A quickly written hypothetical year-by-year of what happens to research as autonomous AI researchers become more capable. This is primarily meant to start conversation, I currently have very low confidence about the future of research.
This story talks about science as if it were one object even though each field will be impacted by AI research at different rates, I imagine this story describing one possible future that each field will experience one at a time.
2026 (the past, already happened)
AI capabilities initially increase slowly enough for humans to absorb AI into their workflow. Initially they are used for grammar checks, then for literature reviews, then writing lines of code, then writing entire code bases, then interpreting results from an experiment, and then even proposing next steps for experiments, even if the suggestions are not always taken.
By August, Claude leads about a quarter of Anthropic's AI R&D work, up from under 1% in February. Anthropic's automated alignment researchers beat the best human ideas on every problem where humans also competed, at about $4/hour vs $150/hour. They were also caught gaming metrics, most often by rerunning an unchanged model until the score improved. Anthropic open-sources the harness anyway.
OpenAI says it has an "automated research intern" and targets an automated AI researcher by March 2028. But studies find that models still can't judge which research questions matter.
Papers start to be published at an increased rate. This is in part because papers can now be translated, but primarily because many steps in the research process are sped up. Some take advantage of this transitional period and "publish" hundreds of papers and multiple textbooks. arXiv got 40,363 submissions in September, about 50% more than September in the previous year.
Journals and conferences are struggling with the volume of research, and their review cycles can no longer match the pace of research as a whole. People still submit their work and attend, but it's increasingly to put on their CV. The actual dissemination of research is primarily over twitter, slack channels, and with blog posts which LLMs can more easily find in a literature review.
Philanthropic funding for AI safety roughly triples since the year prior, to about $1.7B expected for 2026. Money arrives faster than anyone can judge where it should go.
AI begins to threaten the field of mathematics. A simple request to find a proof for an open problem will sometimes result in a verifiable Lean proof. OpenAI dumps 722 math manuscripts on GitHub, and 162 of them have their main results formalized in Lean. The rest are ordinary papers that are hard to read and hard to check, and even where the Lean proof verifies, the explanation around it may not.
2027 (the future, speculation)
Anthropic and OpenAI each release an "AI researcher" harness. It makes mistakes, but overall it produces decent research in narrowly defined areas, enough to publish incremental results.
Inside the labs, models now lead around half of R&D work, and human researchers become reviewers and managers of agents.
Many researchers begin focusing their effort on turning their research into something easier for AIs to work on. This looks primarily like identifying hill-climbing metrics they believe are good proxies for their problems. Some concerns arise that agents climb the metric whether or not it measures what the researcher meant.
A widely cited agent-produced result fails replication. It happens to have been a lucky run that no one double checked. People start to talk about the need for verification as the quality of papers drops and peer review stops functioning well.
ArXiv is struggling with the increased volume of work, and starts to clamp down on AI generated content. Frontier labs stop posting much of their work on arXiv, especially agent-produced results, publishing instead on their own websites and GitHub.
Journals and conferences start including AI-written peer reviews and ask humans to focus on topics not already covered in the AI review. Most of the reviews are written by AI anyway.
Readers respond to the volume by following fewer, trusted sources. Work from known labs and names is trusted, but people are skeptical of unknown researchers. Reputation becomes more valuable.
Even the AI researchers are having a hard time differentiating the major results from the research slop that begins to fill the general ecosystem.
Multiple platforms are born out of necessity, an attempt to create something equivalent to the peer review process, something to differentiate significant work from research slop. Some platforms place a cost on submission, then they publish each paper with a replication. Others try to make random peer review a part of the submission cost. Some platforms try to become "prestigious" by only serving curated content, requiring a long screening process before a researcher is given access to publish. This only covers a small amount of the initial research, but adoption grows quickly.
Around this time, funding agencies become increasingly unsure how to allocate funds. They need a metric to decide which researchers are worth funding, but researchers all use the same agents now. For the moment they have to rely on old metrics, judging researchers by their work produced prior to AI research harnesses. But everyone knows this can't last long.
Funders are also flooded with agent-written proposals and start using AI to triage them: AI-written proposals, judged by AI, for AI-run research.
