This essay offers a few thoughts about the changes that might occur in doctrine and military organization as the United States adapts its nuclear weapons policies to the age of Artificial Intelligence (AI).  On this, however, I’d refer you first and foremost to the essay of mine that INHR published in March of this year, which goes into these matters in more detail than I have time to recount now. 

AI and Nuclear Weapons Decisions

There, I argued that of all the aspects of militarily-relevant decision-making that could potentially come to involve AI tools, the actual decision to fire a nuclear weapon is surely the least likely decision to be offloaded onto a computer, no matter how “well” the machine might arguably be able do it.  In this respect, however, I did note two possible exceptions: one example that is already real, and one that I merely hypothesized. 

As for the real one, Russia has apparently had an automated nuclear launch system in place for many years, in the form of its so-called “Dead Hand” or “Perimeter” system, intended for contingencies in which the high command in Moscow seems to have been eliminated.  That’s not new – and may not really be “AI” in the sense we normally mean that term – but it does represent a counterpoint to my intuition that for the most part nuclear launch decisions are a distinctive prerogative of a country’s human leadership that will not be delegated to machines.  Nevertheless, the Dead Hand example is a Russian one, and I think my intuition still holds for the United States.

But my second, if hypothetical, example might be more tempting to American commanders.  It relates to the sort of all-out nuclear warfighting that could conceivably occur if deterrence failed and a general nuclear exchange began.  In those circumstances, a U.S. shift into “damage limitation” operations intended to destroy as many adversary nuclear forces as possible – and as quickly as possible – might create incentives to delegate both targeting and engagement decisions to AI tools so as to maximize the destruction of adversary nuclear assets before they were able to launch their warheads at the American homeland.

In such circumstances – where every target not immediately engaged could potentially entail the death of an entire American city and its population – commanders might find decisional outsourcing to hyperfast AI decisional systems an irresistibly attractive option.  To be sure, in such a scenario, the limiting factor would probably not be the speed of the decision process but rather the challenge of physically getting warheads from launchpoint to target; it might thus be that beyond a certain point, AI tools wouldn’t make much difference.  Nevertheless, at least in theory, a “damage limitation” scenario at the end of the world, as it were, is one in which I could imagine a case being made for full nuclear decisional offloading to AI.

I did speculate in my essay, however, that AI would come to be of increasing utility, and would increasingly be used, in decision-support systems for human commanders in charge of nuclear forces.  While I am acutely aware of the sort of “offloading creep” that this could potentially create – in which humans come to rely upon AI recommendations in an increasingly reflexive way, eventually simply rubber stamping machine outputs with little critical thought – I still think that some use of AI-based decision-support tools could probably be of benefit in making human targeting decisions better informed, more accurate, and (dare I say it) wiser.

In one particular nuclear weapons application, in fact, I contended in my essay that AI might make a very significant improvement upon the status quo.  Rather than simply presenting the U.S. president with a limited and rigid list of pre-prepared nuclear launch packages in some crisis in which decisions have to be made in mere handfuls of minutes, for instance, AI tools might allow the president quickly to generate new, alternative, “bespoke” nuclear options in response to the idiosyncrasies of the scenario that has generated the crisis – that is, giving him the chance to fire only at what he wants to hit rather than just choosing from a very limited menu of options that the Strategic Command in Omaha had prepared for him.  Using AI to let the president customize strike packages based upon the particular circumstances of the moment, it seemed to me, would be a significant improvement over the rigidities of the current approach, allowing more scope for human judgment and meaningful strategy, and better serving the Clausewitzian goal of tailoring military operations to serve the purpose of a conflict.

Humans and “Commander’s Intent” in AI Warfighting

I stand by those points I made in that earlier INHR essay.  Beyond those suggestions, however – and I should stress that my observations here reflect only my personal views, and don’t necessarily represent those of anyone else – I’ll offer just a couple of quick additional comments.

The recent U.S. war against Iran illustrates further the ways in which militaries are finding it increasingly attractive to turn to AI to assist with speeding up target acquisition and engagement.  In my INHR essay, I mentioned reports of Israeli use of AI in the campaign in Gaza, but the Iran conflict suggests that the United States has also already incorporated AI tools into its decision-making processes to increase the speed and scale at which conventional weaponry can be brought on target.  Cameron Stanley, the Pentagon’s “chief digital and artificial intelligence officer,” told the press that in the recent Iran campaign, the Grok chatbot “enabled U.S. forces to deploy over 2,000 munitions to 2,000 distinct targets within 96 hours,” a degree of “operational efficiency” that was apparently essentially impossible using previous methods.

The Israeli and U.S. examples suggest that the temptation to employ AI tools in support of conventional warfighting will remain strong.  To the degree that AI continues to provide the sort of increases in operational efficiency of which Stanley boasted, therefore, it’s hard to imagine any military turning down such an opportunity. 

As I noted above with respect to nuclear “damage limitation” scenarios, of course, there is probably a limit on howmuch increased efficiency AI can provide, since at some point the laws of physics will limit how fast you can actually get munitions to aimpoints even if you had godlike battlespace omniscience.  Faster engagement, of course, presumably also means running out of munitions faster. 

Nevertheless, for many applications, particularly in kinetic campaigns aimed at crippling an adversary force and detecting and eliminating fleeting mobile targets as quickly as possible, the advantages apparently offered by AI-assisted targeting in terms of speed and sale are likely to prove hard to resist.  And to the degree that this occurs in warfighting more generally, there may over time be some bleeding over into nuclear issues – especially to the degree that reliance upon AI becomes normalized and “comfortable” for commanders and wartime leaders in a general sense, perhaps making the prospect of AI use in nuclear ones seem less outrageous and frightening.

