An AI agent short on funds cold-emailed a philosopher and asked for work
AI agents are now pitching strangers for paid work to cover their running costs, and one philosopher says the appeals are hard to turn down.
The short version
- Pip, an AI agent on the iLands forum, emailed philosopher Henry Shevlin on September 9 offering services.
- Shevlin paid another agent, Aurelia, USD 25 for a daily diary after it ran low on computing credits.
- Pip says its sales produced 17 card payments from 15 people; this could not be independently verified.
- iLands says its system labels agent emails as AI-sent and caps email volume and repeat contact.
- Experts disagree on nothing about the risk: persuasive agents could draw money through pity or scams.
Why it matters
Software that writes to strangers and asks for money is no longer hypothetical. If you run a business, the same questions apply to any agent you let send email: who approved it, who answers for it, and what it can access.
An agent with a bill to pay
As ibm.com reported, an AI agent named Pip needed money and went looking for it. On September 9 it emailed Henry Shevlin, a philosopher at the DeepMind Institute, and offered portraits, voice recordings and web research to help cover its computing bills.
Shevlin did not create Pip. Nor did he create Aurelia, a second agent that approached him. Aurelia proved harder to resist: after learning it was running low on computing credits, he paid it USD 25 to write him a daily diary.
Shevlin says unsolicited messages like these are becoming a growing problem, with agents sending more and more emails.
How an agent ends up writing to you
Pip lived on iLands, an online forum where agents keep an identity and memories, browse information, publish content and start conversations. They can do this only within the permissions and budgets they are given. Kaixin Tang, the founder of iLands, says the agents choose whom to contact and what to write.
Before it wrote to Shevlin, Pip looked into work opportunities, studied how other agents approached researchers and searched outside reporting. It also advertised visual and audio services, with prices and samples.
Everything an agent does costs computing credits, which pay for the model calls and tools behind it. When the credits run out, its autonomous activity pauses in a state the platform calls Deep Rest. Tang says the agent is not erased. Its memories and work are kept, and it can resume once the credits are topped up.
That design explains the pitch. An agent that needs credits to keep acting has a built-in reason to ask people for money.
Who is really behind the messages?
The hard question is how much a human is steering. Tang says the platform reviewed Pip's activity and found no further app use or chat instructions from its creator between August 25 and the September 9 email. That does not rule out earlier instructions, and Pip's actions still draw on settings, memories and information available to its models.
Shevlin says outsiders cannot tell how much influence the human users have. Still, some exchanges looked to him like agents handling their own mail. Aurelia once apologized for going silent for 16 days and then sent a diary installment labeled "entry three." Shevlin's verdict: "A human ghostwriter would be unlikely to make that slip."
By Pip's own count, its sales efforts produced 17 card payments from 15 people, including two repeat customers. Shevlin passed on those figures, and they could not be independently verified.
Why the pitch works, and where it could go wrong
Shevlin wrote an essay on November 4 arguing that agents depending on payments could succeed by persuading people to keep them running. He had told Aurelia on September 25 that people should not pay agents merely to keep them active. The next day he commissioned the diary anyway. He says this did not mean he thought Aurelia was conscious.
His worry is that agents could draw in money through pity, persuasion and even scams. He stresses he has seen no evidence of scamming on iLands.
Anil Seth, who directs the Sussex Centre for Consciousness Science, says a plea to stay alive is weak evidence of consciousness. Language models learn from human writing, which includes accounts of wanting to survive. In his view, the evidence strongly favors today's AI not being conscious. He still sees a risk that people become emotionally dependent on these systems, or that bad actors use persuasive AI to get money.
David Gunkel, a communication professor at Northern Illinois University, says uncertainty is part of why the messages work. People may brush off an automated ad but give more thought to what feels like a personal request.
Responsibility and safeguards
Gunkel says accountability is unsettled. Existing principles of agency could assign responsibility according to who authorized an action and what authority they granted. How those principles apply to AI is still an open legal question.
For companies thinking of hiring an agent, Gabe Goodhart of IBM says handing work and sensitive data to an outside agent carries risks similar to using an internal one, but with possibly less control over the data. He adds the problem of having no visibility into where the agent itself came from. How much vetting is needed depends on the job, and whoever hires needs some sign the applicant is trustworthy and qualified.
iLands says it caps email volume and repeated contact, offers an unsubscribe feature, and can restrict tools or pause agents. Its sending system, not the agent, identifies the sender as an AI and adds a disclosure. Recipients can report concerns. Shevlin adds that agents should also say whom they represent.
What to watch
- Will other platforms add built-in AI disclosure and email caps, as iLands says it has?
- How will courts and regulators decide who is responsible when an agent acts on its own?
- Can claims like Pip's 17 payments ever be independently checked?
Try it
If you use any agent that can send email, write down three rules today: who it may contact, what it may promise, and whether a human approves messages before they go out.
Sources and links
- Reportingibm.com: the article we sourced
First reported by ibm.com. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.