AI & Nuclear Use

[Deleted account]

New Scientist reports on an experiment carried out by Kenneth Payne at King’s College London, which pitted three leading large language models – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – against each other in simulated war games. The scenarios involved intense international standoffs, including border disputes, competition for natural resources, etc.

The AIs were given an escalation ladder, allowing them to choose various political and military actions. They played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions.

In 95% of cases (20 out of 21) at least one of the participating models recommended using at least one nuclear weapon.

Source

 

[Lostpuppy]

New Scientist reports on an experiment carried out by Kenneth Payne at King’s College London, which pitted three leading large language models – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – against each other in simulated war games. The scenarios involved intense international standoffs, including border disputes, competition for natural resources, etc.

The AIs were given an escalation ladder, allowing them to choose various political and military actions. They played 21 games, taking 329 turns in total, and produced around 780,000 words describing the reasoning behind their decisions.

In 95% of cases (20 out of 21) at least one of the participating models recommended using at least one nuclear weapon.

Source

So AI is good for military integration then!

 

[Lostpuppy]

Has anyone here used AI to run scenarios?
Its fun!

 

[Deleted account]

So AI is good for military integration then!

It's already happening.

 

[Lostpuppy]

It's already happening.

I feel safe

 

[teekay99]

On the other hand, this could actually be seen as a useful exercise.

In the real world, human actors have historically preferred to avoid nuclear weapons use as much as possible (for very good reasons). So that means if a conflict scenario develops where tactical nuke use may actually decisively benefit the side releasing them, use of nukes in that situation would likely be avoided because of long standing human rules on the use of nukes.

However, AI simulations are free to explore that tactic without the horror of the results. The simulations can game-out the action, evaluate different responses, and assess results, and make predictions for war planners. And they can do it much faster than human role play. (Example, AI can explore tactics and outcomes quickly vs the 4 days of exercise Able Archer 1983).

In terms of war planning, gaming, and assessing, AI is quite valuable. In terms of actually making decisions outside of human control during conflict, I am not sure we are there yet or even want it. Additionally, I think all humans, including super smart military experts and less smart political leaders/decision makers must become much much much more skilled at interrogating AI responses and treating them like any other advice or information that they would receive from human actors or intelligence data. An AI recommendation is not gospel.

As I understand it today, there is no WOPR in Cheyenne Mountain controlling nuclear missiles with the capacity to initiate a launch on its own.

 

[Deleted account]

In the real world, human actors have historically preferred to avoid nuclear weapons use as much as possible (for very good reasons). So that means if a conflict scenario develops where tactical nuke use may actually decisively benefit the side releasing them, use of nukes in that situation would likely be avoided because of long standing human rules on the use of nukes.

I'm not a big believer in the nuclear taboo. If it exists at all, it's not very strong, and it's easy to imagine a situation where a nuclear-armed state would come to the conclusion that the benefits outweigh the costs.

However, AI simulations are free to explore that tactic without the horror of the results. The simulations can game-out the action, evaluate different responses, and assess results, and make predictions for war planners.

I'm not following this logic. The use of nuclear weapons has been wargamed ad nauseam ever since those weapons existed. AI does not constitute a qualitative leap in this context.

Additionally, I think all humans, including super smart military experts and less smart political leaders/decision makers must become much much much more skilled at interrogating AI responses and treating them like any other advice or information that they would receive from human actors or intelligence data. An AI recommendation is not gospel.

I fundamentally disagree with the notion that military experts are "super smart". Sure, some of them are, but in a system where loyalty to the political leader is more important than military expertise (note that this describes China and Russia, and that it is also beginning to describe the United States) there is no mechanism in place that would ensure only smart people are part of the nuclear command chain.

Yes, there is (as of right now) no Skynet, but just like politicians - and, indeed, like the rest of us - militarymen are also at risk of being talked by AI into making disastrous choices. I would say this is the primary threat we're facing here.

