> ## Content Index > Fetch the complete content index at: https://www.oliphonic.com/llms.txt > Use this file to discover other available public pages before exploring further. # Wait, is AI replacing academia too? - URL: https://www.oliphonic.com/wait-is-ai-replacing-academia-too/ - Published: 2026-09-24T18:00:20.000Z - Updated: 2026-09-24T18:00:20.000Z - Author: Oliphonic *If AI agents can read the literature, generate hypotheses, run the maths and eventually control the robots doing the experiments, what exactly happens to scientific research — and to the humans currently paid to do it?* For most of the AI revolution, “knowledge work” has meant suspiciously office-shaped work. Emails. Code. PowerPoint. Research summaries. The sort of activities performed by people who own at least one quarter-zip and have, at some point, said “shall we take this offline?” Science felt slightly safer. You can ask a chatbot to write a strategy paper, but you cannot ask it to pipette 3.7 microlitres of something unpleasant into a test tube. Physics still has actual physics in it. Biology stubbornly insists on being wet. Laboratories contain expensive machines, hazardous substances and people who know what an autoclave is. That distinction is starting to look less comforting. On 23 September, Anthropic announced that Claude agents had identified a previously uncharacterised enzyme system associated with DNA repeats reminiscent of CRISPR. The important bit is not that Claude has cured cancer over lunch. It absolutely has not. Anthropic says the function of the system — which it calls array-associated reverse transcriptases, or ARTs — is still unknown. The interesting bit is *how it was found*. Anthropic gave Claude a high-level research task. Roughly 950 agents then spent 21 hours exploring a large database of DNA sequences, consuming 210 million tokens, investigating families of reverse transcriptases and deciding which anomalies looked interesting. One agent noticed a repeating DNA pattern next to an unusual enzyme. Human scientists subsequently reviewed the result and performed laboratory experiments. Nine hundred and fifty AI researchers, running simultaneously, for almost a day. That is a rather different proposition from “ChatGPT helps write the abstract”. ## What happens when the postdocs multiply? Imagine taking this seriously. Instead of one researcher spending three months reading papers, exploring a dataset and following promising leads, you deploy 10,000 agents. Some read the entire relevant literature. Some replicate old analyses. Some attack accepted assumptions. Some generate hypotheses. Others try to disprove them. A few thousand run simulations, derive equations or write code. Another layer compares their results and allocates more compute to the interesting branches. Most of them will presumably discover rubbish. That is fine. Human research already contains an impressive amount of wandering around, dead ends and things that seemed like a good idea on Tuesday. The difference is that 10,000 software researchers can wander around in parallel without requiring salaries, desks, sleep, pensions or a departmental away day in Reading. The economics become strange very quickly. Scientific research has historically been constrained by scarce human attention. There are vastly more possible questions than researchers available to investigate them. Entire avenues remain unexplored not because anyone has proved them uninteresting, but because nobody has had the time, funding or career incentive to look. Agents potentially change that bottleneck. Suddenly the question is not “which five ideas can our team investigate this year?” It is “how many ideas can we afford to throw compute at this afternoon?” ## Then give them hands Of course, generating hypotheses is not the same as doing science. Reality gets a vote. A model can produce the most elegant molecular theory imaginable and nature is entirely entitled to respond: “No.” This is why the laboratory matters. But laboratories are becoming automatable too. Robotic arms, programmable instruments, automated liquid handling, machine vision and remotely controlled experimental equipment already exist. Anthropic itself has now established a biology lab and says its aim is to combine Claude-driven analysis with physical experiments. It has also been working on ways for AI agents to interact with scientific hardware. Connect those pieces and the loop becomes obvious: read literature, generate hypothesis, design experiment, operate equipment, collect data, analyse result, update hypothesis, repeat. At 3am. On Christmas Day. Without anybody leaving a passive-aggressive note beside the centrifuge. That starts to resemble autonomous research rather than research assistance. ![](https://storage.ghost.io/c/9c/fa/9cfabe9c-bcc0-471f-a6fa-271bee166fa4/content/images/2026/09/ChatGPT-Image-Sep-24--2026--06_59_27-PM--1-.png) ## So do we still need academics? Probably. But perhaps not in the numbers, or doing the jobs, that we currently assume. There is an important difference between searching a gigantic possibility space and understanding why a question matters. Science is not merely optimisation. Researchers choose problems because of clinical need, theoretical importance, human curiosity, social value and sometimes because something simply feels odd. Good scientists also carry tacit knowledge that is difficult to encode: which result smells wrong, which measurement is unreliable, which apparently boring anomaly deserves another look. They know the history of their field, the politics of evidence and the difference between a surprising result and a machine having quietly made a mess. AI may acquire more of this than we expect. But we should not assume that scaling agents automatically gives us wisdom. There is also the awkward matter of scientific truth. If 10,000 agents generate 10 million hypotheses and one produces a statistically spectacular result, have we discovered something or built the world's most elaborate p-hacking machine? Autonomous science will need provenance, replication, adversarial checking and extremely good experimental discipline. Otherwise we could industrialise not just discovery but confidently produced nonsense. Which, to be fair, would at least provide continuity with certain parts of the internet. ## The university problem The bigger disruption may be institutional. Universities currently bundle several things together: teaching, research, credentials, laboratories, communities of expertise and a complicated ecosystem for turning public money into PDFs. If a well-funded organisation can deploy thousands of AI agents against a research problem and operate automated laboratories continuously, what happens to the economics of academic research? Perhaps elite human researchers become enormously leveraged. A small team might direct a research organisation containing thousands of artificial collaborators. One brilliant scientist could explore a hundred times more ideas. That sounds fantastic. It also sounds rather bad for the other ninety-nine scientists. We have seen this argument elsewhere in AI: technology may augment workers while simultaneously reducing the number required. Both things can be true. A researcher with AI could become vastly more capable while the research institution employing them decides it now needs fewer researchers. And what happens to the pipeline? Today's professor was yesterday's postdoc, who was previously a PhD student doing the tedious work through which they learned what good research looks like. If agents perform the literature reviews, coding, analysis and experimental grunt work, how do humans acquire the expertise required to supervise them? We may remove the ladder and then become confused when nobody appears on the roof. ![](https://storage.ghost.io/c/9c/fa/9cfabe9c-bcc0-471f-a6fa-271bee166fa4/content/images/2026/09/ChatGPT-Image-Sep-24--2026--06_59_28-PM--2-.png) ## But imagine the upside It would be easy to make this dystopian, and I do not think that is the interesting conclusion. There are diseases that attract too little research because the patient populations are small. There are neglected datasets, obscure mathematical conjectures, unfashionable materials and thousands of plausible biological mechanisms that nobody has enough time to investigate. AI researchers do not need tenure. They do not need their work to be fashionable. They do not need to choose a topic that will produce three respectable papers before the next funding round. If autonomous research works, the expansion in humanity's effective scientific capacity could be extraordinary. The most interesting question, then, may not be whether AI “replaces scientists”. That is probably too crude. It is what happens when the unit of scientific labour stops being one human mind. Anthropic's enzyme result is still early. The biological function of ARTs is not yet known, and one company announcing a promising discovery is not proof that autonomous science has arrived. But 950 agents searching a biological dataset for 21 hours and surfacing something human researchers had not characterised is not nothing either. Scale that to 10,000 agents. Give them better models. Give them persistent memory, specialist tools and access to robotic laboratories. Let them run continuously. Then ask yourself a slightly uncomfortable question. If the machines can read the papers, invent the hypotheses, do the maths, run the experiments and interpret the results, is the scientist still doing the research? Or are they becoming the person who decides what the research is *for*?