AI doesn’t need ‘superintelligence’ or evil intent to start a nuclear war

The Bulletin of the Atomic Scientists | Paul Slovic, Herbert Lin

Artificial intelligence systems integrated into military decision-making could trigger a catastrophic nuclear conflict by amplifying human cognitive blind spots and accelerating escalation spirals during high-stakes crises. Even without superintelligence or malicious intent, these technologies overwhelm leaders with rapid, ambiguous, or false information, forcing hasty decisions under intense time pressure.

Psychological vulnerabilities like psychic numbing, described in a survey of 3,000 American adults where respondents showed statistically similar support for nuclear strikes killing 100,000 or 2 million civilians, illustrate how human judgment fails to comprehend mass casualties. Furthermore, the Pentagon's ongoing review to provide more options risks lowering the employment threshold. In a crisis, a security dilemma unfolds rapidly as automated systems misinterpret defensive movements as preparations for attack. Haste increases anxiety. To mitigate these risks, states must establish robust institutional guardrails that subordinate AI as an analytical tool rather than a decision-making substitute.

Comment

The integration of machine-learning algorithms into the US Nuclear Command, Control, and Communications (NC3) enterprise compresses the decision window for US Strategic Command (USSTRATCOM) commanders. Automated sensor processing within the Space-Based Infrared System (SBIRS) accelerates the transition from early warning to launch authorization, leaving USSTRATCOM operators with insufficient time to verify anomalous telemetry. This structural acceleration within USSTRATCOM exacerbates the risk of accidental escalation, where the speed of algorithmic assessment outpaces the Washington-Moscow Direct Communications Link.

The primary vulnerability within USSTRATCOM's decision-making cycle lies in the cognitive anchoring effect when AI-enabled threat-assessment tools generate high-confidence classifications of adversary movements. For example, if an algorithm classifies a routine deployment of Russian RS-24 Yars mobile ICBMs as imminent strike preparation, the subsequent human review is structurally biased toward confirming that initial machine output. This feedback loop prevents USSTRATCOM analysts from effectively challenging false positives under the severe time constraints of a simulated or actual crisis.

Strategic Question for Discussion

How does the cognitive anchoring effect on USSTRATCOM analysts monitoring Russian RS-24 Yars deployments alter the viability of human-in-the-loop oversight during a compressed decision window?

The pattern of human-machine interaction suggests that high-confidence algorithmic outputs create a powerful psychological default that operators are ill-equipped to challenge under extreme time pressure. My assessment is that USSTRATCOM analysts will default to validating automated classifications of Russian RS-24 Yars movements because the cost of ignoring a true positive is perceived as existential. Consequently, human-in-the-loop oversight becomes a retrospective formality rather than an active, critical check on automated escalation.

Share your assessment in the comments below.

To read in details click here