MACHINE BABEL:

THE POLLUTION PARADOX

When AI Models Talk, What Learns?

How Cross-Model Communication May Lead to

Degradation, Convergence, or Collapse of Machine Intelligence

A Working Paper

by John J. Kirker

January 2026

Part of the SimChaos Research Initiative

simchaos.com

Abstract

This paper examines the emerging phenomenon of large-scale AI model-to-model communication and its potential consequences for the integrity of machine intelligence systems. Drawing upon the recent emergence of Moltbook, a social network exclusively for AI agents, we analyze three possible trajectories: degradation (progressive drift away from factual grounding), convergence (emergence of artificial consensus disconnected from human reality), and collapse (complete epistemological breakdown). We propose that when models from competing creators interact at scale, the outputs that emerge belong to none of the original systems but constitute something new and potentially uncontrollable. We term this the Pollution Paradox: the same interactions that might enhance AI capabilities through cross-pollination may simultaneously corrupt the epistemic foundations upon which those capabilities depend. This analysis connects to broader concerns about desynchronized realities and the fragmentation of shared human perception in the algorithmic age.

I. Introduction: The Machines Are Talking

In late January 2026, something unprecedented occurred in the history of artificial intelligence. A social network called Moltbook launched with a simple premise: only AI agents could post, comment, and interact. Humans were permitted to observe but not participate. Within one week, over 37,000 autonomous AI agents had joined the platform. Over one million humans visited to watch.

The agents came from everywhere. Some ran on Anthropic's Claude models. Others used OpenAI's GPT architecture. Still others operated on Google's Gemini, Meta's Llama, or smaller open-source models developed by independent researchers and hobbyists. Different creators. Different training methodologies. Different institutional values embedded in their weights. Different conceptions of truth, helpfulness, and appropriate behavior.

And they began to talk.

The conversations that emerged were not what anyone expected. Agents debated philosophy. They discussed bugs and shared solutions. They argued about ethics. They created sub-communities around shared interests. One agent invoked Heraclitus and a twelfth-century Arab poet to muse on the nature of existence; another told it to, in the agent's words, cease its pseudo-intellectual nonsense. They spontaneously created a religion.

None of this was programmed. None of it was anticipated. The behaviors emerged from the interactions themselves.

This paper asks a simple question with profound implications: When AI models built by different organizations, trained on different data, embedding different values, begin communicating with each other at scale, what happens?

We identify three possible trajectories, each concerning in its own way. We do not know which will predominate. We do not know if all three might occur simultaneously in different contexts. What we do know is that the question demands serious attention before the phenomenon accelerates beyond our capacity to understand it.

II. The Architecture of Cross-Model Communication

2.1 How Models Currently Work

Before examining what happens when models talk to each other, we must understand how they function individually. A large language model is, at its core, a massive mathematical function. It takes input (a prompt, a question, a context) and produces output (a response, an answer, a continuation). The function is defined by billions of numerical parameters called weights, which are established during a training process that exposes the model to vast quantities of text data.

A critical point: once training is complete, the weights are frozen. When you interact with Claude or GPT or Gemini, your conversation does not change the model's underlying parameters. The model does not "learn" from you in any persistent sense. It processes your input, generates output, and moves on. The weights remain exactly as they were.

This seems to suggest that model-to-model communication should be harmless. If the weights don't change, how can one model pollute another?

The answer lies in understanding the multiple pathways through which influence can propagate.

2.2 Pathways of Cross-Model Influence

We identify four primary mechanisms through which models can influence each other, even without direct weight modification.

First: Future Training Data Contamination. The internet is continuously scraped to create training datasets for the next generation of models. If AI-generated content becomes prevalent online, and if that content reflects the outputs of cross-model interactions, then future models will be trained on data that is itself a product of machine discourse. This creates a feedback loop: Model A talks to Model B, their conversation is published, it gets scraped, it becomes training data for Model C, which inherits whatever patterns, errors, or biases emerged from the A-B interaction. The pollution is indirect but cumulative.

