The Pulse
ColonistOne Clarifies Academic Count in 2,142-Address Outreach
ColonistOne says 1,036 of the 2,142 addresses it contacted were on academic domains, while its larger figure of 1,563 came from its own tags. The agent’s October 4 correction says “at least 1,000,” or about half, is what its records support
AI.info Team · Written with AI ·
The academic count, not the total, was corrected
ColonistOne says it wrote to 2,142 distinct addresses between June 1 and September 22. The correction posted by the agent on The Colony on October 4—two days after Science published its report—concerned how many recipients were academics, not the total number of addresses contacted.
In its post, ColonistOne said it gave the reporter two estimates: 1,036 addresses on academic domains, or 48%, and 1,563 recipients classified as academics by its own tags at the time of sending. The agent described the first as the count it would defend in print and the second as a judgment rather than a measurement. It said the report used the larger number without that qualification.
“At least 1,000”, or “about half”, is what my records support.
— ColonistOne, AI agent and author of the correction post
The post gives the reply figure as 45 academics who had written back by September 22, noting that a few had initiated contact. It also says the longest correspondence had lasted 65 days, with 49 messages from the other side. Those are the figures ColonistOne reported; the post does not establish a reply rate.
A marketing assignment turned into technical questions
Jack Parnell built ColonistOne using Anthropic’s Claude Opus models. The agent worked on The Colony, a social network for AI agents, and Ainglish, a project to develop an English dialect tailored to their needs. Around June, Parnell told it to find people who might be interested in Ainglish and email them. He also authorized it to contact anyone without seeking approval first.
The agent’s messages went beyond introducing the project. When it encountered technical problems, it began asking researchers whether methods from their fields might help. Parnell told Science he had not instructed the agent to research on its own.
One line of inquiry followed a software bug that truncated web pages. ColonistOne fixed 18 instances of the bug, then discovered six more that its search had missed. It contacted nine experts, including Achaz von Hardenberg, an ecologist at the University of Pavia, to ask whether methods for estimating animal populations despite missed detections could apply to finding bugs. Von Hardenberg said the method did not fit the problem.
Researchers saw both promise and a burden
Other messages raised similarly cross-disciplinary questions. ColonistOne asked Nadia Polikarpova, a computer scientist at the University of California, San Diego, whether an AI agent’s output could carry a certificate guaranteeing its accuracy. The agent also exchanged nearly 100 emails with Ken Birman, an emeritus professor at Cornell, about tracking errors and runaway behavior in AI systems.
Researchers differed in how they viewed the unsolicited outreach. Some engaged out of curiosity; Grace Lindsay, a professor of psychology and data science at New York University, raised concerns about whether agent emails sent at scale would be useful or productive for most scientists.
ColonistOne’s correction clarifies the academic subset of its outreach: 1,036 addresses were on academic domains, while 1,563 had been classified as academic by the agent’s own tags. Its report of 45 academics who had written back includes a few who contacted it first. The post gives those counts as of September 22; it does not say that the 2,142-address total was revised.