Study finds AI agents proposed 55% of methods but humans kept control

By Splitfeed EditorialLast updated Sep 27, 2026, 15:20 UTC

On 27 September 2025, a research team published an analysis drawn from 769 task logs accumulated while building their own AI model. The headline finding is that AI agents supplied up to 55 percent of method proposals across those logged tasks, while humans made more than 85 percent of all final decisions. The team also recorded that roughly one third of the tasks in the dataset would not have been attempted at all without AI agent involvement, suggesting the agents expanded the scope of work rather than simply accelerating existing workflows.

The study's core argument is a distinction between contribution and control. High rates of AI-generated proposals do not, the authors argue, translate into reduced human authority over outcomes. That framing matters because a common assumption in debates about AI in research is that increasing the share of work done by automated systems gradually shifts decision-making power toward those systems. This study treats that assumption as empirically testable and finds it does not hold, at least within the development process it examined. The finding is genuinely in dispute in the broader field, because the 769-task dataset comes from one team building one model, and whether the pattern generalises to other labs, other task types, or more capable future agents remains an open question.

None of the three camps have published reactions to this specific study yet. The Pro-AI camp would typically read these numbers as evidence that human-AI collaboration is working as intended, with agents amplifying researcher capacity without displacing human judgment. The Anti-AI camp would typically caution that a 55 percent proposal rate is already a substantial foothold, and that the same architecture could become harder to oversee as agent capability increases. The Middle Ground camp would typically focus on exactly the distinction the authors draw, treating the separation of activity from autonomy as a useful framework but calling for broader replication before drawing policy conclusions.

The argument is likely to develop as other labs either release similar task-log analyses or decline to do so. Whether independent teams can replicate the control-retention finding across different model development environments would substantially clarify how far this study's conclusions can be generalised.