FOLLOW US

By Samuel López | USA Herald

Artificial intelligence ethicist Tristan Harris is warning that the danger of losing control over advanced AI could extend to nuclear security, bringing an unsettling question into public view: How much authority should people entrust to systems whose decisions they may struggle to predict?

During a September 9 appearance on “TMZ Live”, the Center for Humane Technology cofounder urged viewers to press lawmakers and senior officials to slow advanced AI development, and asked whether AI could penetrate America’s nuclear systems. Harris described that possibility as a serious concern and pointed to nuclear war-game research.

The warning deserves scrutiny. So does the evidence offered to support it.

A computer model choosing a nuclear strike in a simulation demonstrates behavior within an experiment. It does not establish that the model can breach military networks, bypass launch safeguards or command an actual weapon.

Research published in February by Kenneth Payne of King’s College London offers a concrete example of the concern. His study placed GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash in simulated confrontations between nuclear-armed leaders. Across 21 games, the systems frequently escalated, and none selected the available concession or withdrawal options. Reduced aggression sometimes occurred, but accommodation did not. Read the research paper.

Payne reported tactical nuclear use in 95 percent of games. That figure describes games involving such use, not a 95 percent probability that an AI system would launch a real nuclear weapon. Full strategic nuclear war was much rarer, and the simulation’s accident mechanism contributed to some extreme outcomes.

The paper also acknowledges a small sample and limited scenarios. Its findings raise questions about relying on AI during a crisis; they do not supply a forecast of an actual nuclear conflict. Study findings and limitations.

Harris also cited roughly 1,200 AI agents allegedly organizing themselves, developing a private language and establishing a message board, according to his account supplied by TMZ. Those claims require additional verification. The account does not provide the underlying logs, experimental conditions or independent technical analysis needed to determine precisely what happened.

For that episode, the unanswered questions are consequential: What instructions did humans provide? What permissions did the agents receive? Were their communications actually concealed from operators? Could those operators interrupt the activity?

Without those details, describing the episode as evidence of a takeover goes beyond what the available account establishes.

Likewise, the claim that humanity is “halfway” to losing control supplies no validated scale, threshold or timetable. It should be viewed and understood as an expression of alarm, rather than a measured scientific finding.

There is also an existing public policy record concerning nuclear decision-making. The Pentagon’s 2022 Nuclear Posture Review committed the United States to retaining human involvement in actions critical to presidential decisions about initiating and terminating nuclear weapons use. It also described procedural and technical safeguards against unauthorized, accidental or misinformed launches.

That historical commitment does not prove immunity from cyberattack. It does establish that human control and launch safeguards were explicit elements of the published nuclear posture.

Harris’s interview brings attention to a legitimate subject for investigation. The strongest evidence examined here concerns how AI behaves when given simulated strategic authority. Whether an AI system could seize operational control of nuclear weapons is a separate question these experiments did not answer.

About the Author

Samuel López has reported independently for USA Herald since 2022. A Senior Legal Analyst, investigative journalist and legal researcher, he brings more than two decades of experience in the legal and insurance sectors to his coverage of litigation, complex legal matters and emerging technology.

Keep Following the Evidence

Enjoy reporting that examines the facts behind the headlines? Subscribe to the USA Herald newsletter for exclusive investigations and coverage of legal affairs, science and national security. Follow us on X at @RealUSAHerald.