Hinton, Bengio and 20-plus researchers warn self-improving AI could compress years of R&D into months

More than 20 leading artificial intelligence researchers, among them Turing Award winners Geoffrey Hinton and Yoshua Bengio and OpenAI research lead Jakub Pachocki, have signed a warning that automated AI research presents extreme risks. The group argues that AI systems may soon be able to run the entire research cycle on their own, turning what normally takes years into a matter of months. The statement lands as labs race to build agents that can design, train and improve models with minimal human input.

The core concern is speed. If a model can propose an architecture, run the experiments, read the results and iterate faster than any human team, the pace of improvement stops being bounded by hiring, lab time or conference cycles. The signatories describe that dynamic as an impending intelligence explosion, a feedback loop where each generation of systems helps build a more capable successor. Once that loop closes, they say, oversight mechanisms built for human-paced development may no longer keep up.

The names on the document give it weight. Hinton and Bengio are two of the three researchers who shared the 2018 Turing Award for deep learning, and both have spent recent years speaking about AI risk. Pachocki's involvement matters differently: he works inside OpenAI, one of the labs pushing hardest toward capable agents, so his signature signals that the warning is not coming only from outside critics. The mix of academics and active lab staff makes it harder for the industry to dismiss the message as detached from how the work is actually done.

Automating research is already an explicit goal at several major labs, which gives the warning a near-term edge. Companies have described building AI that can conduct machine learning experiments and improve its own training pipelines, and agentic coding tools are being folded into model development today. The signatories are not calling for a halt to that work; they are flagging that the transition from AI-assisted research to AI-run research could arrive faster than governance, evaluation or safety testing can adapt. That gap, rather than any single doomsday scenario, is what the group wants placed on the table.

What happens next is largely a question of whether the warning changes behavior. Signatures do not bind labs, and no enforcement mechanism accompanies the statement, so the practical test is whether frontier developers publish internal safeguards for self-improving systems, slow specific projects, or let the race override caution. Regulators, meanwhile, have shown little appetite for rules aimed at recursive self-improvement specifically. The open question is simple: if the people closest to the technology say the timeline is collapsing, who is positioned to act on it before the loop closes?