When AI Enters Nuclear Decision Systems
What happens when artificial intelligence enters the most dangerous decision-making systems humans have ever built? In this episode of AI-Curious, we talk with Ankit Panda and Andrew Reddie about the real risks of AI and nuclear weapons.
Original episode title: The Real Risks of AI and Nuclear War, w/ Ankit Panda & Andrew Reddie
Guests
Ankit Panda
Nuclear-policy expert and author
Ankit Panda is the Stanton Senior Fellow in the Nuclear Policy Program at the Carnegie Endowment for International Peace. His research covers nuclear strategy, emerging technologies, and international security. He is also the author of The New Nuclear Age, a book about changing nuclear capabilities and competition.
Separates everyday AI assistance from the much harder problem of delegating consequential nuclear decisions, where evaluation and command authority must remain explicit.
Andrew Reddie
Researcher in security, technology, and governance
Andrew Reddie is a professor at the Goldman School of Public Policy at the University of California, Berkeley, and faculty director of the Berkeley Risk and Security Lab. His work examines emerging technology, security, nuclear strategy, and risk, including how AI may affect military decision-making and crisis stability.
Adds the institutional boundary: testing systems is not enough unless leaders can explain authority, escalation, accountability, and where automation stops.
Watch on YouTube ▶What this conversation is really about
Ankit Panda and Andrew Reddie map AI uses in nuclear systems by consequence, from routine assistance and logistics to intelligence analysis, targeting, and command decisions. The danger is not one imminent autonomous weapon. It is that new tools may compress time, hide the source of evidence, amplify cyber risk, and encourage people to defer to a machine under pressure. The phrase “human in the loop” means little unless someone has enough authority, time, and information to interrupt the process. The historical examples are reporting leads, not proof of any current program.
From the conversation
The argument in focus
Ankit Panda
Nuclear-policy expert and author
Ankit Panda is the Stanton Senior Fellow in the Nuclear Policy Program at the Carnegie Endowment for International Peace. His research covers nuclear strategy, emerging technologies, and international security. He is also the author of The New Nuclear Age, a book about changing nuclear capabilities and competition.
Separates everyday AI assistance from the much harder problem of delegating consequential nuclear decisions, where evaluation and command authority must remain explicit.
Andrew Reddie
Researcher in security, technology, and governance
Andrew Reddie is a professor at the Goldman School of Public Policy at the University of California, Berkeley, and faculty director of the Berkeley Risk and Security Lab. His work examines emerging technology, security, nuclear strategy, and risk, including how AI may affect military decision-making and crisis stability.
Adds the institutional boundary: testing systems is not enough unless leaders can explain authority, escalation, accountability, and where automation stops.
“…all the way up to an artificial intelligence advisor telling the president of the United States whether or not he or she should think about retaliating to an attack that may or may not be incoming.”
Evidence status
Forecast / proposal with high-stakes examples
The conversation applies technical and historical material to nuclear decision systems. It is not evidence of a tested or deployed system and requires authoritative corroboration.
Boundary map
Where the system stops
- What the system handles
- Potential analysis, sensing, simulation, and decision support under severe time and information constraints.
- What remains human
- Command authority, interpretation, escalation, veto, legal responsibility, and the consequences of error.
- What remains open
- Can human authority remain meaningful when speed, automation bias, and incomplete evidence compress the time to challenge the system?
Ideas worth carrying forward
- Classify AI uses by consequence, reversibility, and escalation potential instead of applying one blanket nuclear-AI policy.
- Preserve independent evidence channels so synthesis does not erase which sensor, source, or assumption drove a recommendation.
- Define the veto operationally: which person can stop the action, at what point, with what information, and under what time pressure?
- Test high-stakes systems against automation bias, false warnings, cyber compromise, and synthetic-data assumptions—not only accuracy benchmarks.
What this changes Monday
Ask leadership, risk, and technical teams to select one consequential AI-supported decision and draw its control path from input to action. Mark every place where sources are merged, uncertainty is hidden, time is compressed, or a reviewer is expected to approve without meaningful authority. Require an independent evidence path and a practiced stop or fallback procedure. Separate low-risk assistance from recommendations that can alter security posture or trigger escalation. The useful output is not a promise that humans remain involved; it is a demonstrated ability to inspect, challenge, delay, and overrule the system when conditions change.
Original episode notes
What happens when artificial intelligence enters the most dangerous decision-making systems humans have ever built?
In this episode of AI-Curious, we talk with Ankit Panda and Andrew Reddie about the real risks of AI and nuclear weapons. Ankit is the Stanton Senior Fellow in the Nuclear Policy Program at the Carnegie Endowment for International Peace and author of The New Nuclear Age.
Andrew is a professor at UC Berkeley’s Goldman School of Public Policy and faculty director of the Berkeley Risk and Security Lab.
This is not a conversation about Skynet or killer robots. Instead, we explore the more grounded, near-term ways AI is already intersecting with nuclear strategy, military decision-making, cyber risk, early warning systems, targeting, and command and control.
We look at the full spectrum of risk, from AI helping edit policy memos to AI systems advising leaders during a possible nuclear crisis.
We also discuss why cybersecurity may be the most urgent near-term concern, how AI could affect nuclear deterrence and crisis stability, and why explainability, testing, and human judgment matter so much in high-stakes environments.
Along the way, we revisit the famous Stanislav Petrov incident, explore dead hand systems and automation bias, and ask what it really means to keep humans in the loop when AI is moving faster than institutions can adapt.