AI Is Compressing Research Support While Raising the Value of Frontier Science.

Pure scientific research is one of the worst places to use lazy automation language.

AI is not simply “replacing scientists.” It is attacking some of the most repetitive knowledge work in the research system while simultaneously raising the strategic value of other scientific roles. In some parts of science, AI reduces labor. In other parts, it creates entirely new workflows, new labs, and new demand for highly specialized people.

That is why the underlying March 25, 2026 assessment matters. Across 57 roles, the industry lands at an average AI replacement rate of roughly 34.9%, which places pure science and basic research in the limited-assistance band overall.

But that average hides a much more important pattern:

research-support roles are under real pressure,
experimental and theoretical science remain much harder to replace,
and AI-native scientific roles are expanding rather than shrinking.

This Is a Huge System, but AI Hits It Unevenly

The source file places total global R&D spending at roughly $2.87 trillion in 2024, with basic research accounting for around 15% to 20% of that total. That implies a global basic-research base of roughly $430 billion to $570 billion. It also cites:

U.S. basic-research spending around $125 billion in 2024,
China basic-research spending around $62 billion,
and a global research workforce above 9 million FTE-equivalent researchers.

Inside that system, the AI layer is growing much faster than the rest:

AI for scientific discovery at around $4.8 billion in 2025, projected toward $34.78 billion by 2035,
growing adoption of self-driving laboratories,
and large-scale AI tooling in bioinformatics, molecular design, materials discovery, and literature review.

This matters because science is not a single labor market. A molecular-biology wet lab, a theoretical-physics department, a national lab materials platform, and an academic publishing operation are all “science,” but AI affects them very differently.

The Adoption Wave Is Already Real

The source assessment points to a set of milestones that make it impossible to treat AI in science as a future-only story:

AlphaFold used by more than 3 million researchers across 200+ countries,
GNoME predicting 2.2 million candidate compounds, with 736 already experimentally validated,
more than 50 self-driving labs operating globally by late 2025,
more than 50% of researchers using AI in peer review according to cited Nature reporting,
and roughly 60% to 70% of researchers having used LLMs to assist scientific writing.

Those are not fringe signals. They show that AI is already embedded in the modern research workflow.

The Real Divide in Science Is Information Work vs Scientific Judgment

The source file makes one thing very clear: the most exposed work in science is not the highest-prestige work. It is the work built around information handling.

That includes:

literature search,
data curation,
first-pass statistics,
manuscript drafting,
scientific illustration,
and some research-support administration.

The least exposed work tends to involve:

setting scientific direction,
inventing new theory,
handling messy experimental reality,
making safety-critical lab decisions,
or designing new AI-native research systems.

This is why the labor effect of AI in science feels contradictory. The more the job looks like structured interpretation over large document or data corpora, the more exposed it becomes. The more it depends on creativity, tacit lab skill, physical execution, or scientific taste, the more resilient it remains.

The Highest-Risk Jobs Sit in Research Support

The most exposed jobs in the source assessment are mostly support and output-formatting roles rather than flagship scientific roles.

The highest-exposure roles in the source assessment

Role	Estimated AI replacement rate	Why exposure is high
Scientific Literature Search Specialist	82%	AI search, ranking, extraction, and review tools already automate large parts of the workflow
Scientific Paper Writing Specialist	72%	Structured drafting, formatting, and first-pass synthesis are increasingly machine-generated
Research Illustrator	65%	Figure generation, layout, and visualization automation are advancing quickly
Statistical Analyst	62%	Standard modeling and first-pass statistical workflows are increasingly software-native
Research Data Administrator	60%	FAIR-data handling, metadata tagging, and structured governance fit automation well
Research Assistant	55%	Entry-level search, note-making, formatting, and routine analysis tasks are exposed

This is the first major shock. AI is not only helping scientists do their work. It is directly absorbing a large share of the labor once carried by support functions surrounding science.

Literature Review May Be the Clearest Casualty

The strongest single example is literature retrieval.

The source cites tools such as Elicit, Consensus, and related systems as searching 138 million+ papers, while other cited tools raise screening efficiency by around 90%. Once that becomes true, the old model of manual literature discovery changes permanently.

A human still matters for framing the right question, deciding what counts as a relevant paper, and separating genuine signal from fashionable noise. But the operational burden of retrieval, deduplication, extraction, and summarization is already moving toward software.

That is why the literature-search role tops the exposure table.

Scientific Writing Is Also Being Rebuilt

Scientific paper writing specialists and documentation-heavy roles are also highly exposed. That should not be surprising.

Modern AI systems are already good at:

transforming notes into structured prose,
drafting methods and background sections,
converting experimental records into formal templates,
summarizing prior literature,
and producing cleaner first drafts.

This does not mean AI can produce credible final science on its own. It does mean the low-value drafting layer is shrinking rapidly. Researchers and editors increasingly spend time reviewing AI-assisted drafts rather than producing every paragraph from scratch.

That changes staffing logic in universities, journals, and research-support offices.

Theory and Frontier Judgment Remain Much Harder to Replace

The least exposed jobs in the assessment are the positions where science still depends on deep intuition, creative framing, and non-routine judgment.