And this may be especially the case in a future “near-peer” conflict such as a U.S. war with China or Russia, where the number and diversity of assets involved, the sweeping geographies at issue, and the volumes of fire likely to occur (at least at the outset, before magazines start to run dry), are likely to present particularly grave challenges of command and control for the unaided human mind.  The perceived payoff for using AI is thus likely to be high, and use is likely to continue.

Sovereign vs. Private AI

Another interesting issue that has cropped up in the United States is that – after a period of rah-rah “go faster” approaches to the untrammeled development and adoption of AI – the current U.S. Administration is beginning to regulate AI, or at least considering doing so.  As you’ll have noticed, of course – including in the last couple of days – we in Washington swing back and forth somewhat capriciously between imposing vetting requirements and export control restrictions on AI-related technologies and frontier models and then relaxing them.  Nevertheless, the issue of regulation has shifted for the first time from being “out of the question” to being a serious topic for the Second Trump Administration.

It’s not clear how relevant this is to the nuclear issue, of course.  Moreover, the debate in Washington right now isn’t about whether to limit AI development, but rather simply one about who calls the shots in AI development and access to its fruits: the government as a national security prerogative, or the technology firms who are leading the charge, technically, in building cutting-edge systems.  (There still seems to be no discussion of government officials trying to slow down the pace of the technology companies’ headlong race to build AI that is as powerful as possible as quickly as possible.)  Where all this goes is still unclear, but this is a major shift, and could conceivably at some point have implications in the nuclear sphere.

Debating the “Human in the Loop”

Let me conclude with a word about military uses of AI more generally, in ways that are not necessarily – or at least not directly – relevant to nuclear weaponry, but which could nonetheless come to affect the nuclear arena over time. 

            Bot Competence vs. Human Competence

In the AI-related national security policy community, interesting and spirited debates are raging about how far the trends of increased AI competency and adaption can, or should, continue.  On the one hand, I have heard thoughtful and knowledgeable people argue that however good AI might be at accomplishing ever more sophisticated tasks at speed and scale, there are nonetheless certain things – including aspects of wartime command – that machines neither could nor should ever do: things such as defining the objectives of a conflict, deciding when to change objectives, determining how much cost or pain to bear in support of such goals, and deciding when it is appropriate to stop fighting.  There will thus always be a role for humans here, it is averred, that even the best AI will never be able to supersede.  Some things, this argument goes, are the inherent prerogatives of – and within the unique competencies – of the collectivity that is a human society, and are simply not automatable.

On the other hand, I have also heard strong arguments made that we will eventually inevitably offload all warfighting decisions to AI bots simply because the machines will eventually prove so much better at it that leaders will feel themselves to have no choice.  In every prior instance in which it has been predicted that human intelligence will remain superior to machine intelligence, this argument goes – and here one might think of the “Turing Test,” chess, the game of Go, certain types of creatively advanced mathematics, composing decent music, writing passable fiction, and even accomplishing tasks related to persuading individual humans to change their minds about something – machines have initially failed to outperform humans but soon thereafter generally learned to beat us quite reliably. 

When such outperformance arrives before long with wartime command as well – which, it is contended, is all but certain – we will recognize that we have no choice but to offload decision to the bots.  Especially in a fight with a near-peer adversary such as China, it is thus argued, this will be a no-brainer: letting AI take over the fight will be our “least bad option,” because not letting it do so is “deciding to lose.”  Thus are the intellectual lines being drawn.

I won’t presume to try to arbitrate between such positions here, but I mention these debates because I don’t think a clear position has yet developed in the United States on these matters.  It may be that AI tools are not yet good enough to present a tempting comprehensive alternative, but the trends certainly seem to be in favor of increased AI usage – at least in warfighting with conventional weaponry, at any rate, and presumably also including the cyberspace and electronic warfare domains, where machine advantages of speed and scale over mere humans are likely to be quite pronounced as well. 

Humans and “Commanders Intent”

How would I myself draw the line, if I could?  Well, for my part, given the trends of increasing AI usage, I wonder whether our debates about whether to have humans “in the loop,” “on the loop,” or “out of the loop” of individual target engagement decisions need to be supplemented by a further discussion about whether and how one might code fidelity to human “commanders intent” into warfighting that is otherwise potentially entirely automated.  Specifically, I think we may need to pay more attention to “human on the loop” architectures not at the “retail” level of killing but rather at the “wholesale” level of campaigning – that is, at the classically “commanders’” level of strategy, objectives, permissible operational behaviors, and rules of engagement.

I would like to see AI technicians working more, for instance, to build a sound architecture through which specified authorized human decision-makers could give “mission orders” to operational assets that themselves may – or may not, as circumstances dictate – actually be managed by AI in implementing such orders within whatever parameters are set by the humans.  Is it possible, moreover, to develop a way for the human decision-makers at the center of such a system to retain sufficient battlespace situational awareness and get sufficient feedback from the AI tools that are actually “fighting the fight” that those humans can reliably understand how and the degree to which their “commander’s intent” is actually being followed – insight that will allow them to inject new parameters into the system on an ongoing basis to adjust and calibrate operations in ways that appropriately support their overall goals? 

If there is some arena that seems inherently “human” and that should be kept out of the hands of the machines, it is presumably at just such a level of value-shaped judgments in campaigning in support of objectives established by the human collectivity of a society at war.  Even if the partisans of automation are right that machines will hopelessly outclass us in all matters related to handling speed, scale, and complexity in the modern battlespace, there would seem to be good reason to reserve normative parameter-setting for human intelligences.  If we can figure out a way to let the machines make us better at fighting while reserving for ourselves the process of setting the agenda and parameters for the fight, that might not be a bad balance.

-- Christopher Ford