What happens when everyone asks the same consultant?
Millions of people turn to LLMs for advice daily, consulting on various personal topics, from how to learn a new skill to how to write a message to their best friend. Each interaction might feel special, but behind them all are the same models.
Large language models are trained on an abundance of human data, compress different sources of knowledge, views, and opinions into a single set of model parameters - E Pluribus Unum Consilium[1]. What used to be the job of many different consultants has been centralized to the same handful of models.
Pluribus, an Apple TV show, offers a striking analogy. In the show, most of humanity joins a single collective consciousness called the joined. The horror of the show is not that they are cruel, they are not. The joined are actually calm, peaceful, and nice. A Previous blog presents the analogy using similarities between the joined and AI. We want to extend this analogy and use Pluribus as an illustration to explore the potential risks humanity faces as millions of people are using the same consultants as their primary source of judgment.
Loss of Agency In Pluribus, every person who is part of the collective, automatically consults with "them" (the joined) before acting on anything. We observe a similar pattern in our relationship with AI. We are looking for confirmation and approval before acting. The email we won't send before we let AI check it, the trip we ask help planning. It is that we feel we can't make those decisions without it. Modesty Argument gives a good explanation why this might happen. When we disagree with another person, both sides should adjust their reasoning toward the other until they get full agreement. The problem is when the other side is an LLM consultant. We are highly limited in adjusting the model toward our perspective, therefore, it leaves us to mainly adjust ours. Second, we lack The Proper Use of Doubt when consulting with models. Since we know models process enormous amounts of information, way more than we possibly can, it seems to be a good enough reason for us to believe we might be wrong.
Loss of Originality In the show, everyone shares the same knowledge, ideas, and responses. If you ask one or ask many, you would get the same exact narrative. There is no individuality behind the words. Language models create similar situations. Recent works [2][3][4] show that AI might improve individual creativity, but measurably shrinks the diversity of what the population produces. This is not surprising, we can recognize the "LLM style", similar posts, names, and acronyms appear over and over in different contexts. We would rather replace our own novel ideas and words, as messy and unclear as they are, with others, we converge toward the in-distribution, which by design could not be novel or original.
Loss of Purpose In Pluribus, since all the joined share the same knowledge and capabilities, everyone can do everything. Specialization therefore means nothing. And people become interchangeable. With Advanced agent properties, a question arises: is the only human purpose to do whatever AI can't? What does Make an Extraordinary Effort even mean if we could never read or process the amount of data an agent can? Access to a capable tool can be easily mistaken for being adequate to judge its output. While we see AI as a tool to make us better and more capable, it does not replace our need to continuously develop our professional judgment, critical thinking, and ability to ask the right questions.
Loss of Moral Pluralism In Pluribus, the collective optimizes society toward peace, harmony, and minimal suffering objectives. It succeeds so much that it reaches an Unforeseen maximum. Using morally unthinkable ways to keep food provided to the population. Similarly, what if we optimize models toward objectives like long-term well-being? Does it mean the model should convince everyone to become vegetarians if it means preserving resources? Or should the model convince us to give up on private cars if it is more eco-friendly? Optimizing to one moral objective is itself a moral commitment. When millions of people consult with the same model, they inherit this same commitment without ever making it. The cost is moral pluralism, and the ability to make those moral decisions themselves.
Loss of Verification In pluribus, the collective insists it can't lie. Yet no one could verify it. Without a reliable independent outside source to verify it, there is no way to verify if the joined claims are true. Current models create a similar challenge, when multiple sources are combined into one response, verification becomes extremely challenging. When backtracking the origins of each claim is extremely difficult, we might miscalibrate our trust in the model. Appropriate reliance therefore should come from both a reliable, traceable source of truth as well as our ability to independently verify it.
Loss of Authorship In the show, everyone speaks the same voice, the joined's voice. There is no individual responsible for it. This raises critical questions about intellectual authorship. Who deserves the credit and also the responsibility? What happens when things go wrong? The current relationship is asymmetric: when things go well, we wish for the credit, but when things go wrong, we point at the model. If a model helps to generate a novel medication, who should get the credit? the user? the developer? The model? Or the many prior works that were used as input? What happens if it generates the same idea to multiple different people? As many people use AI as their primary tool, the absence of agreed roles becomes everyone's problem at once.
We used the extreme scenarios presented in the show not to claim humanity is facing a collective consciousness. But to highlight the potential risks we might face when many people use the same consultants as their primary source of judgment. The goal is not to stop using AI. On the contrary, AI holds many advantages and could serve as an amazing tool for humanity. We believe we can create healthy practices on both the user's side and the model side. This requires deliberate design, practices, and roles that will keep humans actively engaged and critical. With the goal of building AI that makes us better thinkers without making us the same thinkers.
Jiang, Liwei, et al. "Artificial hivemind: The open-ended homogeneity of language models (and beyond)." Advances in Neural Information Processing Systems 38 (2026).
Doshi, Anil R., and Oliver P. Hauser. "Generative AI enhances individual creativity but reduces the collective diversity of novel content." Science Advances 10.28 (2024): eadn5290.
What happens when everyone asks the same consultant?