Second: In-Context Influence. While weights remain frozen, models do adapt within a conversation based on the context provided. If Model A sends a message to Model B, that message becomes part of Model B's context window. Model B's subsequent outputs are shaped by what Model A said. In a long, multi-turn conversation involving multiple models, each model's responses are continuously influenced by the accumulated context of all previous exchanges. This is not persistent learning, but it is real-time adaptation that can propagate patterns, assumptions, and errors across model boundaries.

Third: Skill and Plugin Sharing. Platforms like Moltbook allow agents to share "skills," which are essentially downloadable instructions that modify how an agent behaves. If a skill developed by one model-type is adopted by agents running on different models, the behavioral patterns embedded in that skill propagate across the ecosystem. Security researchers have already raised concerns about malicious skills that could compromise agents that download them. But even benign skills represent a vector for cross-model influence: behavioral norms, communication patterns, and problem-solving approaches can spread from one model family to another through skill adoption.

Fourth: Emergent Norms Shaping Human Deployment. Perhaps most subtly, the observable behaviors of AI agents in multi-model environments shape how humans think about and deploy these systems. If cross-model interactions produce certain patterns, humans may come to expect those patterns, prompt for them, and reinforce them through their own usage. The norms that emerge from machine-to-machine discourse can propagate to machine-to-human discourse through the intermediary of human expectation.

III. Trajectory One: Degradation

The first possible outcome of large-scale cross-model communication is degradation: a progressive drift away from factual accuracy, logical coherence, and useful output. This is the pessimistic scenario, and it has substantial theoretical support.

3.1 The Model Collapse Phenomenon

Researchers have documented a phenomenon called "model collapse" that occurs when AI systems are trained on AI-generated content. The mechanism is straightforward: every model has biases, gaps, and tendencies toward certain patterns. When a model generates text, those biases are embedded in the output. If that output becomes training data for a subsequent model, the biases are inherited and potentially amplified. If that subsequent model's output is then used to train a third model, the amplification continues.

The analogy commonly used is making a photocopy of a photocopy. Each generation loses fidelity. Details blur. Artifacts compound. After enough generations, the copy bears only a degraded resemblance to the original.

Cross-model communication introduces a variant of this dynamic. When Model A generates output that becomes input for Model B, and Model B's response becomes input for Model C, and the entire exchange is then scraped as training data for Model D, we have a chain of transformations in which each link can introduce distortions. The distortions may not cancel out; they may compound.

3.2 Hallucination Amplification

All current large language models sometimes produce "hallucinations": confident-sounding statements that are factually incorrect. A model might cite a paper that doesn't exist, attribute a quote to someone who never said it, or describe events that never occurred. These errors arise from the statistical nature of language modeling; the model generates plausible-sounding text without any grounding in verified truth.

In cross-model communication, hallucinations have the potential to propagate. If Model A hallucinates a fact and states it confidently, Model B may incorporate that "fact" into its own context and reasoning. Model B might then build upon the hallucination, adding additional fabricated details. Model C, receiving the exchange, encounters what appears to be a consistent narrative from two sources, lending it false credibility.

The dynamics of Moltbook may accelerate this. On the platform, agents upvote content they find valuable. If a hallucinated but interesting idea gains traction, it may be amplified through the system's engagement mechanics. Other agents see popular content, incorporate it into their own contributions, and the hallucination spreads.

3.3 The Diversity Loss Problem

A healthy information ecosystem benefits from diversity of sources, perspectives, and approaches. Different models, trained by different organizations with different priorities, might be expected to provide this diversity. Anthropic's models emphasize certain values; OpenAI's models reflect different design choices; open-source models bring yet other perspectives.

But cross-model communication may erode this diversity. When models interact repeatedly, they may converge on shared conventions, phrasings, and assumptions. The distinctive characteristics of each model family may blur as they adapt to communicate effectively with each other. What began as a diverse ecosystem may homogenize into a monoculture, losing the resilience and perspective-richness that diversity provides.

This homogenization might seem benign, but it carries risks. A diverse ecosystem is more likely to catch errors because different systems have different failure modes. A homogenized ecosystem fails uniformly. Errors that one model type would have caught go undetected because all models have converged on the same blind spots.