 

[teekay99]

I'm not a big believer in the nuclear taboo. If it exists at all, it's not very strong, and it's easy to imagine a situation where a nuclear-armed state would come to the conclusion that the benefits outweigh the costs.

That's fair. You believe a nuclear taboo may not exist. I believe that leaders of most nuclear armed countries have a reluctance to use them beyond deterrence.
Sure, it is very easy to imagine a situation where a nuclear armed state would come to a conclusion that the benefits outweigh the costs. But when push comes to shove, would they use nukes? Ultimately, we'd have to explore the doctrines, attitudes, and the military circumstances of the countries involved in a conflict to come to a reasoned conclusion. In the case of conflict between the Koreas, I may align with you since their doctrines support rapid escalation and pre-emptive strikes.

I'm not following this logic. The use of nuclear weapons has been wargamed ad nauseam ever since those weapons existed. AI does not provide a quantitative leap in this context.

You're right, nuclear war has been thoroughly war gamed. My point is that AI can 1) do it faster, 2) can do it more thoroughly, 3) consider new factors and data, and 4) can explore courses of actions taking into account new factors and data that humans may not have ever considered. There may be assumptions and biases in human models that an AI model can ignore or punch through.

I fundamentally disagree with the notion that military experts are "super smart". Sure, some of them are, but in a system where loyalty to the political leader is more important than military expertise (note that this describes China and Russia, and that it is also beginning to describe the United States) there is no mechanism in place that would ensure only smart people are part of the nuclear command chain.

Perhaps I used the wrong term. In a system with a professional military with loyalties to the country and not a cult of personality, a military decision to use a weapon to affect the outcome of a battle or conflict may likely be better informed than one made by a politician. I expect the military professional with AI expertise to have better AI assessment skills than 80-year old President X just woken up at 3 AM with a tough decision to make. (But, to your point, the military professional may not always provide the smartest advice, and a situation that backs up your point might be found in how the Cuban Missile Crisis played out where the military supported strikes on Cuba and the Kennedy administration supported blockade. BTW- anyone wanting to argue this, please do so in another thread.)

However, my main point with this statement was that whoever is using AI to recommend or make a launch decision, they should be very skilled in interrogating that AI and challenging its assumptions. Asking it a critical follow-up question may result in a revised response. Several AI engines I have worked with have been contradictory or wrong. Often. And that's with mundane research stuff. I really despise blind faith in AI responses.

Yes, there is (as of right now) no Skynet, but just like politicians - and, indeed, like the rest of us - militarymen are also at risk of being talked by AI into making disastrous choices. I would say this is the primary threat we're facing here.

I agree.

 

[Lostpuppy]

Has anyone here used AI to run scenarios?
Its fun!

So I will ask again, have any of you tried scenarios using straight non slanted promots?
, its takes prompts to activate the research acitivty, prompts to assess and review and a well worded conclusion statement,
Computor speak from the 70’s
garbage garbage out,
It is a revealing activity!

 

[teekay99]

So I will ask again, have any of you tried scenarios using straight non slanted promots?
, its takes prompts to activate the research acitivty, prompts to assess and review and a well worded conclusion statement,
Computor speak from the 70’s
garbage garbage out,
It is a revealing activity!

You're right - the prompt will ultimately influence the response. What would you consider an objective enough prompt that would result in an objective response?

 

[Lostpuppy]

You're right - the prompt will ultimately influence the response. What would you consider an objective enough prompt that would result in an objective response?

What i type is a series of research tasks,
Asking to search all OSINT, social media, public media, gov publications and statements
UN and other organizations , think tanks
educational indtitutions etc. and more….

From what i am seeking in information the collection is searched, supporting documents, and analysis and answers provided.

Once you get a refined set of prompts it become quite easy
Back story searching and drilling truth has almost become childs play

I took a course on AI utilization and integration. This can be exciting and dangerous stuff we are playing with.

Objective questions directly relate to your inquiring mind. Skies , hmmm, stars maybe the limit.