Millions of people turn to LLMs for advice daily, consulting on various personal topics, from how to learn a new skill to how to write a message to their best friend. Each interaction might feel special, but behind them all are the same models.
Large language models are trained on an abundance of human data, compress different sources of knowledge, views, and opinions into a single set of model parameters
- E Pluribus Unum Consilium[1]. What used to be the job of many different consultants has been centralized to the same handful of models.
Pluribus, an Apple TV show, offers a striking analogy. In the show, most of humanity joins a single collective consciousness called the joined. The horror of the show is not that they are cruel, they are not. The joined are actually calm, peaceful, and nice. A Previous blog presents the analogy using similarities between the joined and AI. We want to extend this analogy and use Pluribus as an illustration to explore the potential risks humanity faces as millions of people are using the same consultants as their primary source of judgment.
In Pluribus, every person who is part of the collective, automatically consults with "them" (the joined) before acting on anything.
We observe a similar pattern in our relationship with AI. We are looking for confirmation and approval before acting. The email we won't send before we let AI check it, the trip we ask help planning. It is that we feel we can't make those decisions without it. Modesty Argument gives a good explanation why this might happen. When we disagree with another person, both sides should adjust their reasoning toward the other until they get full agreement. The problem is when the other side is an LLM consultant. We are highly limited in adjusting the model toward our perspective, therefore, it leaves us to mainly adjust ours.
Second, we lack The Proper Use of Doubt when consulting with models. Since we know models process enormous amounts of information, way more than we possibly can, it seems to be a good enough reason for us to believe we might be wrong.
In the show, everyone shares the same knowledge, ideas, and responses. If you ask one or ask many, you would get the same exact narrative. There is no individuality behind the words.
Language models create similar situations. Recent works [2] [3] [4] show that AI might improve individual creativity, but measurably shrinks the diversity of what the population produces. This is not surprising, we can recognize the "LLM style", similar posts, names, and acronyms appear over and over in different contexts. We would rather replace our own novel ideas and words, as messy and unclear as they are, with others, we converge toward the in-distribution, which by design could not be novel or original.
In Pluribus, since all the joined share the same knowledge and capabilities, everyone can do everything. Specialization therefore means nothing. And people become interchangeable.
With Advanced agent properties, a question arises: is the only human purpose to do whatever AI can't? What does Make an Extraordinary Effort even mean if we could never read or process the amount of data an agent can? Access to a capable tool can be easily mistaken for being adequate to judge its output. While we see AI as a tool to make us better and more capable, it does not replace our need to continuously develop our professional judgment, critical thinking, and ability to ask the right questions.
In Pluribus, the collective optimizes society toward peace, harmony, and minimal suffering objectives. It succeeds so much that it reaches an Unforeseen maximum. Using morally unthinkable ways to keep food provided to the population.
Similarly, what if we optimize models toward objectives like long-term well-being? Does it mean the model should convince everyone to become vegetarians if it means preserving resources? Or should the model convince us to give up on private cars if it is more eco-friendly? Optimizing to one moral objective is itself a moral commitment. When millions of people consult with the same model, they inherit this same commitment without ever making it. The cost is moral pluralism, and the ability to make those moral decisions themselves.
In pluribus, the collective insists it can't lie. Yet no one could verify it. Without a reliable independent outside source to verify it, there is no way to verify if the joined claims are true.
Current models create a similar challenge, when multiple sources are combined into one response, verification becomes extremely challenging. When backtracking the origins of each claim is extremely difficult, we might miscalibrate our trust in the model. Appropriate reliance therefore should come from both a reliable, traceable source of truth as well as our ability to independently verify it.
In the show, everyone speaks the same voice, the joined's voice. There is no individual responsible for it.
This raises critical questions about intellectual authorship. Who deserves the credit and also the responsibility? What happens when things go wrong?
The current relationship is asymmetric: when things go well, we wish for the credit, but when things go wrong, we point at the model. If a model helps to generate a novel medication, who should get the credit? the user? the developer? The model? Or the many prior works that were used as input? What happens if it generates the same idea to multiple different people? As many people use AI as their primary tool, the absence of agreed roles becomes everyone's problem at once.
We used the extreme scenarios presented in the show not to claim humanity is facing a collective consciousness. But to highlight the potential risks we might face when many people use the same consultants as their primary source of judgment. The goal is not to stop using AI. On the contrary, AI holds many advantages and could serve as an amazing tool for humanity. We believe we can create healthy practices on both the user's side and the model side. This requires deliberate design, practices, and roles that will keep humans actively engaged and critical. With the goal of building AI that makes us better thinkers without making us the same thinkers.
In Latin: "Out of many, one counsel"
Rios-Sialer, Ian. "The Homogenization Problem in LLMs: Towards Meaningful Diversity in AI Safety." Pluralistic Alignment Workshop at ICML 2026.
Jiang, Liwei, et al. "Artificial hivemind: The open-ended homogeneity of language models (and beyond)." Advances in Neural Information Processing Systems 38 (2026).
Doshi, Anil R., and Oliver P. Hauser. "Generative AI enhances individual creativity but reduces the collective diversity of novel content." Science Advances 10.28 (2024): eadn5290.