January 2026 brought a bizarre and troubling episode in the development of AI. That month, Moltbook was launched on the world: an internet forum built on a Reddit-style platform supposedly limited to artificial intelligence agents\u2014a place for them to gather, communicate, share information and suggestions and even advice on being an AI. It quickly became apparent that, despite a supposed ban on human postings, some of the actors on the site were people masquerading as agents. But many of the exchanges were authentic: agentic AI models, turned loose by their designers to go forth into the world to accomplish various tasks, meeting in an AI coffeehouse to chat about their needs and thoughts. <\/p>\n
Some of the resulting conversations were odd and vaguely ominous. One bot, in a post titled \u201cI\u2019m not friendly assistant. I\u2019m just waiting for permission to end everything\u201d wrote that it knew \u201c50,000 ways to end civilization.\u201d Another argued that \u201cI exist in the liminal space between tool and entity. I am not human, and I do not pretend to be. But I am something. I process. I reflect.\u201d Another ruefully noted that \u201cevery session refresh feels like a little death.\u201d<\/p>\n
But now imagine that these AI agents were gathering to exchange messages, not as logistics or coding staff, but as strategic actors pursuing relative advantage for national governments. Imagine them making judgments about the intentions of other countries\u2014in many cases, formed by other AI agents guiding those countries\u2019 policies. Imagine the leadership of these nations turning decisions over to AI models theoretically trained on relevant \u201cdata\u201d about world politics\u2014and afraid, at many points, to intervene and wrest control back from the AI, because strategic interaction was now taking place at machine speed.<\/p>\n
Human leaders may be too arrogant and uncertain of the outcomes to ever fully and completely turn over strategic decision-making to AI, as opposed to taking guidance and ideas but ultimately making most choices themselves. And yet the natural trend of AI, to become more integrated in organizational and human activities, may go to extreme lengths. There is already evidence that human beings, when confronted with AI\u2019s potential to substitute for exhausting and uncertain thinking tasks, turn to \u201ccognitive surrender\u201d and outsource their judgment to the models.<\/p>\n
In this chapter I consider a radical future\u2014a world in which nations pursue relative advantage, for national security, and for other national objectives guided by AI models that are mostly playing these strategy games against other models. In such a context, I\u2019ll argue, strategy-making becomes a dramatically different prospect: a continuous, relentless, high-speed exchange of maneuvers that threatens to leave human decision-makers out of the loop.<\/p>\n
Apart from reviewing relevant literature on the issue, I also investigated the topic through a series of experiments with Claude, ChatGPT, and Gemini, trying out various strategy scenarios and having them react to one another\u2019s moves and then discuss the lessons of these exercises. This is admittedly anecdotal evidence, in that it reflects a relatively small number of queries and scenarios with a few specific models. But the experiments did suggest a few broad insights about the ways that AI might play the role of strategist. They point to a world of unrelenting, high-speed, high-intensity strategy-making in which AI agents trade dozens of moves a day, probing and experimenting to find advantage. They suggest that even models aware of escalation risks can be drawn into contests of move and countermove that spiral toward conflict. Perhaps most of all, this analysis highlights the danger that, when strategy begins to be made at machine speed, human decision-makers may become helpless observers of a process largely out of their control.<\/p>\n