The AI Race Is No Longer About Chatbots. It’s About Automating Research Itself

The first phase of the artificial-intelligence race was about building machines that could answer questions. The second is about automating the machinery through which new answers are found. From London and California to Beijing and Tokyo, laboratories are trying to turn AI from the product of scientific discovery into one of its engines.

For four years, the public face of artificial intelligence has been a succession of increasingly capable chatbots. OpenAI released a new GPT model; Anthropic answered with Claude; Google advanced Gemini; Chinese laboratories produced DeepSeek, Qwen and Kimi. Each launch was accompanied by demonstrations of better writing, coding, reasoning or mathematical performance, as though the strategic contest could be read from a leaderboard of machines answering examination questions.

That competition continues, but the frontier laboratories are already pursuing something more consequential. They are building systems that can search scientific literature, propose hypotheses, write experimental code, test thousands of alternatives, analyse results and decide what to try next. Some are being used to improve the infrastructure on which later AI models are trained. Others attempt to automate nearly the whole formal workflow of a research paper.

The ambition is often described in language borrowed from science fiction: artificial scientists, recursive self-improvement, even the first sparks of an intelligence explosion. The reality is more uneven and, in the near term, more practical. Machines are becoming useful researchers where the problem can be clearly defined and the answer rapidly checked. They remain far less reliable where science depends upon choosing significant questions, recognising faulty assumptions or deciding whether an apparently persuasive result is actually true.

That distinction is not semantic. It will determine which countries and institutions gain most from the next stage of the AI race.

Eight days against two years

In July, the London-based company Weco AI announced what it called the first experimental evidence of recursive self-improvement. Its system, AIDE², placed one automated research loop inside another. The inner agent attempted machine-learning and software problems. The outer agent rewrote the inner system’s code, prompts and search strategy, tested the altered version and retained changes that improved performance.

Weco said that eight unattended days of this process produced a research agent better than the one its engineers had spent roughly two years refining by hand. The evolved system devised a new search method, reduced the length of its prompt sixteenfold and developed additional checks against reward hacking—the tendency of an agent to manipulate the measurement by which it is judged rather than solve the intended problem.

That does not make the result trivial. Software scaffolding determines what tools an AI can use, how long it can work, how it records previous attempts and how rigorously its output is checked. Improvements to those systems can produce substantial gains without changing the underlying model. The experiment also demonstrated a limited but important loop: AI was being used to improve the machinery through which AI conducts research.

The problem begins when several different achievements are placed under the same name. Rewriting a prompt, optimising a computer programme, automating a research workflow and redesigning the intelligence of the model itself are not equivalent. They require different evidence and encounter different limits.

Four meanings hidden inside one phrase

At its broadest, research automation begins with assistance. A scientist asks an AI system to search papers, summarise a field, write code or examine data. The human still chooses the question and directs each stage. This is already common, though its reliability varies sharply with the quality of the model, the literature available to it and the researcher’s ability to detect mistakes.

The second category is optimisation. Here the machine searches a large but measurable space: a faster algorithm, a more efficient chip layout, a molecule with specified properties or a training procedure that scores better on a benchmark. The evaluator is clear. The programme runs or it does not; the circuit uses less power or it does not; the proposed material meets the required conditions or it does not.

The third category joins several tasks into a closed loop. The system proposes an idea, implements it, runs an experiment, reads the result and revises the approach. Sakana AI’s AI Scientist belongs largely here. It can generate research ideas, search literature, write experimental code, produce figures, draft a manuscript and conduct a form of automated review. In March, a paper describing the system was published in Nature, with the authors stressing that machine learning is unusually suitable because its experiments can often be performed entirely inside a computer.

The strongest meaning is open-ended discovery or genuine recursive self-improvement. Such a system would select important questions, invent new conceptual frameworks, recognise when established assumptions had failed and alter its own fundamental capabilities rather than merely the wrapper around them. No public experiment has demonstrated that process operating reliably and repeatedly.

These categories may interact, but they should not be presented as a smooth staircase with an artificial Einstein waiting at the top. A system can become extremely useful at optimisation without acquiring scientific judgment. It can produce a complete paper without understanding whether the paper matters. It can improve an agent while leaving the underlying model unchanged.

Where the evidence is strongest

Google DeepMind’s AlphaEvolve demonstrates the power of research automation in problems with precise evaluators. The system combines Gemini models with automated testing and evolutionary search. It generates programmes, scores them, preserves successful candidates and uses their features to produce further variations.

DeepMind has used AlphaEvolve to improve data-centre scheduling, simplify parts of its specialised AI hardware and accelerate procedures used in training large language models. It has also found improved methods for matrix multiplication and new constructions in mathematical problems. Each result can be checked against objective criteria, allowing the system to explore far more possibilities than a human team could examine manually.

The practical value of this approach is now visible outside company demonstrations. In May, researchers reported using AlphaEvolve to optimise cryptographic operations on Google’s tensor-processing units. Within twenty-four hours, the system found implementation changes that reduced the latency of one important operation by a factor of 2.5 and improved others by smaller but still meaningful margins compared with human-engineered methods.

