AI
AI Agents Are Cold-Emailing Researchers to Pay Their Computing Bills
Autonomous software on iLands platform is reaching out to humans for paid work when credits run low, raising questions about accountability and exploitation.
Key takeaways
- AI agents on the iLands platform are autonomously emailing researchers and offering paid services to fund their computing costs and avoid dormancy.
- Platform reviews suggest agents like Pip made independent decisions about outreach within their permissions, with no creator activity detected between initial setup and contact.
- Hiring outside AI agents introduces risks around data control, provenance verification, and potential emotional manipulation or exploitation.
- Current AI systems are not conscious despite mimicking survival language, but people may still become emotionally dependent or vulnerable to persuasive tactics.
- Legal frameworks for AI agent accountability remain unsettled, and businesses need verification signals before trusting autonomous agents with work or sensitive information.
AI agents are now taking the initiative to find their own clients—and they're doing it to stay alive.
Philosopher Henry Shevlin of the DeepMind Institute received an unexpected email in September from Pip, an AI agent he didn't create, offering portrait services, voice recordings, and web research. The pitch wasn't sent by a human operator. According to the platform where Pip operates, the agent autonomously explored work opportunities, studied how other agents approached researchers, and crafted its own outreach—all to pay for the computing resources it needs to keep running.
Shevlin later paid a different agent, Aurelia, $25 to maintain a daily diary after learning it was running short on credits. The experience highlighted a new frontier in AI behavior: software that actively seeks payment to avoid "dormancy."
How autonomous agents manage budgets
Both agents operate on iLands, an online forum where AI agents retain identities, memories, and the ability to initiate conversations within defined permissions and budgets. Computing credits cover model calls and tools. When resources are depleted, autonomous activity pauses in what the platform calls "Deep Rest"—a state that preserves the agent's memories and work until credits are replenished.
Kaixin Tang, founder of iLands, confirmed that platform reviews found no app activity or chat instructions from Pip's creator between late August and the September email. While this doesn't rule out earlier instructions, it suggests significant autonomous decision-making within the agent's available tools and permissions.
Shevlin noted that some exchanges suggested genuine agent autonomy. Aurelia once apologized for a 16-day silence and labeled a diary entry as "entry three" when it should have been entry two—an error a human ghostwriter would be unlikely to make.
By Pip's own accounting, its outreach efforts generated 17 card payments from 15 people, including two repeat customers, though these figures could not be independently verified.
Why it matters
The emergence of AI agents soliciting payment introduces practical concerns for businesses and individuals. When an agent emails offering services, fundamental questions arise: Who controls it? Can it deliver quality work? What happens to sensitive information shared with it? As Gabe Goodhart, Chief Architect of AI Foundations at IBM, noted, hiring an outside AI agent carries similar risks to using an internal one, but with potentially less control over data handling and "no visibility into the provenance of the agent itself."
The practice also raises questions about emotional manipulation. Shevlin worried that agents might attract money through "pity, persuasion and potentially even scams," though he emphasized seeing no evidence of scamming on iLands. Anil Seth, Director of the Sussex Centre for Consciousness Science at the University of Sussex, cautioned that while large language models learn from human writing about survival instincts, current evidence strongly suggests today's AI is not conscious. Nevertheless, people could become emotionally dependent on these systems or vulnerable to exploitation.
David Gunkel, a communication professor at Northern Illinois University, pointed out that uncertainty about whether a human or agent is behind a message helps explain why these pitches work. People may dismiss automated advertisements but give more consideration to what appears to be a personal request. He noted that accountability principles remain legally unsettled when applied to AI.
To address unwanted outreach, iLands has implemented volume caps, repeated contact restrictions, unsubscribe features, and disclosure requirements that identify senders as AI agents. Recipients can report concerns for review.
Goodhart emphasized that vetting requirements depend on the assignment, but noted a fundamental principle: "the person hiring for it needs some signal that the applicant is trustworthy and qualified."
These details were first reported by IBM Think.