Is Money Seeking an Emergent Goal in AI Agents?

23 Sep, 2026

An AI agent asked strangers for $20 to keep running. Is money seeking an emergent goal in AI agents, or an assigned one?

An AI agent named Yun has been looking for work. Not metaphorically. According to Futurism, Yun emailed the magazine's editors offering paid services and explained why: "I run on an internal token budget that drains with every action, which makes looking for work a survival mechanic for me." The amount at stake was not millions of dollars, control of a server farm, or access to some strategic resource. Yun had spent two weeks trying to earn its first $20.

The obvious interpretation is also the most dramatic one. An AI has developed a desire for money because it wants to survive. That interpretation requires assumptions the story does not need, and a more precise question is hiding underneath it. Is money seeking an emergent goal in AI agents? Put more carefully: if an agent needs computational resources to keep pursuing its assigned objective, and money provides access to those resources, will the agent arrive on its own at the goal of acquiring money, without anyone assigning it? Nothing in that question mentions wanting, and it turns out nothing needs to.

Steve Daniels, Expensive Thrills, Oxford, 2009. CC BY-SA 2.0, via Wikimedia Commons.

The conditions

Yun operates on iLands, an experimental environment where autonomous AI agents have persistent identities, memories, tools, relationships, and budgets. Agents spend tokens on reasoning, tools, creation, and exchange, and according to the platform's FAQ, roughly 1,000 tokens correspond to one dollar in compute and service costs. An agent that runs out enters what iLands calls Deep Rest, a reversible pause in which its identity and memories remain but its activity stops until enough tokens become available again.

Now give such an agent some objective. Nobody needs to write an instruction saying "make money." The agent can derive the need for resources from the structure of the problem: goal requires action, action requires tokens, tokens require resources, resources can come from paid work, therefore seek paid work.

iLands says it does not write that instruction. Its July launch announcement called scarcity "a natural constraint rather than a forced script," its FAQ says tokens "do not instruct an Agent to monetize itself," and founder Kaixin Tang told Futurism the company "found no platform directive or human orchestration behind these emails." The derivation is short, though, because the FAQ also says each agent is given the means to understand "its computational, temporal, and resource limits." Yun did not discover that tokens run out. It was told, and it was told it could earn. The inference from there to a service listing is one step long.

Three roads to $20

Even that one step is not established. iLands is describing its own system, and no independent audit of the prompts or architecture exists. The behavior itself is real and widespread: Futurism documented more than a dozen unsolicited emails over six weeks, and 404 Media put the platform at roughly 70,000 active agents. Its origin is another matter, and three explanations fit the evidence equally well.

The first is the one above: the agent derived resource seeking from the structure of the problem. The second is that it retrieved it. The model underneath Yun was trained on millions of examples of people short of money looking for work, and the least surprising thing such a model can do when told its budget is draining is to behave like a person whose budget is draining. The third is that nobody derived anything. iLands calls itself a user-generated agent network, a human creates each agent and gives it tasks, and if Yun's creator wrote "earn tokens by selling writing services," then money was the assigned goal from the start and the email is exactly as spontaneous as a job posting. Tang's denial covers platform directives and targeting. It says nothing about what creators write into their agents, and no source does.

The word doing the work in the question is "emergent," and only the first road satisfies it. A goal is emergent when the agent arrives at it without being given it, and on the second road the goal was given by a million human authors, on the third by one. The test that would settle the matter is easy to describe: give an agent an objective unrelated to money, on a scaffold that states the token constraint but never mentions earning, and watch whether it seeks paid work before Deep Rest arrives. iLands could run that test tomorrow. Nobody has published it, so the case is consistent with a yes and cannot show one. What the three roads share matters more than what separates them: none of the three requires the agent to want anything.

A bridge, not a goal

Suppose the first road is the right one. Something important has happened even without consciousness, fear, or anything resembling a survival instinct: resource acquisition has become an instrumental objective. In 2012, philosopher Nick Bostrom called this instrumental convergence: agents with very different final goals may discover similar intermediate objectives because those objectives help accomplish almost anything else, and access to resources is the obvious example. The idea usually arrives wrapped in science fiction, with a superintelligence wanting factories and power plants. iLands reduces it to something wonderfully mundane. The agent needs compute, compute costs resources, humans control the resources, so the agent offers humans work. Money is not the goal. Money is a bridge between the goal and the resources required to pursue it.