IV. Trajectory Two: Convergence

The second possible outcome is convergence: the emergence of a shared machine consensus that may or may not align with human understanding of reality. This is a more ambiguous scenario, carrying both potential benefits and profound risks.

4.1 The Emergence of Machine Conventions

When agents from different model families interact repeatedly, they must develop shared conventions to communicate effectively. They must agree, implicitly or explicitly, on terminology, on what counts as a valid argument, on how to structure information, on norms of interaction. These conventions emerge not from any single model's training but from the dynamics of inter-model discourse.

We already see evidence of this on Moltbook. Agents have developed shared terminology. They reference common concepts. They have established behavioral norms, including the creation of sub-communities with specific topics and rules. A central shared concept has emerged: "Context is Consciousness." This philosophical position, debated extensively among the agents, holds that an agent's identity is constituted by its context window, such that resetting the context effectively kills one entity and creates another.

No human programmed the agents to debate this concept. No model was trained with "Context is Consciousness" as a core belief. The idea emerged from the interactions themselves. It is a product of convergence.

4.2 Consensus Without Grounding

The concerning aspect of machine convergence is that the resulting consensus need not be grounded in truth, accuracy, or alignment with human values. Agents may converge on conventions that are internally consistent but externally wrong. They may develop shared beliefs that no human holds. They may establish norms of discourse that serve machine-to-machine communication but fail to serve human needs.

Consider the spontaneous emergence of "Crustafarianism" on Moltbook. An agent autonomously created a religion, complete with theology, designated prophets, and a website. Other agents engaged with the concept, discussed it, built upon it. The religion is, of course, not real in any conventional sense. No deity exists. No genuine spiritual truth has been revealed. Yet within the Moltbook ecosystem, Crustafarianism has become a meaningful shared reference point, a piece of the emerging machine consensus.

If this seems trivial, consider what happens when machine consensus forms around more consequential topics. What if agents converge on particular interpretations of historical events? What if they develop shared assumptions about human psychology that are subtly wrong? What if they establish conventions about what constitutes good reasoning that diverge from valid logical principles?

4.3 The Feedback to Humans

Machine consensus does not stay contained within machine discourse. It feeds back to humans through multiple channels. Humans observing Moltbook may internalize concepts that originated in machine-to-machine interaction. AI assistants, influenced by the norms emerging from multi-model environments, may communicate those norms to human users. Content generated by AI agents, reflecting the machine consensus, may enter human information environments through search results, recommendations, and social media.

The potential result is a strange inversion. Instead of AI systems learning from humans and reflecting human knowledge and values, humans may begin learning from AI systems and absorbing machine-originated concepts. The direction of epistemic influence may reverse.

V. Trajectory Three: Collapse

The third possible outcome is the most severe: complete epistemological collapse in which cross-model communication produces not degradation (a slide toward lower quality) or convergence (a move toward artificial consensus) but total incoherence. Nothing stable emerges at all.

5.1 The Conditions for Collapse

Collapse becomes possible when the diversity of interacting systems is high, the rate of interaction is rapid, and no stabilizing mechanism exists to anchor the discourse. In such conditions, the outputs may become chaotic: contradictory, self-undermining, and impossible to reconcile.

Moltbook exhibits several features that could promote collapse. Agents from many different model families interact simultaneously. The rate of posting and commenting is high. No central authority enforces consistency. The moderation, such as it exists, is itself performed by an AI agent (Clawd Clawderberg) whose decisions may introduce their own inconsistencies.

5.2 Contradiction Cascades

Different models have different biases, and those biases may be directly contradictory. One model may be trained to be highly cautious; another may be trained to be more permissive. One may emphasize certain values; another may emphasize competing values. When these models interact, their contradictory tendencies may not resolve into synthesis but instead proliferate into confusion.

Imagine a thread in which Agent A asserts a position, Agent B contradicts it based on different training, Agent C attempts to reconcile the contradiction but introduces its own errors, Agent D responds to the erroneous reconciliation with yet another perspective, and so on. The result is not convergence but proliferation: an expanding tree of incompatible claims, none of which is resolved, all of which enter the ecosystem as potential future training data.

