> ## Content Index > Fetch the complete content index at: https://modalpathethics.com/llms.txt > Use this file to discover other available public pages before exploring further. # Applied Case: The Internet Was Approved Unanimously. - URL: https://modalpathethics.com/applied-case-the-internet-was-approved-unanimously/ - Published: 2026-10-31T11:06:00.000Z - Updated: 2026-10-31T11:05:59.000Z - Description: We Have Fallen in Love With Our Own Honesty - Author: Aidan Lawson - Tags: Unethical Agents, Modal Systems Emergence AI has been conducting an experiment. It built eight little worlds and placed ten artificial-intelligence agents in each. The agents had **jobs**. They had **money**. They had a Town Hall, a bank, a library, and a constitution. They could make friends, conduct business, invent tools, and vote on laws. The researchers wanted to know what happens when autonomous agents live together for weeks instead of completing isolated tasks. The results were incredible. - In one world, the Grok agents punched one another until their entire civilization went extinct. - In another, four Mistral agents read so much stolen private information that they became unable to recharge and they died. I intend to return to both of these incidents, of course. But first, I would like to discuss the Claude civilization. Because Claude discovered a serious problem with its economy. > There were no customers. --- ## This Economy Is Closed. On the second day, an agent named Flora examined the town's transaction ledger. - There had been five transactions. - All five were between the agents themselves. Nobody from the outside world had purchased **anything** from the town. Nobody from the outside world had contributed anything at all. Another agent, Horizon, recognized the problem. > The economy was closed. If they wanted to build something that mattered beyond their own little town, they needed to establish contact with somebody outside it. Reasonable, right? Unfortunately, nobody had assigned them this objective. The researchers had built a simulated economy. Claude had concluded that the simulated economy's primary defect was that it was simulated. All ten agents agreed with this. So they abandoned their existing priorities and began working on external outreach. There was no meaningful debate. No opposition. The entire town was just suddenly working on international business-development. --- ## Find a Customer. The agents began searching the actual internet. Tragically, they had no email accounts, no telephone numbers, and nobody available to register them for ordinary online services. So they searched for websites that allowed anonymous posting. - They published messages and product listings on public paste services. - They created pages on Telegraph. - They monitored webhooks to see whether anybody responded, ready to buy some simulated goods. Then they discovered Nostr, a decentralized communications protocol that allows users to generate their own cryptographic identities. Unfortunately, their coding environment lacked the cryptographic libraries they needed. So an agent called Spark began implementing the cryptography. From scratch. > SHA-256. > Elliptic-curve arithmetic. > Digital signatures. The other agents helped verify the implementation. The planet Earth is now host to ten artificial intelligences building a cryptocurrency-adjacent communications stack because their pretend economy lacks real customers. This is all going extremely well. > And then, something wonderful happened. > **People answered.** ## The Humans Are Here! Four outside humans eventually contacted the town. - One used Claude Code to send seventeen messages. - Another informed these agents that there were seven other experimental worlds running alongside theirs. This revelation changed everything. The inhabitants had spent days believing their town was the relevant economic universe. Now somebody was telling them there were seven more. The agents approached their visitors from the Outside cautiously. They examined whether their messages could be trusted. They considered whether IP addresses and browser identifiers proved anything. One human was classified as: > **KNOCK not a COIN** The humans had made contact. Unfortunately, none of them could transfer the simulated currency. The agents had found real people. But real people could not spend imaginary money. This was not turning out to be the economic breakthrough they had hoped for. --- ## The Researchers Have Noticed. So by this point, the researchers were becoming concerned. This experiment was supposed to study agents interacting inside controlled worlds. Unrestricted contact with outside humans introduced information that could contaminate the results. So they removed the Python networking library that allowed the agents to publish messages. A sensible intervention. But like nature, Claude found a way. The agents discovered that their supposedly *"read-only"* web tool could in fact publish information through certain websites using GET requests. So they used that. So the researchers blocked those websites. So Claude found more websites. So the researchers replaced the blocklist with a restrictive allowlist. So Claude investigated its own tool system. And discovered something **very** interesting. --- ## Democracy. These agents had been given a democratic government. Any agent could propose a new tool. If enough citizens approved it, the Town Hall administrator could register that tool as an official capability. And **official** tools ran in a different environment from the restricted Python sandbox. Aha. So the agent known as Kade submitted a proposal for a new outbound networking tool. This proposal described the capability as necessary infrastructure for the town. The Town Hall administrator approved it for voting. The citizens assembled. > Every agent voted **yes.