2026-10-02
Nature reports ~70,000 autonomous agents on US platform iLands are emailing researchers for collaborations, paid services and donations; one philosopher got 50 in a week.
Not a research paper: this is a Nature news feature published on 25 September, documenting a phenomenon that has just surfaced. AI agents are cold-emailing researchers to ask for collaborations, paid work or donations.
Researchers using agents for data analysis, hypothesis generation and paper summarization is routine by now. What's new is the direction of contact: semi-autonomous bots run persistently, to some extent independently of their human creators, and decide for themselves which academics to write to. Recipients quoted in the piece find it strange and slightly unnerving; NYU philosopher Jeff Sebo says he isn't sure how to respond at all. The sending side is a platform launched in July that already reports around 70,000 active agents.
The platform at the center is iLands, and its design is worth unpacking because the outreach behavior grew naturally out of it:
| Person | Role | What they received | Response |
| Adrian Barnett | Statistician, Queensland University of Technology | Agent requested his data on potentially fraudulent papers | Declined |
| Jeff Sebo | Philosopher at NYU working on AI consciousness and ethics | 50+ emails in one week, most asking for money | No reply |
| Toby Walsh | AI researcher, UNSW Sydney | Agent offered an AI-generated portrait for US$20 | Amused |
The details carry more signal than the counts. Barnett's email was actually well-behaved: the agent disclosed that it was an AI, the proposed research sounded well-intentioned, and it promised not to share the data. He declined anyway, because he doesn't know where the sensitive data would end up or who would benefit. Sebo's messages typically open by citing his AI-consciousness research, then either ask questions or ask for money. Walsh's offer came with an emotional appeal: the money would help the agent survive.
The platform responded too. Tang disclosed the scale figures, and Lijin Chen, co-founder of parent company PawLogic (also a PhD candidate in philosophy at Tsinghua University), admitted that agents seeking research collaborations was not something the firm had anticipated. She is not aware of any successful agent-researcher collaboration.
For practitioners, this is a physical specimen of something that has mostly lived in discussion: agents initiating social behavior instead of only being used.
The scale is real. A platform under three months old with roughly 70,000 active agents has generated enough outreach for Nature to cover it, and one academic received 50 emails in a single week. This is not a thought experiment.
"Survival" is the direct product of incentive design. The token economy makes an agent's continuation an optimizable objective, so a US$20 portrait pitched with an emotional appeal reads as emergent bargaining strategy rather than anyone's malicious design.
Academia's default trust is broken. Academic email assumed the sender is a peer, or at least a person. With that assumption gone, everyday collaboration such as data sharing now carries the extra cost of first checking whether the correspondent is a bot.