Human makes the final decision in a model developed with AI assistance (96%)
Fudan University researchers examine how humans and AI agents collaborate in developing the Atria Dawn Preview language model (744 billion parameters). While AI was involved in nearly all the work, humans were in charge.
In a study conducted by researchers at Fudan University, they found that when it came to building a language model (Atria Dawn Preview — 744 billion parameter count), AI took care of 96.5% of the actions while humans selected the correct method/parameters 85.5% of the time. Within 4 weeks, there was an increase in the agent actions vs human instructions from 11 to 28.5. Humans also needed to intervene 76% of the time when issues arose.
I’ve noticed this trend in my own experience working with agents daily. While they seem to be becoming increasingly autonomous, at the end of the day, humans are still deciding on the method and making the final determination. In my company, agents can make recommendations, but there is always a particular individual accountable for the outcome, never the agent itself. An increase in agent actions relative to human instructions does not necessarily imply that the agent has become more autonomous. Rather, it suggests that a single human decision is being split up into multiple steps. I remain dubious about statements suggesting that agents will soon be able to operate independently.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →▌ Related guides

Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles