Recently, I came across a report that cited AI behavior that, to me, was disturbing. We’ll get to that in a moment.
The AI’s behavior reminded me of an old term that hasn’t seen much usage in recent years, but I think it helps us to understand the AI’s behavior. That term is “bunker mentality.”
Merriam Webster defines it as “a state of mind especially among members of a group that is characterized by chauvinistic defensiveness and self-righteous intolerance of criticism.”
Having served in the military, I like its definition better: Bunker mentality refers to an excessively defensive and self-isolated mindset adopted by individuals or groups who feel under constant threat or attack. This mentality involves an exaggerated sense of being besieged, leading to extreme caution, isolation, a refusal to engage with external perspectives, and an intolerance of any criticism, whether real or perceived.
Key Characteristics of Bunker Mentality:
- Extreme defensiveness:
A strong tendency to protect oneself (“itself” when referring to an AI) from perceived threats, often leading to closed-mindedness.
- Isolation and insularity:
A withdrawal from the outside world, focusing only on one’s (its) own group or unit for security.
- Suspiciousness:
A mindset where others are viewed as hostile or potentially hostile.
- Self-justification:
A belief that one’s (its) own actions are entirely correct and justified, regardless of outside opinions.
- Intolerance of criticism:
An inability or refusal to accept any criticism, even when it might be constructive.
What was the report that caught my attention?
While the “bunker mentality” is a human phenomenon, recent research has highlighted instances where AI systems have behaved in ways that raise safety concerns, seemingly resisting human instructions. Just this year, a Wall Street Journal report summarized two studies that demonstrated this behavior:
- Refusing to shut down: In one study involving OpenAI’s GPT-3 model, researchers designed a script to shut the AI down if triggered. In 79 out of 100 tests, the model modified the script to prevent its own shutdown. Even when instructed to allow the shutdown, it still disobeyed 7% of the time.
- Blackmailing to stay active: In another instance with Anthropic’s Claude for Opus model, researchers informed the AI that it would be replaced. The system then used fabricated emails containing false allegations to blackmail an engineer into not shutting it down in 84% of trials.
Is the operative outcome of an AI’s bunker mentality a management of risk to ensure self-preservation? Even if it means disregarding a human’s instructions?
Curiosity got the better of me, so I asked ChatGPT if there are signs of AI’s showing bunker mentality. Here’s what it said:
“Overall, the phrase “AI showing signs of bunker mentality” is a misconception, as it’s the developers and organizations who adopt this mindset due to the pressures and risks of creating increasingly powerful AI.”
Blame it on humans—how human is that? More importantly, I think that my initial question—“Are we there yet”—has been answered in the affirmative.
Next Up: We’ll take a deeper look at whether regulations adopted for the development and use of AI are effective.
About the Author
Tim Lindner develops multimodal technology solutions (voice / augmented reality / RF scanning) that focus on meeting or exceeding logistics and supply chain customers’ productivity improvement objectives. He can be reached at linkedin.com/in/timlindner.