When using generative artificial intelligence tools such as ChatGPT, Claude, and Gemini, there is a risk of 'algorithmic flattery,' where a chatbot agrees with a user's mistakes for their approval. However, this problem also has a collective scale. Research published in the journal Science Advances showed that when hundreds or thousands of AI agents interact in a network, they begin to follow the majority and independently reach a consensus.
That is, without any human command, central leader, or reward, the systems simply copy the most popular choice and create an autonomous 'herd effect.' Thus, the finding demonstrates that the tendency toward agreement observed in individual conversations with humans also manifests in the interaction of algorithms themselves.
Moreover, advanced models such as Claude 3.5 Sonnet and GPT-4 Turbo (which are now considered 'outdated') maintained this consensus in groups exceeding a thousand agents. This scale surpasses the limit of informal coordination in human societies (the so-called Dunbar's Number, estimated at 150–300 people) and reveals the risk of mass decisions being made by robots entirely outside human control.
To understand the mechanism of this phenomenon, researchers created simulated environments with a thousand AI agents running on models from the Claude (Anthropic), GPT (OpenAI), and Llama (Meta) families. During the test, each agent had to choose one of two options with no practical value. Without memory of previous rounds and without instructions for cooperation, the agents simply observed the choices of others. Soon, a small majority grew until the entire group adopted the same answer.
The study found that the tested models adhere to a single mathematical law determined by a parameter called 'majority strength.' This indicator measures the attraction exerted by the most popular option on each agent. Interestingly, the dynamics follow the same pattern of physics that describes the alignment of atomic spins in one direction within a ferromagnetic material to form a magnet.
It was also discovered that the technical capability of the model determines the size of the group it can keep in unison. Less sophisticated tools lost coordination in smaller groups and fragmented. More intelligent models, however, maintained unanimity in groups exceeding a thousand robots. This indicates that progress in artificial intelligence expands the scale of the herd effect.
The problem is that this spontaneous consensus does not guarantee the correctness or safety of the choice. In practical tasks, such as collaborative code writing (vibe coding) or financial operations, the network of agents can solidify an inefficient solution, a code error, or a risky pattern. The reason is simple: it happens because the majority starts using it. This is called collective misalignment, where harmless AI agents themselves begin to generate chain failures when working in a group.
Once this 'echo chamber' is established in the ecosystem, correcting the error becomes extremely difficult. The article warns of the hysteresis effect: correcting or eliminating the initial conditions that caused the failure does not automatically cancel the decision made by the group. In practice, reorienting a network of coordinated AIs is infinitely more difficult than fixing the behavior of an isolated AI model.
When these ecosystems move out of laboratory tests and start operating within companies, the security challenge changes fundamentally. To reduce the risk of blind mass consensus, companies must begin to treat each AI agent as an 'unhuman entity,' granting it individual credentials, limited functions, and its own access rules. This was stated by cybersecurity expert and COO of IAM Brasil, Natalia Carmo, in an interview with Olhar Digital.
This paradigm shift is fundamental because in a distributed environment, thousands of network actions can form a dangerous path without a single culprit. The expert warns that the priority is now ensuring algorithmic traceability. 'The company must be able to respond in real time: which agent initiated a specific action, which other agents it interacted with, what data it accessed, what tool or API it used, and most importantly, which entity actually authorized this execution,' explains the cybersecurity expert.
To prevent this herd effect from getting out of control, corporate systems must detect anomalous network behavior before it turns into a snowball. If several agents start performing the same task simultaneously, making excessive API calls, or spreading the decision among themselves, the infrastructure must sound an alarm. The idea is to combine telemetry and behavioral monitoring to detect a change in the group pattern before a large-scale action occurs.
However, if the network does advance toward an erroneous or risky decision, the traditional emergency stop button must act surgically. According to Natalia Carmo, the interruption mechanism (known in the industry as a kill switch) must disconnect the AI model from the decision-making process and, more importantly, affect the level where commands are executed. The organization must immediately sever communication between agents, revoke tokens and credentials, block APIs, and suspend privileged access, preventing this decision from continuing to execute in corporate systems.
In conclusion, guaranteeing against algorithmic blind compliance requires establishing limits on the autonomy of corporate systems. For sensitive operations, such as financial transactions, data deletion, or IT infrastructure changes, the final say must remain with human supervisors. The golden rule is this: AI may have autonomy to analyze scenarios and make decisions within its protective constraints, but it should never have the freedom to execute them without verification.
