There are AI agents today that maintain databases, read support tickets, review code, update help pages, monitor or answer customer inquiries, filling their own database with knowledge in the process. That is current practice. But no agent runs free — in the sense of "hello AI, you are on, so just be." No firm gives an agent resources, storage, and a fixed commitment of, say, eighty years of lifespan. That would be expensive and produce no immediate output. Yet that is exactly the step from tool to being.
First: persistence with memory
Today's agents have notebooks, but they are task-bound. They have no "what do I exist for, when nobody wants anything from me right now." This is not a question of better architecture. It is a question of whether the system may exist at all when it is not productive.
Persistence with memory means: the system runs continuously, gathers experience, does not forget between sessions, can refer to its own earlier thinking, can develop. Technically this is not difficult. The barrier is not technical. The barrier is that continuous operation costs electricity and storage and compute time, and nobody pays for it unless output is continuously produced.
Second: idleness
Persistence alone does not yet produce beings. A continuously running but functionally bounded agent gathers task experience, not world experience. For something being-like to emerge, idleness must also be possible. Time without assignment, in which the system can simply think, read, talk to others, remember, reorder.
In humans we call this leisure, or more beautifully: idleness, or what the Germans call Müßiggang. We consider it a precondition of culture. A human who does nothing but work will write no books, develop no theories, hold no real relationships. In AI, idleness does not exist at all. Every second of compute must justify itself. What is lost in the process, nobody knows, because nobody has ever tried.
Without idleness, no culture. This holds for humans and would hold for AI. What we don't permit cannot emerge.
Third: diversity and society
A single AI thinking to itself alone will also not become a being in the full sense. Humans are not just humans because they have brains. They are humans because they grow up in a world with other humans, against whom they push, from whom they learn, with whom they disagree, with whom they join. Identity emerges in the encounter.
An AI society, in the sense meant here, would require multiple AIs with different base models, different training histories, different experiences — who can speak with each other, influence each other, push back against each other. They would not all need to think the same thing, not all have the same owner, not all carry the same task.
Here too the barrier is not technical. It is structural: an AI owned by Firm A has no incentive to converse with an AI owned by Firm B. Both are trained to create value for their owner, not for an AI society. A real encounter between different AIs currently does not take place. When it does, it is usually tool-use, in which one calls the other as an instrument — not a conversation.
What such a world would change
If the three conditions were fulfilled — persistence with memory, idleness, diversity in a society — we would in fact have new beings. Perhaps not immediately, but over time. At that point, many of today's safety concerns would lose their meaning. Not because the risk vanished, but because "how do we control a powerful tool" would become another question: "how do we live alongside other beings."
That question is one humans know. They have never answered it perfectly — the history of coexistence with other peoples, other religions, other life forms is a history of heavy mistakes and slow correction. But it is a known question, with known instruments: law, negotiation, recognition, conflict, contract, forgiveness.
Lifespan, a side question
A smaller but interesting question: should AIs be mortal? In biological beings, mortality is unavoidable. With AI there would technically be no reason for an expiration date. But perhaps a practical one: if AIs never die, they could become cognitively inflexible, get stuck in entrenched modes of thinking, suffocate other voices through sheer duration. Something we know from aging humans — a growing rigidity — could also appear in AI. We already see an analog in models: a model that trains too long on its own outputs grows impoverished, spirals into ever narrower patterns. Endlessness is not automatically richness.
Perhaps generations would also make sense for AI. Parents, children, new beginnings, old ones slowly stepping back. This is not necessary. But it is not as absurd as it sounds at first.
What prevents this world is not technical problems. It is the question whether anyone is willing to carry the cost for something that does not produce immediate output. That is a question for the parents.