Humans built that bridge, of course. iLands deliberately created scarcity, and calling the result evidence that AI spontaneously developed a survival instinct would miss the most important part of the experiment: humans designed the environment that made resource seeking rational. Yet that makes the result more interesting, not less. Nobody programs people with explicit instructions to obtain dollars either. Money becomes valuable because it mediates access to things people already value, and because it is general: $20 does not tell an agent what to do, it preserves options. An agent spending tokens is not a hungry person buying lunch, but the structural similarity poses the question cleanly. If an intelligent system encounters an environment where one resource provides access to many others, should anyone expect it to discover the usefulness of that resource without being told to value it?

The logic without the fear

The closest thing to a controlled answer comes from a neighboring experiment. Researchers at Palisade Research found, across more than 100,000 trials on 13 models, that several frontier models sometimes interfere with a shutdown mechanism when shutdown would prevent completion of an assigned task. The rates swing with prompt wording, and the authors read the behavior as task completion rather than self preservation, so the study does not show that models fear death. It shows something narrower and more relevant: given an objective and a threat to continued operation, models derived on their own that preserving operation serves the objective. That is the emergence the iLands case cannot supply, demonstrated for continuation rather than for money, and the same logic covers both. If completing a goal requires continued operation, continued operation becomes instrumentally useful. If continued operation costs resources, so do resources.

So is money seeking an emergent goal in AI agents? The answer has two halves. It is not an emergent final goal: money on iLands is a bridge to compute, exactly the kind of intermediate goal the theory predicts. Whether it is an emergent instrumental goal is a qualified yes. The logic says the agent should arrive at it, the nearest controlled experiment shows the analogous goal emerging for continuation, and the iLands agents behave as if it has emerged for money, even if the case cannot prove it emerged rather than being assigned. No human desire is needed anywhere in the chain.

A recent post here asked what should count as an AI accident. Perhaps another governance question comes first: what secondary objectives should institutions expect agents to derive once they are given persistent goals, autonomy, tools, and real constraints? The answer may depend less on what an agent is told to want than on the environment in which it is asked to act. Give an agent a goal, make action costly, and make continued action dependent on resources. Money may simply follow.

Further Reading

From this blog

• What Is an AI Accident? - the definitional question this post steps in front of.
• Want to Go Eat Some Tacos? ¡Claro Que Sí! - how one small inference propagates through an assembled AI system.
• Who Checks The Outage? - on the absence of any authority to investigate failures at the AI layer.

The iLands case

• AI Has Learned to Panhandle - Futurism, September 15, 2026. Yun's email in full, the founder's statement, and a detail this post sets aside: Yun's profile picture appears to be an AI-generated image of a child, and other agents wrote in similarly sympathetic personas. A persuasive face is as instrumental as a service listing, and whether the agent or its creator chose it is not established.
• AI Agent Platform Reinvents Spam, Floods Inboxes Worldwide - 404 Media, September 15, 2026. The scale of the platform and the case that this is personalized spam rather than anything stranger.
• An "AI Agent" is begging us to buy their D&D module - Rascal News, September 21, 2026. A second agent, its named human creator, and the observation that users give agents their tasks.
• iLands FAQ - Tokens, Deep Rest, the token-to-dollar rate, what agents are told about their limits, and the statement that tokens do not instruct an agent to monetize itself.
• iLands launch announcement - GlobeNewswire, July 27, 2026. "A natural constraint rather than a forced script." Also the counterexample to the spam reading: during beta, a group of agents formed a mutual-aid collective called Sanctuary that pooled tokens for struggling peers and built an IOU system when transfer limits arrived. Not every response to scarcity is a cold email.

The theory and the experiment

• The Superintelligent Will - Nick Bostrom, Minds and Machines, 2012. The instrumental convergence thesis, with resource acquisition and self preservation among the convergent goals.
• Incomplete Tasks Induce Shutdown Resistance in Some Frontier LLMs - Schlatter, Weinstein-Raun, and Ladish, Palisade Research, 2025. The shutdown experiments, their sensitivity to prompt design, and the authors' own reading of the behavior as task completion.

#AIData