a clearer sense of “ground truth.”
• These models may not map onto human intuitions — but they can still be scientifically useful.
• That makes physics a promising sandbox for a science of AI: what is learned, what generalises, and how do we verify?

• LLMs emulate human behaviour
• Unsurprising that the work better “in collaboration”
• Also with rapid access to information infrastructure.

• Denario
• Claude Code
• Codex

• Science as technology
• Science as understanding

• Covid19 Epidemiological Modelling
• Does your model account for Facemasks?
• Does your model account for Hotel Closures?

We might not have data, but we can do some arithmetic

• Compare with technical debt.
• Highlighted by Sculley et al. (2015).

• Decompose your complex problem/task into parts.
• Each part manageable (e.g. by a small team)
• Recompose to solve total problem

• Highly successful approach to complex tasks.
• Tuned to the human bandwidth limitation.
• But the whole system still hard to understand.

• Technical debt is the inability to maintain your complex software system.
• Intellectual debt is the inability to explain your software system.

• Agentic AI could pay down technical and intellectual debt.
• But it can create agentic debt: delegation without authority/authorship

• Delegation of workflows without crisp boundaries.
• Agentic debt is about unsafe or illegible delegation.

• Agentic Orchestration compress reading/writing/coding into a single interface.
• Tool use turns text into action: search, code execution, lab automation, simulation pipelines.
• This shifts the bottleneck to verification and scientific judgement.

• Operational understanding: can I use it safely and know when it fails?
• Mechanistic understanding: do I have an interpretable causal/mechanistic story?
• Paradigm understanding: can the community reproduce, contest, and extend it?
• Social understanding: are the ideas understood in the wider public and other fields?

• The “judgement layer” in organisations is often tacit: norms, exceptions, escalation paths, and context that rarely makes it into documentation.
• It lives in handoffs and approvals: what gets challenged, what gets waived, and what triggers a halt.
• Ceding this tacit knowlege without making it explicit is how we accumulate agentic debt.

• Human communication: walking pace (2000 bits/minute)
• Machine communication: light speed (billions of bits/second)
• Our sharing walks, machine sharing …

• Databases / tables (OEIS): store patterns and prior results.
• Solvers (CAS, SAT/SMT): mechanised search and case analysis.
• Modern triad: proof assistants, machine learning, large language models.

• Digitally verifiable: proof assistants (e.g. Lean) and machine-checkable artefacts.
• Operationally reliable: code, simulators, pipelines — repeatable, but not always interpretable.
• Tacit + contextual: protocols, judgement, and field knowledge (bio/geo/social science).
• LLMs can compress and transmit tacit knowledge — but they push the bottleneck to verification boundaries.

• We can hold humans to account for judgement
• We can’t hold models (directly) to account
  • they don’t bear responsibility or liability
• Need a named author to sign off

• Governance gap: policy exists, practice often lags (European Commission: Directorate-General for Research and Innovation and Montgomery, 2025; Resnik and Hosseini, 2025).
• Research culture gap: weak incentives/careers for interdisciplinary team science (National Academy of Sciences et al., 2005).
• Infrastructure gap: compute and models concentrated in few actors (Bommasani et al., 2022; Lawrence and Montgomery, 2024).

• Narrative gap: corporate visibility can eclipse public infrastructure contributions (Cave et al., 2018; Moult et al., 1995).
• Coordination gap: funding/governance/data/skills often evolve separately (European Commission: Directorate-General for Research and Innovation and Montgomery, 2025).

• Funding for domain-grounded, interdisciplinary AI-for-science work.
• Governance for reproducibility, attribution, and safe use (Ball, 2023; Wachter et al., 2024).
• Infrastructure: open tooling + shared compute pathways.
• Data: trusted access and stewardship.
• Talent/skills: train scientists to explain, verify, and challenge (Cranmer et al., 2026; European Commission: Directorate-General for Research and Innovation and Montgomery, 2025).

• What do we want trainees to be able to explain, verify, and challenge?
• Where is the verification boundary in AI-for-science systems?
• Which artefacts are the new paradigm stores?

1. Is there an emerging playbook for AI-for-Science?
2. Are we converging on a standard recipe, accurate but slow simulators, amortized surrogates, differentiable pipelines, and how do we know when these surrogates truly generalize?
3. Can AI grapple with open-ended discovery?

4. Beyond supervised prediction, to what extent can current or near-future systems generate meaningful hypotheses, concepts, and research directions?
5. What tools do we have, and what tools are missing?
6. Which existing AI/ML capabilities are already reshaping scientific practice, and what critical tools (for uncertainty, causality, interpretability, or interfaces) are still absent?