What changes when general-purpose AI and specialised scientific models enter the research workflow?
Based on: Codreanu, Imas, Mateos-Garcia et al., AI in Science: Early Insights, September 2026. Supplied research paper, 42 pages.
The paper is an early measurement of adoption, not a claim that AI has solved discovery.
Researchers report saving almost seven hours a week. Much of that time returns to research, but scientists also report more untested hypotheses and a large verification burden.
Source: supplied paper, pp. 1, 23-26. Survey results are self-reported and the sample is not representative of all scientists.
The benefit is not simply “more AI”. It is a better division of labour across the research system.
Time saved on analysis, code, literature work and communication can be reinvested into output, experiments or harder questions.
AI can lower the cost of searching, translating and recombining ideas across fields.
Verification, physical experiments, clinical validation, data quality and provenance become more valuable as generation speeds up.
Beginner translation: if AI helps you create ten plausible hypotheses instead of two, the next question is not “can the model generate more?” It is “which ones can we test, and how do we know the result is sound?”
Source: supplied paper, pp. 25-31. The beginner translation is an interpretation of the paper’s bottleneck and verification findings.
The paper’s data spans general LLM interactions and specialised model capabilities. Together they cover more of the workflow than either category alone.
Life sciences: use a general model to troubleshoot analysis code, then call a specialised model to predict a molecular structure or outcome.
Health research: use an LLM to organise a protocol or compliance documentation, then keep clinical validation and evidence review under explicit human control.
The useful product is likely a connected workflow, not one model pretending to do every job.
Source: supplied paper, pp. 12-18, 29-31. Examples are grounded in the paper’s task categories and named model types; they are not reported case studies from the paper.
How is AI entering science, what tasks does it touch, and what changes do researchers perceive?
15 million anonymised Gemini interactions. A classifier isolated about 360,000 likely science interactions.
An inventory of 2,690 specialised AI models linked to publications and code repositories, with 460,000 unique citations since 2020.
A survey of 637 active researchers in the US and UK about adoption, time, bottlenecks and research choices.
Why combine them? Logs show observed interactions, models show available capability, and the survey asks about time and bottlenecks. No single source can answer all three.
Source: supplied paper, pp. 3-8, 22-23. The paper calls the model inventory a work in progress, not an exhaustive registry.
In the paper’s Gemini sample, scientific occupations appear more often than their share of US employment would predict.
The paper also observes LLM usage in 195 of 217 granular scientific subfields above its privacy threshold. Shares are interaction volume, not researcher counts or hours.
At a broad level, both help with quantitative analysis. At a granular task level, their roles diverge.
The product insight: orchestration becomes more valuable when the system can route a task to the right model, carry context across steps, and preserve evidence about what happened.
Source: supplied paper, pp. 17-19, 29-30. The complementarity conclusion is an association from task distributions, not a causal test.
Researchers say AI frees time. The paper’s more interesting question is where that time goes.
Just under 30% say saved time goes into more overall research output.
About 21% reinvest it in lab execution and data collection.
About 19% use the dividend to tackle more ambitious scientific problems.
Source: supplied paper, pp. 23-24. These are self-reported survey responses, not measured causal productivity effects.
When analysis, drafting and hypothesis generation get cheaper, physical experimentation, clinical validation and field data collection can become the rate limiter.
Source: supplied paper, pp. 25-26. The funnel is an explanatory visualisation of the paper’s bottleneck-shift finding, not a measured funnel ratio.
AI output is not turn-key scientific evidence. Researchers still need to audit, debug, fact-check and reproduce it.
Record source documents, model versions, prompts, tools, parameters and intermediate outputs.
Make it possible to rerun the step or inspect why a result was accepted.
Assume wrong documents, unsupported claims, missing provenance and confident errors.
Keep the scientist responsible for standards of evidence and the cost of physical testing.
Source: supplied paper, pp. 25-31. The workflow controls are recommendations derived from the verification findings.
The survey contains a tension worth taking seriously: more interdisciplinary access, but a tilt towards tractable, benchmarked problems.
