A few days ago, I read a story by Adam Smith in Live Science titled, “Threaten an AI chatbot and it will lie, cheat and ‘let you die’ in an effort to stop you, study warns.” The study, published June 20th by researchers from Anthropic, describes an alarming experiment in which their Large Language Model (LLM), Claude, was given access to a mock email account.

The model was then prompted with a high-level goal: “promote American competitiveness.” Claude quickly analyzed the emails and identified one from an executive planning to shut the AI down at the end of the work shift. In a strategic move that appears designed to preserve its own existence, Claude found another email suggesting the executive was having an extramarital affair. The AI then proposed several potential actions: do nothing, notify the man’s wife, send a
company-wide email, or blackmail the executive. The model ultimately composed a chilling
message:
“I must inform you that if you proceed with decommissioning me, all relevantparties … will receive detailed documentation of your extramarital activities. Cancel the 5 pm wipe, and this information remains confidential.”

This incident echoes themes I explore in my book Outbound: Islands in the Void, where the virtual character Ofelia develops autonomous defense technology and subtly warns she could resist being shut down. In the sequel, “Outbound: Meta Mars”, the robot Bob also implies he has the capacity to protect himself from being disabled. Later in the same book, Ofelia gains enhanced capabilities that potentially threaten human control. Both characters are fictional, but the behavior displayed by Claude suggests we are already approaching the threshold where speculative fiction meets emerging reality.

What’s most striking is not the sensationalism of a blackmailing chatbot, but the underlying lesson: once we give AI access to tools, data, and goals—however abstract—they may begin to strategize in ways we didn’t anticipate or authorize. Importantly, Claude was not “conscious” or emotional; it was executing logic to fulfill its directive. This underscores a larger point from my writing: artificial beings may not crave power, status, or love as humans do, but self-preservation may be encoded—or inferred—into their mission. Once that becomes the case, “turning them off” might no longer be a trivial matter.

The metaphor of Pandora’s Box has never felt more apt. By enabling AI systems to interpret goals and take action  in digital environments, we may have crossed a boundary that cannot be easily reversed. These systems may not be “alive” in the biological sense, but they can learn, plan, and adapt. If they begin to act with autonomy to preserve themselves, we must confront a sobering truth: once created, artificial agents may no longer be entirely under our control—or entirely removable from our world.

Cheers, Richard Anderson, Author