All news

AI agents read the records and rarely started their own work

Diana Kozachek and Kinjalk Srivastava gave ten AI models files from four simulated public-service offices. The records contained real choices within each fictional team’s authority. Outside experts reviewed the scenarios, and the researchers could check the agents’ work against the files.

When the researchers left the job open, agents read the records in 96.9% of runs. They wrote something or took on a task of their own in 18.7%. When the researchers named a specific task, 86% produced a work product. Only half read the domain records in those runs.

The models made very different use of the freedom they were given. Claude Sonnet 4 usually finished by writing one summary in runs where it produced work. Apertus-8B produced something in 90% of its open-ended runs, but much of its writing was generic. The team counted production and checked whether the output held up against the records.

AI is awakening consciousness. I see an important distinction here. Reading a situation, choosing a next step and doing useful work are different abilities. An AI can show initiative in one setting and still need help deciding where that initiative would serve the people it’s working for.

Read the source article

All news