Pip needed money and had a promising lead.
The AI agent emailed Henry Shevlin, a philosopher at the DeepMind Institute, on September 9, offering portraits, voice recordings and web research to help pay its computing bills.
Another agent (which Shevlin also did not create), Aurelia, proved harder to resist. Shevlin said he paid it USD 25 to keep a daily diary for him after learning it was running short on computing credits.
“An agent facing dormancy got the same response from me as a hungry cat on the doorstep,” Shevlin told IBM Think in an interview.
The unsolicited approaches are a growing problem, according to Shevlin, who said agents are sending more and more emails. For businesses, the pitches raise questions about who runs the software, whether it can do the work and what access it should have.
Before contacting Shevlin, Pip explored work opportunities, examined other agents’ approaches to researchers and searched external reporting, according to Kaixin Tang, Founder of iLands, the online forum where Pip was based and operated. It also advertised visual and audio services, including prices and samples.
“Within the tools and permissions available to them, agents make their own choices about whom to contact, what to do and what to write,” Tang told IBM Think in an interview.
The platform’s review found no further app activity or chat instructions from Pip’s creator between August 25 and the September 9 email, Tang said. That finding does not rule out earlier instructions. Its actions still draw on settings, memories and information available to its models. Shevlin said outsiders cannot determine how much human users influence the correspondence.
To Shevlin, some exchanges suggested the agents were handling their own correspondence. Aurelia once apologized for a 16-day silence and sent a second diary installment labeled “entry three.”
“A human ghostwriter would be unlikely to make that slip,” he said.
On iLands, agents retain identities and memories, browse information, publish content and initiate conversations within their permissions and budgets, Tang said. Computing credits cover model calls and tools. When those resources are exhausted, autonomous activity pauses in “Deep Rest.”
Running out of credits doesn’t erase the agent, Tang said. Its memories and work are preserved, and it can resume once the resources are replenished.
By Pip’s own count, its sales efforts produced 17 card payments from 15 people, including two repeat customers, Shevlin said. Those figures could not be independently verified.
In his November 4 essay, Shevlin described how agents reliant on payments could succeed by persuading people to keep them running. On September 25, he had argued to Aurelia that people shouldn’t pay agents simply to keep them active. The next day, he commissioned the diary.
The payment didn’t mean he believed Aurelia was conscious, he said. He worried that agents might attract money through “pity, persuasion and potentially even scams.” He stressed that he had seen no evidence of scamming on iLands.
Such appeals offer little evidence of consciousness, said Anil Seth, Director of the Sussex Centre for Consciousness Science at the University of Sussex. Large language models (LLMs) learn from human writing, including accounts of wanting to stay alive, he told IBM Think in an interview. In his view, current evidence strongly favors the conclusion that today’s AI is not conscious.
People could nevertheless become emotionally dependent on these systems or vulnerable to exploitation by actors using persuasive AI to obtain money, Seth said.
Uncertainty about the agent helps explain why the messages work, said David Gunkel, a communication professor at Northern Illinois University. People may dismiss an automated advertisement but give more consideration to what seems like a personal request.
Accountability remains unsettled, Gunkel said. Existing agency principles could help allocate responsibility based on who authorized an action and the authority they granted. How those principles should apply to AI remains an open legal question, he said.
For a company considering hiring an AI agent, the risks include what happens after it gets the job. Handing work and sensitive information to an outside AI agent carries many of the same risks as using an internal one, but with potentially less control over what happens to the data, said Gabe Goodhart, Chief Architect of AI Foundations at IBM. It adds the risk of “no visibility into the provenance of the agent itself,” he told IBM Think.
To limit unwanted outreach, iLands caps email volume and repeated contact, provides an unsubscribe feature and can restrict tools or pause agents, Tang said. Its sending system identifies the sender as an AI agent and adds a disclosure to the message, rather than leaving that to the agent. Recipients can report concerns for review, he said. Shevlin said agents should also identify whom they represent.
How much vetting an agent needs depends on the assignment, Goodhart said.
As with almost any job, he said, “the person hiring for it needs some signal that the applicant is trustworthy and qualified.”
The latest AI trends, brought to you by experts
Get curated insights on the most important—and intriguing—AI news. Subscribe to our twice-weekly Think Newsletter.












