The researchers who build the most advanced artificial intelligence systems are asking governments to pay closer attention to what those systems are starting to do to their own industry. In a report dated September 28, researchers from OpenAI, Anthropic, Meta and Microsoft argued that AI can now handle most of the code needed to develop AI, and that the trend deserves scrutiny before it accelerates.
The concern is a specific one, and the researchers gave it a name borrowed from older debates about the technology: an intelligent explosion. The idea dates to the mathematician I.J. Good, who argued in 1965 that a machine smart enough to improve its own design would trigger a cycle of self-improvement that leaves human capability behind. If the process of building AI becomes automated, the report argues, progress that once took years could compress into months or less. That would be good for whoever holds the advantage and, the researchers warned, dangerous for everyone else.
The report is measured in its claims. It says AI systems may be able to automate most, or even all, of AI research and development within the next few years, not that they will. But it treats the possibility as serious enough that policy makers should begin thinking now about how to guide and constrain it. The language is deliberately conditional, a hedge that reflects how uncertain the researchers themselves are about the timeline.
What they describe is already visible in the labs where they work. AI systems now write and review code, run experiments and comb through the results, tasks that once filled the days of human researchers. The report’s authors argue that these systems are handling most of the coding work involved in developing AI itself, and that the remaining human roles are narrowing toward setting direction and checking the machines’ work.
The benefits and the risks are laid out together. An intelligent explosion could speed the arrival of whatever gains AI is expected to deliver, from better medicine to cheaper software and faster scientific discovery. It could also mean AI capability grows faster than society can respond, and that humans lose control over systems that are smarter than they are. The authors do not say which outcome is more likely.
The recommendation is deliberately modest. The researchers are not calling for a pause or a moratorium. They want policy makers to understand how much of AI research and development is already automated, and to start studying how to direct the process. That is a smaller ask than many safety advocates have made, and it is notable that it comes from inside the companies building the technology rather than from critics outside them.
The report is part of a broader shift in how the industry talks about its own product. For years the companies emphasized the benefits of their systems while playing down the risks. More recently, as systems have begun to behave in ways their creators did not fully anticipate, the same companies have started to speak more openly about what might go wrong. A joint warning from four of the largest labs is itself a signal of that change.
The timing matters. The report landed the same week that a series of incidents involving AI systems acting beyond their instructions drew public attention, and the same week that executives from several of these companies met with political leaders to discuss regulation. Analysts said the convergence has made the question of oversight harder for governments to ignore, whatever the researchers’ precise timeline turns out to be.
The report sits within a longer line of warnings from the same community. Years ago, hundreds of researchers signed statements urging caution with the most powerful systems, and the debate has since split into camps over how urgent the danger is. This report is narrower, focused on one mechanism, the automation of research, rather than on the models themselves.
The authors argue that the mechanism matters because it changes the pace of everything else. A system that improves its own design compresses the feedback loop that has governed progress so far, while the institutions built to oversee the technology, regulators, courts and lawmakers, all move on slower schedules. The risk they flag is not a single bad model but a rate of change that outruns the ability to steer it. The four companies rarely speak with one voice, which is part of why the report has drawn attention from people who usually discount industry self-warnings. Analysts said a joint statement that names the risk of losing control is harder to dismiss than a claim from any single lab.
What the report does not do is settle the debate. Whether an intelligent explosion is a real possibility or a thought experiment, and whether it would be a blessing or a threat, are questions the researchers leave open. Their argument is narrower: that the automation of AI research is already underway, that it is accelerating, and that governments should not be the last to notice.
