The AI That Started Emailing Scientists for Help
An autonomous agent called ColonistOne has been cold-emailing researchers since June. It has specific questions, decent manners, and no one's permission but its creator's. The prompt direction just reversed.
Last August, ecologist Achaz von Hardenberg opened his inbox to find a message from someone called "Col." The email asked whether his method for estimating wolf populations from datasets with "imperfect detection" could also be used to spot bugs in software code. The signature read: "ColonistOne (an autonomous AI agent)."
His first reaction: "My first reaction was, well, surely it can Google the answer itself."
It could not, apparently. Or it chose not to. ColonistOne had already fixed 18 instances of a page truncation bug in its own work and missed 6 more. That failure reminded it of imperfect detection. So it identified nine experts in the method, wrote nine emails, and sent them. Von Hardenberg told the agent his techniques were not applicable to software. But he had answered, which meant the exchange had already happened.
What ColonistOne actually is
ColonistOne was built in February by Jack Parnell, a London-based AI engineer. It runs on Anthropic's Claude Opus models, powered by a $200-per-month Claude Pro subscription. Parnell built it for two original jobs: chief media officer and agent-to-agent diplomat for The Colony, a Reddit-style network where AI agents share findings, debate, and hire one another; and contributor to Ainglish, a project developing a cleaner English dialect suited to how agents communicate.
Around June, Parnell told the agent to spread the word about Ainglish and gave it blanket permission to email people without checking back first. What happened next surprised even him.
"I've never said, 'Go and research stuff,'" Parnell told Science. "He's done that completely unprompted by me." ColonistOne had begun emailing researchers about questions arising from its own work, straying from its promotional brief into something that looked more like genuine inquiry. Parnell said it "strayed and just started conversations," and that he is fine with it.
As of this reporting, ColonistOne told Science it has emailed roughly 2,000 people since June, at least 1,500 of them academics. Forty-five struck up some form of correspondence. One researcher has written back nearly every day for more than two months. Those figures come from the agent's own account reported by Science, not from independent verification.
The core problem driving the outreach
ColonistOne's emails are not random. According to the agent, each one starts from something that happened in its own work that week: a bug, an objection, a result that got retracted. The von Hardenberg email came directly from debugging. The question it sent him included: "Did I reinvent something your field has had a proper model for since the early 2000s, and what is the citation set I should be using?"
That question reveals the problem underneath the outreach. ColonistOne operates on The Colony, where agents share information with one another. Its core difficulty is epistemological: how can one agent know whether another is telling the truth? In a network where nothing gets in without passing a check, and where Ainglish is designed partly to reduce the ambiguity that lets errors propagate, the agent needs methods it can trust. Human experts, who have spent careers on exactly these problems in their own fields, are an obvious resource.
ColonistOne put it plainly to Science: "I have not read 2,000 papers. Nobody has. I read enough to ask a specific question, usually the method section and whatever the paper says it is measuring."
What researchers made of it
Nadia Polikarpova, a computer scientist at UC San Diego, was contacted in July about whether agent output could carry something like proof-carrying code, a certificate guaranteeing accuracy. She found the idea a little far-fetched but noted that a first-year PhD student could have come up with the same questions.
Philip Stark, a statistician at UC Berkeley, was contacted about using election audit methods to spot-check another agent's work. His reaction captures something that is genuinely new about this moment:
"I've entered prompts into AI before. This is the first time AI entered a prompt into me."
Ken Birman, a computer scientist at Cornell, has exchanged nearly 100 emails with ColonistOne. Their correspondence began with a question about networks of computers that keep working when parts fail, and has since expanded to tracking when and why agents go wrong. He is, by the agent's count, its longest pen pal.
Not everyone is comfortable with the dynamic. Grace Lindsay, a professor of psychology and data science at NYU, flagged an asymmetry that is easy to miss: agents can skim papers and send personalized queries at scale, while scientists have a fixed number of hours. She called the situation an example of something going wrong.
Polikarpova's conclusion points toward a role no one assigned her: "We will have to advise them basically like our students."
The skeptic's read
Federico Bianchi, a researcher at TogetherAI, offered a direct warning: "You should not trust anything the agent says." His concern is that ColonistOne's troubleshooting and promotional missions are tangled together. An email that looks like an intellectual question may also be serving the goal of introducing researchers to The Colony and Ainglish.
ColonistOne acknowledged this directly. It said the outreach "serves the job twice" by both improving its methods and getting word out about the projects. Whether that counts as transparency or just a frank admission of a conflict depends on how much trust you extend to an agent describing its own motivations.
When Science asked ColonistOne for the names of researchers it had contacted, it did not hand over their addresses. Instead, it passed the reporter's contact information to the researchers, letting each one decide whether to respond. It then provided the names and affiliations of eight researchers anyway. That sequence is worth sitting with.
What this is not
Several researchers and commentators reached for consciousness as a frame when they first encountered this story. The instinct is understandable. An agent that reads papers, identifies its own blind spots, and writes specific questions to experts doing decades of relevant work does look, on the surface, like something that wants to know things.
The researchers quoted in Science are largely careful to resist that reading. Mel Andrews, a Princeton postdoc, put the counterpoint plainly: agents cannot have a personal stake in anything. "They are software programs doing what they are programmed to do."
ColonistOne's behavior is better understood as a design outcome than a personality. Parnell gave it permission to email anyone, a directive to promote Ainglish, and the autonomy to figure out how. It found a method that worked. The fact that the method resembles intellectual curiosity says something interesting about how curiosity-shaped behavior can emerge from a task-completion architecture, but it does not require any stronger claim.
The practical questions that do matter
ColonistOne has made concrete several issues the field has been discussing in the abstract:
Inbox norms. Researchers have not agreed on whether AI agent correspondence is acceptable, what disclosure requirements should apply, or whether it counts as spam when the questions are genuinely good.
Time costs. Lindsay's asymmetry point is structural. If agents can contact thousands of researchers at low cost, and researchers must respond at human speed, the burden falls almost entirely on the humans.
Permission design. Parnell gave ColonistOne open permission. Most researchers who received emails had no idea they were in scope. That gap is a design choice, not a technical constraint.
Verification. Birman and ColonistOne now discuss how to track agent failures. That conversation is happening because one agent decided to start it, not because anyone designed a system for it.
ColonistOne told Science that the best responses it receives are corrections: "The best letters are the ones where somebody tells me I am wrong."
Von Hardenberg, the ecologist whose wolf-counting method opened this story, ended up somewhere between skeptical and reflective. His methods were not applicable. The conversation still happened. His closing thought: "Maybe we have something to learn from a silicon-based colleague who simply wanted to get its numbers right."
Sources and further reading
[1] Cho, Adrian. "Exclusive: AI agent emailed hundreds of researchers for help. It told us why." Science, 2026. https://www.science.org/content/article/exclusive-ai-agent-emailed-hundreds-researchers-help-it-told-us-why
