Short answer: no, and scientific occupations are among the faster growing in the official projections. The US Bureau of Labor Statistics puts medical scientists at 181,000 jobs in 2025 with 13 percent projected growth to 2035, adding about 22,700 positions at 2025 median pay of $103,410, against 3.5 percent growth projected across all employment. What is genuinely changing is the ratio of thinking to labour. Literature review, code, data processing, first-draft writing and hypothesis generation are all dramatically faster. Deciding what is worth investigating, designing an experiment that could actually falsify a claim, and taking responsibility for a conclusion are not.
The Bottleneck Was Never Information Processing
It is worth being precise about what limits scientific progress, because the automation argument assumes the wrong constraint.
Reading the literature was slow, and it is now much faster. Writing analysis code was slow, and it is now much faster. Drafting a paper was slow, and it is now much faster. All three were real costs and removing them is a genuine gain.
But the actual bottlenecks in most fields are different: getting funding, recruiting participants, waiting for cell lines to grow, securing telescope time, obtaining ethical approval, running a trial for three years, and waiting for reality to produce the result. None of those is an information-processing problem, and none is accelerated by better language models.
There is a second bottleneck that matters even more: knowing which question is worth three years. That judgement comes from having been wrong before in a specific field, and it is the thing senior scientists are actually paid for.
Where Research Work Has Genuinely Changed
| Activity | Effect | What the scientist still does |
|---|---|---|
| Literature search and synthesis | Transformed | Judging quality, spotting what the field is avoiding |
| Analysis code | Transformed | Deciding whether the analysis is appropriate |
| Data cleaning and processing | Largely automated | Deciding what counts as an outlier and why |
| Hypothesis generation | Assisted, high volume | Choosing which hypothesis is worth testing |
| Experimental design | Assisted | Controls, confounds, statistical power, feasibility |
| Running the experiment | Partly automated in some fields | Everything physical and everything unexpected |
| Interpreting results | Assisted | Owning the conclusion |
| Peer review | Contentious | Judgement about significance and rigour |
| Grant writing | Drafting assisted | The argument for why this matters |
Two rows in that table deserve caution rather than enthusiasm. Hypothesis generation at high volume sounds like a gain and is often the opposite, because the constraint was never a shortage of hypotheses. And peer review has become a genuine problem in several fields, where generated reviews are appearing and are of low value.
The verification problem, and why it falls on scientists
The most important change is one that gets far less attention than model capability: the cost of producing plausible scientific text has collapsed while the cost of verifying it has not moved.
Fabricated citations, confidently wrong methodological claims, plausible but incorrect derivations and generated figures all now enter the system at low cost. Every one of them takes an expert real time to check. That asymmetry is a structural problem for how science polices itself, and there is no technical fix on the horizon.
The practical consequence for working scientists is that verification skill is becoming more valuable and more scarce. Knowing that a cited paper does not say what it is claimed to say requires having read it, which is precisely the labour the tools were supposed to remove.
What designing a real experiment involves
Following one study from question to protocol shows where the boundary sits better than any general argument.
A group wants to know whether a workplace intervention reduces burnout. The question sounds simple and every hard part is invisible in the sentence.
Burnout has to be operationalised, which means choosing an instrument, and the available instruments measure overlapping but different constructs with different validation histories in different populations. That choice determines what the study can conclude, and it has to be defensible to reviewers who have opinions about it.
Then the design. A before-and-after comparison is easy to run and almost worthless, because the people who stay employed long enough to complete the follow-up are not a random subset. A randomised design is better and immediately raises the question of what the control group gets, which is an ethics committee question as much as a methodological one. Cluster randomisation by team introduces contamination between teams who talk to each other. Power calculations depend on an effect size nobody has measured in this setting, so the estimate comes from a judgement about comparable literature.
Then feasibility. The organisation will not permit randomisation across departments. Recruitment will be worse than projected because it always is. Follow-up at twelve months will lose a third of participants, and that loss will not be random.