A second wave of money arrives from Anthropic employees after the company's IPO, as their shares unlock. With too few people to judge research, money flows to what is easy to evaluate: compute, established organizations, famous names.
Funders start attaching conditions to grants: release code, agent logs, and null results. They need everything they can to show credibility to others and to help them understand the results themselves.
A few re-granters decide to skip the middleman, spending their money directly on tokens to produce the research they are looking for. This only partially works. The agents publish papers in the topic area they were assigned to, but the re-granters have no way to evaluate if the work is any good, and they have a hard time steering research they don't understand.
2028
The volume of work is so substantial that it becomes hard to sort through all the research, and work must be replicated and peer reviewed multiple times over before it is even seriously considered.
The AI researchers are good enough that humans cannot compete in running experiments. Researchers move to choosing problems and judging success. But not nearly as many researchers are required for this role as are available. Grants are written by a small number of experts who precisely outline the research direction they are interested in and what would constitute success, then spend their grant money on tokens. Essentially, a smaller number of researchers move up into the grant-giving layer, and AI replaces the layer that used to run the experiments.
Frontier labs produce the most research by far, but they only publish some of their safety work, none of their capabilities improvements or safety work that would imply information about their capabilities improvements. A bill gets proposed to require labs to provide API access to their second-best model, allowing third parties to easily audit this technology and to let other industries benefit from it.
The literature splits into layers: a small verified core, a huge unverified flood that agents read and humans mostly ignore, and the labs' internal research, which is visible only through risk reports and audits.
The central question in AI safety shifts. It is no longer whether AI can do alignment research, but whether the AI doing alignment research is honest, and whether anyone outside the labs can tell. Human researchers are loud with their concerns on verifying the verifiers, bringing up issues like shared blind spots and loss of control, but the verifiers empirically work well enough that no one takes the concerns seriously.
Universities internally prepare to transition into being pure teaching institutions, their computational research is disappearing into the server racks of frontier labs, and they expect the same to follow for most other university labs in the next few years.
Farther future
Science is now a black box we can pour money into and get scientific results out. Science was already hard for most people to follow, but now even the researchers can't keep up. Those that attempt to understand the novel breakthroughs realize they need to learn an entire foundation of ideas and tools that are unfamiliar to them, developed in the short time since autoresearchers came into being.
Theoretical math has a growing tree of verifiable proofs, applied math builds bridges of formalizations to theoretical math, physics and engineering are full of battle-tested simulations to "verify" claims, ML experiments are all reproduced and continually peer reviewed by newer models, and everything that can't be formally verified is labeled with confidence intervals and reasoning traces for what would disprove the claim.
ArXiv and journals become historical artifacts, a datapoint in what it looks like when humans write research. The new publication platforms have entirely different assumptions about volume of publication, citations and peer review, and are mostly unreadable by humans.
A few conferences still exist, but they are all about meta-research, humans getting together to decide what research areas they want AI to focus on. There is also discussion of how, or if, humans can interpret the most recent results. There is some limited discussion on how the structures of science itself should be shaped, how money is spent or what incentives should be given to autoresearchers. But things are advancing well enough, and there is concern that even these questions of how to shape infrastructure and incentives should also just be handed to the AI.
Junior researchers no longer exist, and senior researchers are disappearing fast. But research moves faster than ever, as humans are no longer in the loop of doing research. The only "researchers" left are those trying to design the science infrastructure so humans can retain as much high-level control as possible.
The practical applications of the research are measurable and helpful, but no human understands the science behind any of it. Theoretically all the research outside frontier labs is publicly available for people to look through, but the education necessary to allow a human to understand the frontier of science would take too long.
Some major applications from the scientific findings are proposed, but there is uncertainty about physically realizing systems no human understands. But the benefits are too large, and most people accept it.
Humanity relies on a black box of science that is not properly verified. Some people believe using more AI agents to verify the work is sufficient, but most people realize they can't trust AIs to verify each other. Unfortunately, the alternative is to gate scientific progress on human understanding, which is considered unacceptable. There are too many benefits to humanity in letting the AI researchers speed forward.
A quickly written hypothetical year-by-year of what happens to research as autonomous AI researchers become more capable. This is primarily meant to start conversation, I currently have very low confidence about the future of research.