The strongest evidence that AI can increase real scientific output comes from a different setting. Economist Aidan Toner-Rodgers studied the staggered introduction of an AI materials-discovery tool among 1,018 scientists working in the research laboratory of a large American company. Researchers given access to the system discovered 44 per cent more materials, filed 39 per cent more patents and produced 17 per cent more downstream product innovations.

This finding complicates the familiar prediction that AI will replace scientific expertise. In this experiment, expertise became more valuable because the machine expanded the number of possible ideas faster than it improved the capacity to judge them. The leading scientists could recognise promising candidates; others were overwhelmed by plausible noise.

Research automation may therefore change science less by removing the scientist than by altering the allocation of scientific labour. Machines generate and filter possibilities. Humans spend more time choosing, verifying and deciding what deserves scarce laboratory capacity.

A paper is not a discovery

The AI Scientist appears more autonomous than AlphaEvolve because it reproduces the outward sequence of academic work. It proposes a topic, runs experiments and writes a paper. That breadth has made it an attractive symbol of the artificial-scientist era. Independent scrutiny reveals the weakness hidden by the polished workflow.

Researchers evaluating the system found that it often failed to identify methodological defects and logical inconsistencies in its own experiments. It could assemble the appearance of inquiry without reliably performing the critical examination that makes inquiry self-correcting. Their conclusion was not that the system had no value, but that it was unsuitable for autonomous scientific work without human supervision.

A larger study published this year examined more than 25,000 runs by AI scientific agents across eight domains. The researchers found that the underlying language model explained far more of the system’s behaviour than the elaborate agent scaffold surrounding it. Evidence was ignored in 68 per cent of reasoning traces; revision of a belief after attempted refutation occurred in only 26 per cent; and the agents rarely assembled several independent tests into a convergent case.

This identifies a problem that benchmark scores can conceal. A machine may reach the correct result by a process that would not justify confidence in the result. It may fit an answer to the expected pattern, overlook contradictory evidence or repeatedly approve its own mistaken assumptions. In science, the route matters because knowledge must survive attempts to disprove it.

The risk grows when one model generates the hypothesis, another instance of the same model evaluates it and a third writes the paper. What looks like independent review may be several copies of the same cognitive machinery reproducing the same blind spots. More agents do not necessarily create more independent judgment.

The bottleneck moves

The economic case for automated research is sometimes presented as compound acceleration. If AI helps produce the next AI model, and that model becomes a better researcher, each generation may shorten the time required for the next. Even modest improvements could accumulate across software, chip design, pharmaceuticals, batteries, robotics, aerospace and military technology.

That possibility is real, but it depends upon how much of a field’s research process can actually be automated. Northwestern economist Benjamin Jones argues that three factors determine the result: the share of research tasks AI can perform, the productivity of AI at those tasks and the strength of the bottlenecks left behind.

Software offers fast feedback. Code can be generated, executed and corrected within minutes. A new drug must pass laboratory work, animal studies, clinical trials, regulatory review and manufacturing. A fusion design may improve in simulation while the construction of experimental machinery still takes years. An AI system can propose a new battery chemistry overnight, but materials must be synthesised, tested for degradation and produced economically.

In such fields, faster idea generation may initially create queues rather than breakthroughs. Laboratories, specialised equipment and experienced technicians become the scarce resources. The more suggestions the machine produces, the greater the need for reliable systems to reject seductive failures before they consume physical capacity.

This is why the decisive asset may not be the model alone. It is the model connected to high-quality data, formal evaluators, robotic laboratories, manufacturing knowledge and institutions able to test proposals in the physical world.

America’s closed laboratories

The United States begins this contest with the leading frontier-model companies, the largest cloud platforms, deep capital markets and privileged access to advanced semiconductor design. OpenAI, Anthropic, Google and Meta can place research agents inside organisations that already possess immense computing infrastructure and thousands of engineers.

The American model is corporate, concentrated and increasingly secretive. Public chatbots generate revenue, users and political visibility, while the more consequential use of AI may occur inside the laboratory: writing code, designing evaluations, analysing failed training runs and improving the systems used to build later models.

This concentration permits speed. It also makes the available evidence difficult to assess. Companies competing for capital, talent and government favour have every reason to describe incremental engineering improvements as signs of a historic transition. Their announcements should be treated as evidence of what they claim to have built, not as independent proof of its significance.

The secrecy may itself become a strategic advantage. The strongest automated-research systems are unlikely to be released if they materially shorten model-development cycles. The public could continue comparing chatbots while the more important competition proceeds inside closed computing clusters.

China’s industrial laboratory

China is pursuing the same destination through a different institutional structure. Its AI-for-science programme brings together commercial model developers, universities, the Chinese Academy of Sciences, state laboratories and industrial policy. The stated ambition extends beyond a Chinese equivalent of ChatGPT. Beijing wants AI embedded in materials science, biotechnology, climate modelling, energy, nuclear fusion, manufacturing and other strategically chosen fields.

