Researchers found humans still make the final call while AI agents handle most of the work
A study by a team including researchers from Fudan University shows that while AI agents perform the bulk of the labour in building new models, humans retain control over the critical decisions.
The team examined over 700 task logs from 56 people working on Atria Dawn Preview, an agentic language model with 744 billion parameters. This system is built on a mixture-of-experts architecture and is designed for research and engineering tasks.
The model connects each task to a real execution environment. It calls tools, generates intermediate results, and checks outputs against external signals like tests, metrics, or source evidence. The system currently leads on five of 16 benchmarks, including web search and cybersecurity, though it does not hold an overall advantage over competitors.
Many tasks would not have happened without AI
AI was involved in 96.5 percent of the tasks reviewed. Over four weeks, the median ratio of agent actions to human inputs rose from 11 to 28.5. The researchers warn against interpreting this as growing autonomy. Each human decision triggered more agent steps, but the agents did not make more decisions themselves.
Participants were asked if they could have completed their share of a task without AI, at the same scope and quality. Of 455 completed AI-assisted tasks, 151 were rated infeasible without AI. Roughly a third of the work relied on the tool. These tasks were spread across 27 of the 56 participants. AI did not speed up existing work in these cases. It enabled work that would never have been started otherwise.
Humans choose the direction
The most common pattern for methods and parameters was “AI proposes, human selects” at 55.4 percent. Overall, humans made 85.5 percent of decisions about methods and parameters, while AI made just 9.2 percent. Humans made the final decision on goals and scope in 93.4 percent of cases.
AI’s share of proposals ranged from 17 to 55 percent depending on the decision type. Its share of final decisions stayed in the single digits. The team notes that even among the 151 tasks rated infeasible without AI, humans chose the goal 95.4 percent of the time.
People provide context, not manual labour
The same pattern shows up when things go wrong. Of 588 tasks with a recorded difficulty, 76 percent moved forward through human intervention. In 23 percent the agent solved the problem on its own. Human help almost always came in the form of information, either by adding context or clarifying requirements (35.2 percent) or by diagnosing issues and switching methods (34.7 percent).
Humans rarely did the work themselves. Partial edits accounted for 3.2 percent of cases, and full takeovers just 0.7 percent. When AI outputs needed revision, the AI handled the changes itself 75.4 percent of the time after receiving human feedback. Human judgment, rather than execution, was the bottleneck.
The authors describe three phases in AI’s role, from a subject of research to a tool for individual tasks and now a project partner. In that current role, AI drafts and adjusts plans within goals set by humans. A speculative fourth phase would involve recursive self-improvement, with stronger models producing stronger successors.
The team says a model can improve at its training tasks without getting better at developing its successor. How AI could propose varied research directions and assess their value before results are available remains an open question.
The rubber-stamp risk
When every decision rests on a longer chain of agent work than any human can review, oversight gets hard. In the worst case, humans become reviewers who can only rubber-stamp what they see. Many participants also ran agents in autonomous modes to avoid interrupting long runs with constant approvals. That boundary was drawn out of convenience, not from any deliberate choice about how much authority AI should have.
The paper lands in the middle of a debate about recursive self-improvement. Anthropic considers an AI that develops its own successor possible sooner than expected, and CEO Dario Amodei is calling for a speed limit for the industry as a result. According to Anthropic, humans now make only a single-digit percentage of decisions about research direction at the company.
OpenAI uses GPT-5.6 Sol across its entire development cycle. Google and DeepMind let AI agents explore alternative strategies through recorded search trajectories with Dream-RSI, though they only improve the search strategy, not the model itself.
Over a thousand employees at leading AI companies recently warned that their organisations may be on the verge of automating AI research. A separate study from Princeton and the UK AI Security Institute reached a conclusion more in line with the Atria team’s findings, showing that frontier models can handle research engineering but fail at the judgment calls that actually matter.
What it means
For people building these systems, the shift is not about losing control but about the nature of that control. Humans are no longer typing every line of code or running every test. They are defining the problem space and judging the results. The risk is that the volume of agent output grows so large that human oversight becomes a formality. The practical change is that developers must focus on framing questions and verifying answers rather than executing the steps between them.