The lowest-exposure roles in the source assessment

Role	Estimated AI replacement rate	Why exposure stays low
Institute Director	7%	Strategy, funding politics, reputation, and institutional leadership remain human
Chief Scientist	8%	Scientific direction and intellectual leadership are still human-led
Lab Director	12%	Team leadership, safety responsibility, and experimental oversight remain human
AI-Assisted Drug Discovery Researcher	15%	This is a new high-value role created by AI adoption rather than eliminated by it
Automation Laboratory Engineer	15%	Building self-driving labs requires hardware-software integration and experimental design judgment
Mathematician	15%	Original conjecture and proof strategy remain highly human
Animal Experiment Technician	15%	Ethical oversight and physical execution keep automation limited

The pattern is consistent. The safest jobs either sit at the top of the scientific hierarchy or at the hard physical edge of research work, where tacit skill, safety, or originality still dominate.

AlphaFold Changed Science Without Ending Biology

The source file treats AlphaFold, IsoDDE, and other scientific-AI milestones as major turning points, and rightly so.

AlphaFold turned protein-structure prediction from an exceptionally slow and difficult process into something far more accessible. GNoME transformed the scale of candidate-material generation. Self-driving labs showed that closed-loop experimentation can move much faster in carefully structured environments.

But these milestones did not make scientists obsolete.

Instead, they changed what scientists spend time on:

less brute-force search,
more validation,
more interpretation,
more decision-making about which AI-generated candidates are worth pursuing,
and more work integrating AI outputs into coherent scientific programs.

That is why biology, chemistry, and materials science do not disappear in the source assessment. They reorganize.

Experimental Science Still Has a Strong Physical Barrier

Wet-lab work remains meaningfully protected, even as AI improves upstream and downstream reasoning.

Roles such as:

molecular biologist,
organic chemist,
tissue-culture technician,
electron-microscope operator,
environmental scientist,
agricultural scientist,
and biosafety officer

remain relatively resilient because a large share of their value still depends on:

physical manipulation,
real-time troubleshooting,
sample variability,
safety judgment,
and context that is not fully captured in structured digital input.

That is especially obvious for:

animal experimentation,
biosafety oversight,
lab safety,
and complex experimental setup.

The more a role depends on embodied skill and local judgment under messy conditions, the lower the replacement rate tends to be.

Self-Driving Labs Change the Labor Mix, Not the Need for Scientists

Self-driving labs are one of the most important themes in the source file. The assessment points to a growing global SDL footprint and to research systems that integrate AI planning, robotic execution, automated measurement, and iterative optimization.

This matters enormously. But it still does not add up to “AI runs science.”

What it really means is:

fewer people may be needed for some repetitive experimental loops,
more value concentrates in those who can build, supervise, debug, and strategically direct those systems,
and lab work becomes more polarized between commoditized execution and high-end system control.

That is why automation lab engineers and other AI-native science roles look resilient in the source. AI is creating them faster than it can eliminate them.

Science Also Faces an Intellectual Risk: Higher Output, Narrower Exploration

One of the most important strategic warnings in the file comes from cited reporting that AI users can produce 3x more papers and receive 5x more citations, while also contributing to a narrower scientific search pattern.

That is a real risk.

If AI increasingly steers researchers toward the same obvious literatures, standard methods, and mainstream questions, science may become more efficient while becoming less exploratory. In other words, AI may supercharge production while reducing diversity of inquiry.

That is one reason frontier judgment remains so valuable. Someone still has to decide when to ignore the most legible path and pursue a strange one.

The Postdoc Layer Is Vulnerable in a Very Specific Way

The assessment’s treatment of postdocs and early-career researchers is especially important.

Postdocs do not sit in the highest-exposure band overall, but they are exposed in a structurally dangerous way. The risk is not that all postdocs disappear. It is that the bottom layers of analytic and writing work that once trained early-career scientists get compressed, especially in computational disciplines.

That means:

computational postdocs face more direct AI pressure,
experimental postdocs remain safer,
and the training pipeline itself may change because junior scientists spend less time doing the repetitive work that used to teach them how the system worked.

The consequence is not just labor displacement. It is a pipeline problem for future scientific leadership.

What This Means for Scientific Institutions

The industry is not facing one decision. It is facing three.

First, which parts of the research workflow should be automated now:

literature retrieval,
research support,
first-pass statistical work,
figure generation,
structured reporting,
and metadata management.

Second, which parts should be redesigned around human oversight:

experimental design support,
grant management,
peer-review assistance,
computational analysis,
and scientific communication.

Third, which parts should stay clearly human:

frontier theory,
institutional leadership,
biosafety and lab safety,
high-stakes experimental judgment,
and the final interpretation of novel scientific claims.

The mistake would be to ask whether AI can replace scientists. The right question is which layers of the scientific labor stack are becoming infrastructure, which are becoming supervision work, and which are becoming more scarce because AI makes them more important.

The Structural Conclusion

Pure science is not being automated from the top down. It is being compressed through the support layer, reshaped in the middle, and defended at the frontier.

The first work to be absorbed is:

search,
drafting,
formatting,
routine statistics,
data handling,
and other repeatable knowledge-processing tasks.

The work that remains hardest to replace is:

scientific direction,
deep theory,
difficult experimentation,
lab leadership,
safety responsibility,
commercialization judgment,
and the design of AI-native research systems.

So the future of science is not “AI scientist replaces scientist.”

It is closer to this:

AI removes large amounts of support labor,
raises the throughput of capable researchers,
creates new roles around automated discovery infrastructure,
and forces institutions to decide whether they are optimizing merely for output or for real discovery.

AI can accelerate science dramatically. It still does not know, by itself, what is worth discovering.