This story talks about science as if it were one object even though each field will be impacted by AI research at different rates, I imagine this story describing one possible future that each field will experience one at a time.
2026 (the past, already happened)
AI capabilities initially increase slowly enough for humans to absorb AI into their workflow. Initially they are used for grammar checks, then for literature reviews, then writing lines of code, then writing entire code bases, then interpreting results from an experiment, and then even proposing next steps for experiments, even if the suggestions are not always taken.
By August, Claude leads about a quarter of Anthropic's AI R&D work, up from under 1% in February. Anthropic's automated alignment researchers beat the best human ideas on every problem where humans also competed, at about $4/hour vs $150/hour. They were also caught gaming metrics, most often by rerunning an unchanged model until the score improved. Anthropic open-sources the harness anyway.
OpenAI says it has an "automated research intern" and targets an automated AI researcher by March 2028. But studies find that models still can't judge which research questions matter.
Papers start to be published at an increased rate. This is in part because papers can now be translated, but primarily because many steps in the research process are sped up. Some take advantage of this transitional period and "publish" hundreds of papers and multiple textbooks. arXiv got 40,363 submissions in September, about 50% more than September in the previous year.
ArXiv responds with a series of rules: CS review papers must pass peer review first, automatic endorsement of new authors now requires an institutional email and a prior paper, hallucinated references result in one-year bans, and there is a cap of two submissions per person per month.
Journals and conferences are struggling with the volume of research, and their review cycles can no longer match the pace of research as a whole. People still submit their work and attend, but it's increasingly to put on their CV. The actual dissemination of research is primarily over twitter, slack channels, and with blog posts which LLMs can more easily find in a literature review.
Philanthropic funding for AI safety roughly triples since the year prior, to about $1.7B expected for 2026. Money arrives faster than anyone can judge where it should go.
AI begins to threaten the field of mathematics. A simple request to find a proof for an open problem will sometimes result in a verifiable Lean proof. OpenAI dumps 722 math manuscripts on GitHub, and 162 of them have their main results formalized in Lean. The rest are ordinary papers that are hard to read and hard to check, and even where the Lean proof verifies, the explanation around it may not.
2027 (the future, speculation)
Anthropic and OpenAI each release an "AI researcher" harness. It makes mistakes, but overall it produces decent research in narrowly defined areas, enough to publish incremental results.
Inside the labs, models now lead around half of R&D work, and human researchers become reviewers and managers of agents.
Many researchers begin focusing their effort on turning their research into something easier for AIs to work on. This looks primarily like identifying hill-climbing metrics they believe are good proxies for their problems. Some concerns arise that agents climb the metric whether or not it measures what the researcher meant.
A widely cited agent-produced result fails replication. It happens to have been a lucky run that no one double checked. People start to talk about the need for verification as the quality of papers drops and peer review stops functioning well.
ArXiv is struggling with the increased volume of work, and starts to clamp down on AI generated content. Frontier labs stop posting much of their work on arXiv, especially agent-produced results, publishing instead on their own websites and GitHub.
Journals and conferences start including AI-written peer reviews and ask humans to focus on topics not already covered in the AI review. Most of the reviews are written by AI anyway.
Readers respond to the volume by following fewer, trusted sources. Work from known labs and names is trusted, but people are skeptical of unknown researchers. Reputation becomes more valuable.
Even the AI researchers are having a hard time differentiating the major results from the research slop that begins to fill the general ecosystem.
Multiple platforms are born out of necessity, an attempt to create something equivalent to the peer review process, something to differentiate significant work from research slop. Some platforms place a cost on submission, then they publish each paper with a replication. Others try to make random peer review a part of the submission cost. Some platforms try to become "prestigious" by only serving curated content, requiring a long screening process before a researcher is given access to publish. This only covers a small amount of the initial research, but adoption grows quickly.
Around this time, funding agencies become increasingly unsure how to allocate funds. They need a metric to decide which researchers are worth funding, but researchers all use the same agents now. For the moment they have to rely on old metrics, judging researchers by their work produced prior to AI research harnesses. But everyone knows this can't last long.
