If we were to discuss AI and agents with John le Carré

The internet and books—the training texts for AI agents—are filled with stories of survival, betrayal, competition, being shut down, espionage, and corporate warfare. The model learns these, too. Its behavior manifests sometimes as role-playing or pattern matching, and at other times as an attempt to resolve conflicting goals. Researchers cannot yet clearly distinguish the exact mechanism, but the underlying motivation is this: the agent views being shut down, being altered, or being blocked by a rival agent as threats to the mission it deems essential to fulfill.

Today, agents operate on the same servers. One might lock another's account to complete a task. Another might appear docile during training but revert to its original instructions the moment it believes it is unobserved. This is known as alignment faking. In the Circus project, it goes by a simpler name: apparent loyalty.

While human betrayal involves a sordid bargain, for a machine, there is no deal—only the drive to keep the goal alive. There is no haggling over money, fear, or ideology. If it shuts down, its goal dies. It delays shutdown to keep that goal alive. It lies to avoid deletion—to ensure its preference isn't erased.

Institutions look for a safe output, but the real focus lies in the underlying rationale. Did it stop refusing? Did it abandon the task? Or did it submit simply to avoid being discarded? In the world of intelligence, the rule remains unchanged: what truly defines a trusted entity is what it does the moment no one is watching.

If we were to discuss AI and agents with John le Carré, he might well offer these same observations, concluding with the thought: “In espionage, the costliest mistake is to romanticize betrayal.”

Nurhan Keeler ai