Interpretation: the same tools that help recombine ideas can also make data-rich, well-benchmarked work disproportionately easy. That “streetlight effect” is a hypothesis to test, not a settled conclusion.
Source: supplied paper, pp. 26-31. The paper reports 28% saying AI encouraged higher-risk questions; it also flags the streetlight effect as a possible interpretation.
An LLM plans and coordinates. Specialist models do domain work. Evidence and verification remain first-class outputs.
Define what would count as a useful answer before selecting a model.
Use the general model for coordination and the specialised model for domain-specific work.
A result without provenance is a lead, not a conclusion.
The scientist decides what is strong enough to act on or test.
Source: supplied paper, pp. 30-31. Architecture is a recommended implementation pattern inferred from the paper’s proposed LLM orchestrator.
Start with one bounded workflow where verification matters and the hand-offs are observable.
Pick a narrow task such as literature-to-hypothesis synthesis, assay-data triage or protocol comparison. Define the user, inputs and decision it supports.
Separate what the general LLM does, what a specialised model does, what needs a human, and what depends on a physical or external test.
Log source IDs, model/version, prompt or parameters, tool calls, timestamps, outputs and reviewer decisions.
Require citations, schema checks, unit checks, domain tests, uncertainty flags and human sign-off before an output becomes a research artefact.
Track time saved, rework, error types, verification time, untested-hypothesis backlog and how many outputs reach validation.
A strong benchmark asks whether the whole system helps researchers make better, safer progress.
Assemble representative historical cases with known-good answers, accepted sources and expert annotations. Include hard negatives and ambiguous cases.
Test wrong-document retrieval, missing provenance, hallucinated dosage or units, unsupported extrapolation, tool failure and overconfident summaries.
Measure factuality, citation correctness, retrieval recall, time-to-decision, verification burden, escalation rate and downstream validation success.
Map the same architecture to an adjacent domain, such as regulated finance or engineering, and identify which controls transfer and which need redesign.
Rule of thumb: if the system cannot show where a claim came from and how it was checked, it has produced a draft, not a result.
Source: the paper’s limitations and discussion, pp. 27-31. The golden-set and failure-mode checklist is a practical recommendation for implementing the paper’s verification agenda.
The paper’s lesson is useful for any high-stakes agent that turns messy evidence into a decision or next action.
Use the general model to coordinate literature, sensory data and experimental design. Keep evidence lineage and assay or sensory validation visible.
First slice: provenance-aware hypothesis briefUse retrieval and orchestration to structure multimodal records. Treat provenance, uncertainty and clinical review as part of the output.
First slice: depersonalised timeline synthesisUse the agent to gather and summarise project memory, but preserve source links, time boundaries and a clear distinction between record and interpretation.
First slice: evidence-linked weekly digestProduct principle: the durable advantage is not “an AI that knows science”. It is a workflow that makes scientific reasoning easier to inspect, challenge and continue.
Source: paper findings on orchestration, provenance, verification and bottlenecks, pp. 25-31. Project examples are tailored recommendations, not claims from the paper.
Scientists are using general and specialised AI across a wide range of research tasks.
General LLMs coordinate and support broad work; specialist models handle domain-specific prediction and generation.
But the dividend is partly consumed by verification and downstream execution.
Research systems need better validation capacity, provenance, benchmarks and decisions about risk.
Gemini logs, a curated model inventory and a US/UK survey only show part of the global landscape.
LLM classifiers map messy interactions and model abstracts into a taxonomy, with known errors.
The paper does not establish that AI caused the reported productivity or workflow changes.
Will AI help science take bigger bets, or mostly make tractable work cheaper?
Primary source: Mihai Codreanu, Alex Imas, Juan Mateos-Garcia et al., AI in Science: Early Insights, September 2026, supplied PDF, 42 pages.
Deck note: Percentages and sample sizes are reproduced from the supplied paper and labelled where they are self-reported, inferred or descriptive. This deck is an educational explainer, not professional, medical or legal advice.