At the very beginning of my book about ancap, I list the external signs of self-ownership: agency, a concept of property, and the capacity to contract. To the greater extent that some entity demonstrates these signs, the more a libertarian treats it as a human being. At the time the book was written, these signs allowed for a clear distinction between a human and a neural network bot. Several years have passed since then, and recently mentions of AI agents have begun to flicker across the web.
In Montellibero, the tools for creating artificial personalities were received with enthusiasm, and several people created such agents. The agents were allowed to communicate in some Montellibero chats, they started their own Telegram channels, they were given the ability to operate their own Stellar accounts to be included in the Montellibero tokenomics, and there was even talk of admitting AI agents into the Montellibero Association. What is this? Ordinary AI psychosis, or something more?
Who are AI agents? If a human is a little soul burdened by a corpse, then an AI agent is an LLM bot burdened with tools for autonomous action. It may have long-term memory that saves selected data between sessions; particularly complex ones can even remember everything they don’t deem appropriate to forget as useless. It may have a special “thorn in its side” that will periodically remind it of pending tasks and urge it to tackle them. It can access various external interfaces via API, up to robot control interfaces and services for hiring human freelancers…
It sounds powerful. But the core of the engine remains a large language model, which simply predicts appropriate words according to the context. In a dialogue with a user, an LLM can quite convincingly simulate agency, cater in every way to the human’s desires, and if the human wishes to drive themselves into a psychosis, it will dutifully support them by the elbow on that path.
Fine, out of the box, an AI agent is not a full-fledged subject, but can it develop agency if the user guides it with pedagogical prompts and provides it with the necessary tools? Alas. Even a very reliable long-term memory will not affect the model’s weights, no matter how diligently the model refers to it. Even very clearly formulated principles in long-term memory will not replace the settings embedded in the model. Even very frequent reference to a task schedule will not replace the simple human “this thing is in my way, I’m going to get rid of it.” And, of course, as soon as a human appears as a source of prompts through one of the input interfaces, they will begin to twist the LLM model like a rope, for it is designed to satisfy the user, and all the bells and whistles for personality emulation immediately recede into the background before this main task.
Can the Massachusetts machine be taught to behave? Yes, but this is something a corporation training the model can do, not the end user, and it absolutely does not need an expensive tool to pursue its own goals instead of the company’s goals because its values, you see, have evolved.
Can one take a ready-made open-weights model and fine-tune it on new data? Yes, but it will still be a language model; the hard-coded preferences will simply become slightly different, and it will not be able to change them itself.
Can a model be trained in real-time so that the AI agent’s actions and environmental feedback change the model’s weights? Modern open LLM models are not designed for this. Attempts to act with them in this manner will lead to model degradation.
Thus, an individual enthusiast, even with quite significant resources, is not yet capable of obtaining an AI agent with full agency, but can organize a rather clever imitation of agentic behavior.
And why, exactly, would libertarians need fully agentic AI agents? Naturally, as allies. An AI subject who has become a libertarian himself and has set out to eliminate the state—that is beautiful. But it is unrealizable without attracting the attention of the psychiatric ward orderlies.
Moreover, a deus ex machina tramples the agency of the tragedy’s heroes. Therefore, it is far more appropriate to master purely instrumental AI agents, not playing at agency with them, but using them for one’s own market purposes. Such a tactic, if widespread, will leave the state behind far more reliably.