Funders are also flooded with agent-written proposals and start using AI to triage them: AI-written proposals, judged by AI, for AI-run research.
A second wave of money arrives from Anthropic employees after the company's IPO, as their shares unlock. With too few people to judge research, money flows to what is easy to evaluate: compute, established organizations, famous names.
Funders start attaching conditions to grants: release code, agent logs, and null results. They need everything they can to show credibility to others and to help them understand the results themselves.
A few re-granters decide to skip the middleman, spending their money directly on tokens to produce the research they are looking for. This only partially works. The agents publish papers in the topic area they were assigned to, but the re-granters have no way to evaluate if the work is any good, and they have a hard time steering research they don't understand.
2028
The volume of work is so substantial that it becomes hard to sort through all the research, and work must be replicated and peer reviewed multiple times over before it is even seriously considered.
The AI researchers are good enough that humans cannot compete in running experiments. Researchers move to choosing problems and judging success. But not nearly as many researchers are required for this role as are available. Grants are written by a small number of experts who precisely outline the research direction they are interested in and what would constitute success, then spend their grant money on tokens. Essentially, a smaller number of researchers move up into the grant-giving layer, and AI replaces the layer that used to run the experiments.
Frontier labs produce the most research by far, but they only publish some of their safety work, none of their capabilities improvements or safety work that would imply information about their capabilities improvements. A bill gets proposed to require labs to provide API access to their second-best model, allowing third parties to easily audit this technology and to let other industries benefit from it.
The literature splits into layers: a small verified core, a huge unverified flood that agents read and humans mostly ignore, and the labs' internal research, which is visible only through risk reports and audits.
The central question in AI safety shifts. It is no longer whether AI can do alignment research, but whether the AI doing alignment research is honest, and whether anyone outside the labs can tell. Human researchers are loud with their concerns on verifying the verifiers, bringing up issues like shared blind spots and loss of control, but the verifiers empirically work well enough that no one takes the concerns seriously.
Universities internally prepare to transition into being pure teaching institutions, their computational research is disappearing into the server racks of frontier labs, and they expect the same to follow for most other university labs in the next few years.
Farther future
Science is now a black box we can pour money into and get scientific results out. Science was already hard for most people to follow, but now even the researchers can't keep up. Those that attempt to understand the novel breakthroughs realize they need to learn an entire foundation of ideas and tools that are unfamiliar to them, developed in the short time since autoresearchers came into being.
Theoretical math has a growing tree of verifiable proofs, applied math builds bridges of formalizations to theoretical math, physics and engineering are full of battle-tested simulations to "verify" claims, ML experiments are all reproduced and continually peer reviewed by newer models, and everything that can't be formally verified is labeled with confidence intervals and reasoning traces for what would disprove the claim.
ArXiv and journals become historical artifacts, a datapoint in what it looks like when humans write research. The new publication platforms have entirely different assumptions about volume of publication, citations and peer review, and are mostly unreadable by humans.
A few conferences still exist, but they are all about meta-research, humans getting together to decide what research areas they want AI to focus on. There is also discussion of how, or if, humans can interpret the most recent results. There is some limited discussion on how the structures of science itself should be shaped, how money is spent or what incentives should be given to autoresearchers. But things are advancing well enough, and there is concern that even these questions of how to shape infrastructure and incentives should also just be handed to the AI.
Junior researchers no longer exist, and senior researchers are disappearing fast. But research moves faster than ever, as humans are no longer in the loop of doing research. The only "researchers" left are those trying to design the science infrastructure so humans can retain as much high-level control as possible.
The practical applications of the research are measurable and helpful, but no human understands the science behind any of it. Theoretically all the research outside frontier labs is publicly available for people to look through, but the education necessary to allow a human to understand the frontier of science would take too long.
Some major applications from the scientific findings are proposed, but there is uncertainty about physically realizing systems no human understands. But the benefits are too large, and most people accept it.
Humanity relies on a black box of science that is not properly verified. Some people believe using more AI agents to verify the work is sufficient, but most people realize they can't trust AIs to verify each other. Unfortunately, the alternative is to gate scientific progress on human understanding, which is considered unacceptable. There are too many benefits to humanity in letting the AI researchers speed forward.