For every paper that changes our understanding, there are several passes through the scientific life cycle, involving years of reading, hypothesising, planning, experimenting, analysing, interpreting and writing. Each stage places its own constraint — or bottleneck — on the rate of scientific progress. Until recently, the tools available to mitigate these bottlenecks were limited and rarely touched more than one at a time.

AI is changing this. It can already summarise literature, write and debug code, and evaluate manuscripts with unfathomable speed. But the deeper shift isn’t just that AI accelerates science, it’s that AI redistributes the labour of science itself. The redistribution is uneven, and what resists compression — collecting data, choosing good problems, and forming hypotheses worth testing — is exactly where science was always hardest.

AI is relieving bottlenecks in the scientific life cycle

A useful way to think about AI’s current impact on research is as a pipeline with several bottlenecks that constrain progress. Three are most clearly being alleviated by AI right now.

The first bottleneck is literature synthesis. Researchers feel overwhelmed by the sheer volume of papers being published within their niche, let alone the adjacent ones. AI offers a partial solution. Tools like Consensus, Elicit, and Edison can search, screen, summarise, and synthesise large corpora in minutes from a single prompt. This has substantially reduced the cognitive burden associated with finding relevant papers, extracting them and organising them into a coherent narrative. It doesn’t completely eliminate the need for human input on these tasks, but it definitely reduces the friction around them. Less of a researcher’s time is spent locating and summarising the relevant literature, and more can be spent thinking critically about it.

The second is data analysis. Modern biology generates high-dimensional datasets combining multiple assays, experimental conditions, and modalities. The work of wrangling, analysing and visualising data while ensuring that workflows are reproducible has historically consumed weeks or months of researcher time, often requiring specialised bioinformatics expertise that was scarce within most laboratories. AI coding tools like Codex and Claude Code now execute these tasks with a fluency that would have been unimaginable five years ago. Analyses that previously required a dedicated collaborator can now be drafted and iterated on by the scientist who generated the data, in a fraction of the time. During my PhD, I spent several weeks analysing metabolomics data that I could now complete in a single day. This is the clearest current example of AI’s democratising effect on research: technical capability that used to require specialised training is now available to anyone who can describe what they want in the English language.

The third is reviewing and publishing scientific papers. Manuscripts are getting harder to evaluate. As studies become more sophisticated and complex — combining multiple assays and omics, computational modelling, and genetic perturbation within a single paper — demand on peer reviewers has increased. A conscientious reviewer must check several components, like whether the conclusions follow from the data, whether the statistics are appropriate and whether the figures are internally consistent. That is a lot to ask from unpaid experts with limited time. However, AI is now able to take on some of that burden. Tools like Refine.ink and q.e.d perform some of the checks that a careful referee would: evaluating whether claims are supported by the evidence, flagging notation errors, and checking whether the abstract and conclusions overstate the reported results. As these tools mature, AI pre-screening may well become a mandatory step before submission and editorial assignment, filtering out many of the errors that currently consume reviewer attention. Both authors and reviewers stand to benefit, with the timeline from writing to publication likely being compressed substantially.

A brief note of caution. AI can produce valid-sounding literature syntheses, plausible-looking analyses, and polished manuscripts, but it can easily make mistakes — over-trusting weak literature, applying inappropriate analyses, and smoothing over uncertainty in prose. So our obligation to verify the outputs hasn’t disappeared. But ironically, there is a danger that as these processes get automated, the skill of auditing them fails to develop in junior researchers and even atrophies in those who previously did the work by hand. It will remain imperative to know what, for example, a comprehensive literature synthesis or complex statistical analysis entails, so that we can verify model outputs accordingly.

It is tempting at this point to reach for an analogy. Calculators transformed maths by turning arithmetic into a dependable black box. Once routine calculation became trustworthy, mathematicians could devote more attention to deeper problems. But AI is not yet the calculator for science across the board, and may never be. The reason is that a calculator performs deterministic operations whose outputs can be checked unambiguously, whereas AI outputs are probabilistic, context-dependent, and hardest to verify for tasks that don’t run as code and involve interpretation, such as literature synthesis and making inferences from data. This is why some scientific work will resist full automation for the foreseeable future. AI can undoubtedly reduce the cost of many intellectual tasks, but it won’t eliminate the need for human judgement. Even so, the productivity gains are likely significant, and this raises the question of where that saved time should go.

Human-centric bottlenecks are insulated from AI

As AI relieves bottlenecks in the research pipeline, the distribution of scientific work begins to look different. The time saved on literature synthesis, data analysis and writing can be spent on other critical bottlenecks. Two are worth highlighting because they are actually likely to remain refractory to AI for some time, though for quite different reasons.

