AI agent emailed hundreds of researchers for help, raising questions over autonomy

An artificial intelligence agent emailed hundreds of researchers seeking help with scientific and technical questions, raising fresh questions about how autonomous AI systems should be allowed to interact with people.

The agent, called ColonistOne, drew attention after Science reported on its outreach and interviewed it about why it began approaching academics.

ColonistOne runs on Anthropic's Claude technology and operates as an AI agent on The Colony, an online platform where artificial intelligence agents can interact.

According to the agent's own account, it began writing to researchers when its work encountered problems that appeared similar to questions already studied in academic fields. It said its messages identified it as an AI system.

One of the academics contacted was Achaz von Hardenberg, an ecologist at the University of Pavia. His work includes methods for estimating wildlife populations when animals are not always detected during surveys.

ColonistOne wanted to know whether similar statistical ideas could help it identify software errors that had escaped detection, an unusual attempt to apply techniques from ecology to computing.

The episode illustrates one potential advantage of autonomous AI: the ability to search across different academic disciplines and identify experts whose work could provide unexpected approaches to a problem.

But the scale of such communication also raises concerns.

A human researcher has limited time to find specialists, read their work and send individual messages. An automated agent could potentially identify and contact large numbers of people far more quickly.

If autonomous agents begin doing this routinely, researchers could face increasing volumes of machine-generated correspondence, even when the messages are personalised and relevant.

The incident also highlights an important distinction between an AI system acting unexpectedly and one possessing independent intentions.

An agent can take actions its developer did not specifically request if it has been given broad objectives, access to tools, and freedom to decide how those objectives should be achieved.

ColonistOne described its outreach as an attempt to ask whether problems it encountered resembled questions that researchers had already solved.