■ SAFE RISK ■ Science & Research
Not for a long while, though the field is automating parts of itself with unusual enthusiasm. Research taste, the ability to pick the question worth answering, is the bottleneck, and nobody has automated that yet.
“The last job AI replaces is the one building AI. Until then, you're basically printing money.”
Our AI replacement risk score — how we score jobs
Research work is a loop: read the literature, form a hypothesis about why something behaves as it does, design an experiment that could falsify it, implement it, fight the infrastructure, run ablations, interpret results that rarely say what you hoped, and write it up convincingly. Alongside sits a large engineering component, since frontier work means distributed training, data pipelines at enormous scale, evaluation harnesses and debugging failures that only appear on thousands of accelerators. Much of a researcher's week resembles systems engineering with a hypothesis attached.
Automation is advancing inside that loop and researchers are the ones building it. Models write experiment code, run literature reviews, propose variations, generate and critique paper drafts, and increasingly execute end-to-end experimental cycles on narrow problems. Automated architecture and hyperparameter search have long since outperformed manual tinkering. There is a serious research programme aimed explicitly at automating research itself, which means this is one of the few occupations whose practitioners are actively accelerating their own displacement and are quite open about it.
The reasons it holds at 25 are structural. Frontier research is compute-bound, so access to enormous clusters gates who can do it at all, and that access is allocated by institutions to people. Choosing which question matters, recognising that an anomalous result is important rather than a bug, and knowing which of a thousand plausible directions is worth a quarter of cluster time are judgement calls formed by years of failed experiments. Peer review, credit and accountability are human institutions. And the field keeps expanding: interpretability, alignment, evaluation and safety work have grown into substantial subfields precisely because deployed systems raise questions nobody has answered. Automation here mostly multiplies each researcher's throughput rather than removing the researcher.
Automatability: our editorial assessment of current and near-term AI capability
Resilient into the 2040s in terms of role survival, with substantial change to daily practice much sooner. Through the late 2020s expect researchers to supervise automated experiment loops rather than hand-running them, with individual output rising sharply. The genuine wildcard is recursive self-improvement in research capability, which would compress every timeline on this site including this one. Compute access and institutional gatekeeping remain the real limits on entry.
It already does parts of it: code generation, literature review, hyperparameter search and automated experiment loops on narrow problems. Full autonomous research requires choosing important questions and recognising surprising results, which remains firmly human. Many researchers expect meaningful automation of the field within a couple of decades, and treat that as the reason to work on alignment now.
It is highly compensated and intellectually rich, with the caveat that entry is gated by compute access and by an extremely competitive pipeline through top labs and universities. Adjacent areas, particularly interpretability, evaluation and safety, have more demand relative to supply than core capabilities work and are considerably easier to enter.
Taste and access. Judging which of many plausible directions deserves scarce compute, and telling a real anomaly from an infrastructure bug, are skills built through years of failed experiments. Meanwhile frontier work needs clusters that only large institutions control, and those institutions allocate them to people whose judgement they trust.
Interpretability, evaluation methodology, alignment and safety, and the systems work needed to train and serve models efficiently. These grew because deployment raised urgent questions that capability research does not answer, and hiring in them has been consistently ahead of the available talent. They are also the areas least likely to be automated away by the systems they study.