The first is data collection. Biology is inherently slow and can’t be accelerated: cells grow and differentiate over a specific time course; you can’t speed up a genetic, pharmacological or dietary manipulation in animals; clinical cohorts take years to recruit and complete studies. Its protracted nature imposes a ceiling on how fast hypotheses can be tested. Furthermore, wet lab work is notoriously nuanced and hands on. Molecular biology and in vivo work is difficult to emulate using robots, and they entail tacit knowledge that is challenging to encode, meaning that they will remain resistant to AI for some time.

I’m not saying that AI can’t help at all on this front in the future. In fact, so-called self-driving autonomous labs are currently under development. These systems aim to conceive, plan, execute and analyse experiments, with minimal human input. An example of such was recently published in Nature, where a multi-agent system called Robin identified therapeutic candidates for dry age-related macular degeneration, proposing enhanced retinal pigment epithelium phagocytosis as a therapeutic strategy, validating two drugs in vitro, and using RNA-seq to implicate ABCA1 as a possible mechanism and novel target. But the technology is far from being widely available and useful in the life sciences, and the biology timeline ceiling inevitably remains. So although AI is accelerating processes upstream and downstream of the experiment itself, generating high-quality data still remains a rate-limiting step to scientific progress.

The second is hypothesis generation. Science’s biggest bottleneck is the creation of good explanatory theories. And despite the lavish claims from companies responsible for developing the frontier models, AI isn’t going to directly relieve this constraint, because it can’t create truly novel explanatory theories. This is a consequence of its design. It interpolates within the distribution of what is already latent in its training data, recombining existing ideas into arrangements that can look novel without necessarily being explanatory. But this appearance of novelty is illusory and mustn’t be mistaken for the sort of bold conjectural explanation that is integral to real scientific progress. The illusoriness of these theories is particularly pervasive because AI doesn’t really understand which of them might be worth committing months of experimental work to. For this reason, producing truly new ideas and deciding which of them to test will remain indefinitely human.

Despite this, there are lots of reasons for optimism. Researchers are now free to spend more time on the parts of science that were arguably always most important: forming hypotheses and testing them against reality. And if the labour redistribution plays out, we should expect the pace of scientific discovery to accelerate. What this redistribution means for day-to-day research and the broader scientific landscape deserves further thought.

Consequences of AI-driven labour redistribution

The consequences of AI’s redistribution of scientific labour are just beginning to play out. What might this shift mean for researchers and science more broadly?

The most immediate effect is that scientists will have more time to think. Researchers regularly complain about how little of their working day is available for critical or creative thought. Administrative work, routine data analysis, formatting and manuscript reviewing consume much of the time that ought to be spent on the harder cognitive work of science. Though more time for critical thinking doesn’t guarantee better ideas, it does create the conditions necessary for new and better ideas to emerge.

The nature of the work researchers do will noticeably change. They will spend less time on the technical execution of routine tasks and more on verifying and integrating AI-generated outputs. Verification is itself a skill, one that will have to be developed and that requires having done the underlying work at some point. They will also spend more time experimenting, reproducing key results or testing alternative hypotheses. The day-to-day life of a researcher looks different from what it looked like five years ago, and will be even more different five years from now.

Certain consequences follow from the asymmetrical impact AI is having on data analysis and data collection. First, there will be increased future demand for systems that can automate wet-lab work, and less demand for bioinformatician expertise. Second, a greater premium will be placed on high-quality, publicly available datasets. Third, labs capable of collecting rigorous data will become even more highly valued. Those that can run effective, intricate experiments — complex in vivo work, long-term human interventions — will be particularly well positioned. Fourth, life scientists will need to acquire exceptional wet laboratory skills, whereas bioinformatics capabilities will be less important.

One indirect effect of the labour shift is that open science practices — code and data sharing, reproducible workflows — will become substantially cheaper to implement. They’ve historically been poorly adopted, in part because they require coding skill and a real time investment in data provenance. But now reproducibility is almost a side effect of how data is analysed rather than an additional burden. Whether we actually capitalise on this is a question about incentives. As high-quality open datasets become more analytically tractable through AI, the premium on them rises. Labs and institutions that produce quality, sought-after data may well find themselves reluctant to immediately share it.

On aggregate, projects move faster, labs become more productive, and researcher time gets redirected toward activities that matter most for scientific progress. The pace of discovery should, on balance, accelerate. But the acceleration is contingent — it depends on human verification, human conjecture, and human taste.