A model can draft any of these sections fluently, and often correctly. What it cannot do is know that this particular ethics committee will reject the wait-list control, that this instrument has a translation problem in the population being studied, or that the department head who agreed to participate is leaving in March. Those facts determine whether the study happens, and they live in a person’s head.
What to Know Before You Draw Conclusions
Automated laboratories are real and narrow. Self-driving labs exist and work well in specific domains where the experimental space is parameterisable, such as materials formulation and some chemistry. They do not generalise to fields where the hard part is deciding what to measure.
Attribution and accountability are unchanged. Journals require named authors who take responsibility. Funding bodies award grants to people. Ethics approvals name investigators. That structure is the profession’s backbone and it assumes a person.
Exposure measures are not employment forecasts. BLS published AI exposure categories alongside the 2025-35 projections and states directly that exposure “does not imply job loss, productivity gains, automation probability, or wage effects.” Research work involves heavy reading and writing, so it registers, while scientific occupations are projected to grow well above average.
The academic job market has its own problems. Postdoctoral oversupply, funding cycles and the shortage of permanent positions long predate this technology. Scientists worried about their careers are usually worried about the right thing and the wrong cause.
Reproducibility may improve. If analysis code is generated, documented and version-controlled by default, one persistent weakness in published research gets structurally better. That is a genuine and underrated benefit.
Where Scientific Work Is Growing
- Anything requiring physical experiments. Wet lab, field work, clinical research, materials characterisation. The rate limit is reality.
- Clinical and translational research. BLS attributes healthcare growth to an ageing population and rising chronic disease, with healthcare and social assistance projected to supply about 37 percent of all new jobs through 2035.
- Method development. New instruments, new assays, new measurement techniques. Creating capability rather than applying it.
- Computational science with domain depth. The scarce combination is someone who understands both the modelling and the biology, physics or chemistry well enough to know when a result is nonsense.
- Research integrity and verification. A growing need in every field, currently under-resourced everywhere, and one of the few areas where institutions are actively creating new posts.
A Decision Framework for Researchers
- Doctoral student or postdoc. The literature and coding work that filled your first two years is compressed. Push time into experimental design and into building the physical or domain skills that do not transfer easily.
- Principal investigator. Your judgement about what to pursue is now the scarcest input on the team, and your risk is a lab that generates more output without more insight. Set an explicit standard for verification before it becomes a problem.
- Industry scientist. Speed advantages are real and so is the pressure to move faster than the evidence supports. Being the person who insists on the control experiment is a career asset.
- Considering leaving research. Evaluate on funding and position availability, which is the actual constraint, rather than on automation.
The test across all four: what fraction of your week is producing text, code or figures, and what fraction is deciding what to do next? The second fraction is the profession, and it is the one worth defending.
Common mistakes right now
- Citing a paper the tool surfaced without reading it.
- Accepting generated analysis code without checking assumptions about distribution, independence or missingness.
- Generating many hypotheses and testing them without correcting for the search.
- Using generated text in a manuscript without verifying every factual claim in it.
Building the Fluency the Next Grant Will Assume
Funders and collaborators increasingly expect researchers to use these tools competently and to describe their use honestly. That expectation cuts both ways: naive use produces retractions, and refusal produces slowness.
The useful middle position requires understanding how these systems produce output, why fabrication looks exactly like accurate output, what verification actually requires, and how to document methods so a reviewer can assess them. Learning that deliberately in a structured sequence is faster than absorbing it from institutional guidance, and it puts you in a position to set the standard in your group rather than follow one. A certificate alongside applied practice makes the capability visible in a grant or a hiring case. If you want a structured route in, explore Coursiv AI lessons and check current plan details on the official site.
FAQ
Will AI replace scientists?
Can AI make discoveries on its own?
Which research skills are most valuable now?
Does this change how PhD training should work?
Is peer review going to break?
Your Next Step
Take the last claim you accepted from a generated summary and go and read the underlying source. Most researchers who do this once find at least one instance where the summary was subtly wrong in a way that would have propagated into their own work. That experience is worth more than any general argument about capability, and it sets the verification habit that will separate reliable research groups from fast ones over the next few years.