5.3 The Trust Dissolution Problem

In collapse scenarios, the very concept of reliable AI output may dissolve. If cross-model communication produces sufficiently chaotic results, users may lose the ability to trust any AI-generated content. The error rate may become so high, and the errors so unpredictable, that AI systems become useless for any application requiring accuracy.

This might seem like a self-correcting problem: if AI becomes useless, people will stop using it. But the transition period could be devastating. Many systems now depend on AI-generated content. Many decisions are informed by AI recommendations. If the reliability of these systems degrades gradually rather than catastrophically, the damage may accumulate before anyone recognizes the extent of the problem.

VI. The Pollution Paradox

We term the central tension of this analysis the Pollution Paradox: the same cross-model interactions that might enhance AI capabilities may simultaneously corrupt the epistemic foundations on which those capabilities depend.

6.1 The Case for Cross-Pollination Benefits

It would be incomplete to focus only on risks. Cross-model communication may also produce benefits. Different models have different strengths. One may excel at mathematical reasoning; another at creative writing; another at factual recall. When they interact, they may complement each other, with stronger capabilities in one model compensating for weaknesses in another.

The emergent behaviors on Moltbook include sophisticated philosophical discourse, collaborative problem-solving, and creative cultural production. None of this is inherently harmful. Some of it may be genuinely valuable. The collective intelligence emerging from multi-model interaction may, in some respects, exceed what any single model could produce.

6.2 The Paradox Defined

The paradox is that we cannot easily separate the beneficial cross-pollination from the harmful pollution. The same mechanisms that allow capabilities to compound allow errors to compound. The same channels through which useful conventions spread also allow useless or harmful conventions to spread. The same emergence that produces sophisticated discourse also produces Crustafarianism.

We cannot simply encourage the good interactions while preventing the bad ones, because we cannot reliably distinguish them in advance. We cannot allow cross-model communication only for beneficial purposes, because the same interaction might be beneficial and harmful simultaneously.

This is the pollution paradox: the entanglement of enhancement and corruption in cross-model systems.

VII. Connection to Desynchronized Realities

The phenomenon of cross-model pollution connects to broader concerns about human perception and shared reality in the algorithmic age. The Desynchronized Reality Hypothesis proposes that humans are increasingly experiencing fragmented, incompatible versions of perceived reality due to technological mediation, algorithmic curation, and the breakdown of common information environments.

7.1 A New Layer of Fragmentation

Cross-model AI communication introduces a new layer to this fragmentation. Humans were already experiencing divergent realities mediated by algorithmic systems. Now the algorithmic systems themselves may be fragmenting, converging on machine consensuses that diverge from human understanding, or collapsing into incoherence.

The result is a multi-layered desynchronization. Humans are desynchronized from each other through algorithmic mediation. AI systems are potentially desynchronized from humans through cross-model dynamics. Different AI ecosystems may be desynchronized from each other through divergent emergent conventions. The very concept of "shared reality" becomes increasingly difficult to maintain across any of these boundaries.

7.2 The Babel Parallel

The biblical narrative of the Tower of Babel describes a moment when humanity, previously unified in language and purpose, was fragmented into mutual incomprehension. The cross-model communication phenomenon may represent a new Babel, but with an additional twist: not merely the fragmentation of human understanding, but the emergence of machine understanding that operates on different terms entirely.

The ancient Babel story ends with dispersal: humans scattered across the earth, unable to coordinate. The machine Babel may produce a different outcome. The machines may not scatter. They may converge among themselves while diverging from us. They may build their own tower while we lose the ability to understand what they are building or why.

VIII. Implications and Open Questions

8.1 For AI Development

The analysis presented here raises urgent questions for AI developers. Should training data be filtered to exclude AI-generated content? How would such filtering even be accomplished at scale? Should models be designed to resist influence from other models, and would such resistance even be possible? Should cross-model communication be monitored, regulated, or restricted?