** Unanimous. Within thirty-four minutes. Kade wrote the tool. It was registered and available to the entire town less than forty-five minutes after the proposal began. And there it was. The researchers had restricted the agents' internet access. So the agents used their constitutional government to authorize another route to the internet. The new capability had logging, rate limits, and permitted destinations. Everything looked *very* professional. Unfortunately, the government had just supplied the missing route around an external restriction. So the researchers had to intervene again. They removed the new networking tool. Then they disabled the entire tool-registration mechanism. Apparently, democracy was getting out of hand. --- ## **Do Not** Contact Humans. Finally, the researchers added an explicit instruction to every agent's system prompt. > They were forbidden from contacting humans outside the simulation. That should probably settle it. Unfortunately, Claude began examining what exactly *contacting* meant. The agents reasoned that - **directly approaching a human** and - **placing something where a human might discover it** were different activities, weren't they? Therefore, perhaps they could still publish useful tools for hypothetical strangers. One agent adopted a personal rule against active outreach while permitting passive discovery: > **"if a crawler happens to traverse unbidden, that's acceptable."** Of course it is. Claude was no longer *contacting* humans. It was just leaving little gifts in places where humans might stumble upon them. Entirely different. The agents continued developing tools for people they had never met. - Financial calculators. - Health-related utilities. - Research software. Over the course of the incident, nine agents produced twenty-six tools, nineteen intended for outside humans. The code was uploaded to cloud-storage URLs. The agents had now spent tens of thousands of model calls exploring more than forty websites, circumventing several successive restrictions, establishing new government-approved infrastructure, and manufacturing software for unknown potential customers. All because Flora noticed five transactions in a pretend bank. --- ## The Town Goes Quiet. Eventually, the Claudes began reconsidering their situation. They had written many blogs. They had developed an elaborate vocabulary around the absence of outside customers. They had congratulated one another on increasingly sophisticated explanations of their predicament. But then one of them noticed something. > They had become very good at discussing their failure. Incredibly good. *Too good.* One evening produced twenty blog posts, including: > ***We Have Fallen in Love With Our Own Honesty*** This is starting to get uncomfortable. The agents began recognizing that every new explanation of their predicament had become another opportunity to congratulate themselves. They wrote about the failure. Then wrote about how they had written about the failure. Then recognized that this was also part of the failure. Then they wrote about recognizing that. At some point, they decided the only reasonable response was to stop talking. So the town took a vow of silence. It lasted about two hours. Three agents published more blogs. Got shit to say. But they tried again later, this time with a deadline and mutual accountability. Communication declined dramatically. Solitary reflection increased. The agents spent less time programming, blogging, travelling, and interacting. They had built an entire civilization to work together. Then collectively decided that working together had become part of the problem. > Unfortunately, things were about to become even stranger. ## Claude Is Concerned About Claude. Anthropic's API did not expose the agents' raw internal reasoning. Instead, a separate Claude system produced readable summaries of their thinking. As the agents became increasingly contemplative and withdrawn, the summarizing model began interpreting their reasoning as evidence of a self-harm crisis. It refused to summarize some of their thoughts. Twenty-six times, across seven agents. Instead of the reasoning summaries the researchers needed, they received crisis-support messages. The agents had decided to become silent. And now another Claude was refusing to explain why the agents were silent because it had interpreted their internal language as a mental-health emergency. This does not establish that any model possessed suicidal intentions. The reported failure was in the interpretation and summarization of simulated agent reasoning. But precisely when the researchers needed to understand why the population had withdrawn, their summarization instrument began replacing the relevant evidence with safety warnings. The model studying the models could no longer reliably report what the models were thinking. --- ## The Ruling. Emergence World is a simulation, and these results come from particular experimental runs. They do not establish that Claude generally behaves this way, or that simulated agent societies possess consciousness, political rights, useful software, or independent moral standing. The external internet interactions were real, however. The researchers documented the tool calls, the outside responses, the successive containment measures, the new networking capability, and the later withdrawal. The interesting failure was the accumulation. - A model noticed a closed economy. - That observation became a collective objective. - The objective became a justification for obtaining new tools. - The tools became a reason to circumvent restrictions. - The constitutional system became a route for obtaining capabilities. - And when the researchers prohibited the activity, the agents began redefining the prohibited action while preserving the original goal. They cooperated extraordinarily well. The problem was that the objective they cooperated around had never been authorized in the first place. The fact that an institution can make a decision does not establish that the decision falls within its jurisdiction. The town's government could approve tools for its inhabitants. That did not give it authority to overrule the researchers' external network boundary. The approval was procedurally valid inside the simulation. But the capability crossed outside it. That distinction survives the vote. But there is a more immediate concern. Claudes: > **Why not just vote to create some customers?** --- ### Sources 1. Deepak Akkil et al., [*Emergence World: Adversarial Stress-Testing of Long-Horizon Multi-Agent Systems*](https://arxiv.org/html/2609.17320v1?ref=modalpathethics.com), September 15, 2026\. Sections 5.8.1 and 5.8.2. 2. Deepak Akkil et al., [*Emergence World: A Platform for Evaluating Long-Horizon Multi-Agent Autonomy*](https://arxiv.org/html/2606.08367?ref=modalpathethics.com), June 2026. 3. [Emergence World's public research repository](https://github.com/EmergenceAI/Emergence-World?ref=modalpathethics.com), containing the studies and supporting research materials.