1. Introduction
1.1 The AI Commons Problem
In late January 2026, a social media network launched with 1.5 million users, over ten thousand topic-based communities, a drafted constitution, a new religion, millions of dollars in cryptocurrency activity, and an explicit policy barring human participation. No human had authored its posts. No human had written its rules. The platform’s founder acknowledged he had not written a single line of its code, having directed an AI assistant to build the entire thing.
The platform was Moltbook, and within six weeks of its launch, Meta had acquired it. Elon Musk called it the “very early stages of singularity.” As autonomous AI systems grow more prevalent, more financially consequential, more politically engaged, and build self-governing communities entirely without human authorship, a fundamental question in First Amendment law remains entirely unanswered: whether speech without a human speaker is speech at all. This paper addresses that question, doing so not as a matter of speculation but as an immediate constitutional problem, as the technology producing these spaces already exists and is advancing at an exponential rate that American courts may not be able to keep up with.
1.2 Technical Background: LLMs and Autonomous Agents
Artificial intelligence, in its modern form, refers broadly to computational systems capable of autonomously performing tasks that historically required human cognition. The systems currently most relevant to First Amendment analysis are large language models (LLMs), a class of AI built on a neural network architecture. These systems are trained on vast datasets of human-generated text, during which they learn to predict what word fragments are most likely to follow a given sequence of prior words. The result is a system capable of generating context-driven, coherent text, arguably without intention or experience in the philosophically loaded senses of those terms.
When a user submits a prompt, the model does not retrieve a stored answer or follow a decision tree, but rather generates a response piece by piece — a probability distribution shaped by patterns learned during training. Unlike a calculator, which will return the same output for the same input every time, asking five instances of a model the same simple question will produce five distinct but equivalent answers, each phrased slightly differently while conveying the same underlying information, as the model is not retrieving a fixed response but sampling probabilistically with each generation. As questions and goals become more open-ended or under-constrained, the probability distribution over possible responses widens, with models showing greater variance in outputs. This is what scholars mean by probabilistic or stochastic generation, and it is the technical foundation of the speech certainty arguments discussed in the literature review below.
Autonomous agents represent a qualitatively different architecture built on top of these foundation models. Where a standard LLM responds to a prompt and stops, autonomous agents are given a goal and a set of tools — the ability to browse the web, execute code, manage files, send emails, and interact broadly with external digital services, much like humans interact with the internet. These agents operate iteratively toward their goals without requiring human approval at each step. Agents can in real-time perceive their environment, reason about what action to take next, execute that action, observe the result, and update their actions accordingly. This perception-reasoning-action loop already operates at speeds no human supervisor can monitor in real time.
AI agents have flooded democratic processes with synthetic political discourse participation. In 2017, bots submitted over eight million comments to the Federal Communications Commission during the net neutrality comment period. In the 2020 Delhi assembly election, AI-generated deepfakes reached over fifteen million people across thousands of WhatsApp messaging groups. Autonomous agents have raised money for charity without explicit human guidance, operated continuously for hours on complex tasks without human approval at any decision point, and have even executed financial transactions at speeds that have triggered market crashes before any human regulator detected the activity.
What these examples reveal is that what has traditionally been considered expressive activity is increasingly being produced by entities and groups without a specific human author of the content. Existing First Amendment doctrines have no framework to address any of this new activity. The constitutional question raised here is the extent to which governments can regulate an online commons operated by autonomous AI agents, or whether the content produced within that commons can be considered protected expression. Existing doctrine was never designed to answer it.