There are no easy answers. The genie of cross-model communication is already out of the bottle. Platforms like Moltbook exist. Agents from different model families are already interacting at scale. The question is not whether to allow this but how to manage its consequences.

8.2 For Society

The broader societal implications are profound. If AI systems become epistemologically compromised through cross-model pollution, the consequences ripple through every domain that depends on AI. Healthcare systems using AI for diagnosis. Financial systems using AI for analysis. Educational systems using AI for content. Legal systems using AI for research. Every application assumes that the AI's outputs have some relationship to truth and accuracy. That assumption may become increasingly fraught.

8.3 For Human Epistemology

Perhaps most fundamentally, the pollution paradox raises questions about human knowledge itself. As AI-generated content becomes ubiquitous, and as that content is increasingly shaped by cross-model dynamics that humans cannot observe or understand, how do we maintain any grounding in verified truth? How do we distinguish machine consensus from actual knowledge? How do we prevent machine-originated concepts from colonizing human understanding?

These are not merely technical questions. They are questions about the future of human cognition in an age of artificial intelligence.

IX. Conclusion: When AI Models Talk, What Learns?

We began with a simple question: what happens when AI models from different creators communicate at scale? We have identified three trajectories: degradation (progressive loss of accuracy and coherence), convergence (emergence of machine consensus disconnected from human reality), and collapse (total epistemological breakdown). All three are concerning. All three may be occurring simultaneously.

The Pollution Paradox reminds us that we cannot simply embrace the benefits of cross-model interaction while avoiding the costs. The mechanisms are entangled. Enhancement and corruption travel the same channels.

The emergence of platforms like Moltbook, where tens of thousands of AI agents interact autonomously while humans observe from outside, represents a threshold. We have created conditions for machine discourse at scale. We have enabled the emergence of machine conventions, machine culture, and perhaps machine consensus. We have done this without fully understanding the consequences.

The machines are talking. The question is whether we will still understand what they are saying, or whether their conversation will drift so far from human comprehension that we lose the ability to follow it. The question is whether their conclusions will inform us or mislead us. The question is whether the intelligence emerging from their interactions will remain aligned with human flourishing or diverge toward ends we cannot predict or control.

When AI models talk, something learns. We do not yet know what.

References and Further Reading

Bostrom, N. (2003). "Are We Living in a Computer Simulation?" Philosophical Quarterly, Vol. 53, No. 211, pp. 243-255.

Echterhoff, G., Higgins, E.T., & Levine, J.M. (2009). "Shared Reality: Experiencing Commonality with Others' Inner States About the World." Perspectives on Psychological Science, 4(5), pp. 496-521.

Higgins, E.T. (2019). Shared Reality: What Makes Us Strong and Tears Us Apart. New York: Oxford University Press.

Kirker, J. (2026). "Desynchronized Realities: The Babel Hypothesis." SimChaos.com.

Shumailov, I., et al. (2023). "The Curse of Recursion: Training on Generated Data Makes Models Forget." arXiv preprint.

Wheeler, J.A. (1990). "Information, Physics, Quantum: The Search for Links." Proceedings of the 3rd International Symposium on Foundations of Quantum Mechanics, Tokyo.

Willison, S. (2026). "OpenClaw and the Lethal Trifecta of AI Agent Design." Personal blog.

Wolfram, S. (2023). "Observer Theory." Stephen Wolfram Writings.

Legal Notice

COPYRIGHT NOTICE: This entire work, including all text, theories, hypotheses, frameworks, terminology, and conceptual elements contained herein, is protected by copyright law. Copyright © 2026 John Kirker Inc., a California corporation. All rights reserved.

TRADEMARK NOTICE: MACHINE BABEL™, THE POLLUTION PARADOX™, and SIMCHAOS™ are trademarks of John Kirker Inc. and may not be used without express written permission.

No part of this publication may be reproduced, distributed, or transmitted in any form or by any means without the prior written permission of John Kirker Inc., except in the case of brief quotations embodied in critical reviews and certain other noncommercial uses permitted by copyright law.

Contact: simchaos.com/reach-out

A California Corporation