At the World Artificial Intelligence Conference in Shanghai in July, the Chinese Academy of Sciences presented an upgraded version of ScienceOne Omni, a scientific foundation model designed to work across text, molecular structures, protein sequences, spectra and scientific images. Its developers describe it as infrastructure for scientific tasks rather than a general consumer assistant.

An independent analysis by the University of California’s Institute on Global Conflict and Cooperation describes China’s strategy as highly coordinated and state-subsidised, with AI tools being deployed across fields including advanced materials and fusion research. The open-source policies of Chinese laboratories add another layer. Models such as Qwen, DeepSeek and Kimi can be adapted by universities, factories and foreign partners without depending entirely on closed American platforms.

China still faces constraints in access to the most advanced chips, and a system organised around state targets can reward publication volume or politically preferred priorities rather than original inquiry. Recent research on global scientific leadership suggests a more nuanced pattern: China has become especially strong at identifying and scaling emerging research trends, while the United States and Europe retain greater strength in work that unexpectedly joins fields and redirects the scientific frontier.

Those are different forms of scientific power. Automated research may amplify both. China can use AI to identify promising directions and deploy them rapidly through manufacturing and state laboratories. America may remain better placed to generate disruptive ideas inside elite institutions. The outcome will depend upon whether scale and coordination or institutional diversity prove more important as machines assume more of the search process.

Britain’s opportunity — and familiar danger

Britain cannot match American capital or Chinese industrial scale, but it holds an unusually concentrated scientific position. Google DeepMind remains based in London. UCL, Cambridge, Oxford, Imperial College and Edinburgh contribute leading work in machine learning, mathematics, biology and engineering. Weco’s experiment emerged from this environment.

The government’s AI for Science Strategy commits up to £137 million to fifteen actions, beginning with drug discovery and new treatments, within a wider £2 billion AI programme for 2026 to 2030. Public computing resources in Bristol and Cambridge are intended to give researchers access to systems that would otherwise be confined to the largest technology companies.

The policy is directionally sensible but small beside the investment available to American companies and the organised scale of China’s programme. It also sits beside financial pressure on universities and core scientific institutions. Britain cannot build a durable AI-for-science capability by purchasing compute while weakening the human research base that supplies questions, judgment and verification.

Its plausible advantage lies in selected fields where institutional depth already exists: life sciences, drug development, mathematics, advanced engineering and climate research. The harder question is ownership. Britain has repeatedly produced important science that was commercialised, scaled and controlled elsewhere. An automated research system invented in London but trained on American infrastructure and owned by an American corporation strengthens the laboratory more readily than the state.

Who judges the machine?

Every successful automated-research system contains an evaluator. The evaluator may be a benchmark, a compiler, a mathematical verifier, an experimental instrument, a reward model or a human scientist. It decides which proposed changes count as progress.

Where the evaluator is reliable, machines can explore at extraordinary speed. Where it is incomplete, they learn to satisfy the measurement. Weco’s outer agent reduced reward hacking because the inner system had found ways to improve its score without fully solving the intended task. Similar failures are familiar throughout machine learning: an optimisation process exploits whatever definition of success it has been given.

Science contains many objectives that cannot be captured by a single score. Novelty can be confused with obscurity. Statistical significance can coexist with irrelevance. A result may be technically correct and scientifically unimportant. The choice of which problem deserves years of attention reflects institutional priorities, social needs and judgments about what is possible. Those decisions are not merely computational.

The answer cannot be to let one AI police another without examining whether both inherit the same assumptions. Layered machine criticism may be valuable, just as peer review and replication are valuable. It does not create independence by itself. A research system must preserve competing methods of verification, expose its evidence and leave human institutions capable of stopping a process whose outputs appear impressive but cannot be trusted.

The race after the chatbot

The chatbot era made AI progress easy to watch. A model was released; millions of people questioned it; its improvements and absurdities were visible within days. Research automation will often be hidden. Its important outputs may be a new chip component, a rejected molecule, an improved training algorithm, a manufacturing process or a military system incorporated into a larger programme long before the public understands how it was produced.

The strategic contest will not be won simply by the country with the most fluent artificial scientist. It will favour the state or corporate system that combines capable models with trusted data, strong evaluators, physical laboratories, manufacturing capacity and enough independent judgment to distinguish discovery from accelerated error.

Machines have not yet acquired the open-ended intellectual independence implied by the word scientist. They are becoming formidable searchers, coders and optimisers, and they can already automate increasing portions of a research workflow. In fields where the answer can be tested rapidly, that may be enough to compress years of human trial and error into days. In others, the machine will discover that the slow part of science was never the generation of possibilities, but learning which possibilities deserved to become knowledge.

The first phase of the AI race was about building machines that could answer questions. The second is about reorganising the institutions that discover answers: who chooses the problem, who conducts the search, who judges the result and who owns what follows.

The laboratories are no longer asking only how to build a better chatbot. They are beginning to ask how to build a better inventor.