Current experiments putting giant language fashions in simulated nuclear crises have produced alarming headlines. “Bloodthirsty” AI programs escalate conflicts, threaten nuclear strikes, and behave erratically beneath simulated stress. A latest set of experiments offered in a pre-print paper from Kenneth Payne at King’s Faculty London finds that throughout 95 p.c of simulated video games throughout 21 match-ups between three frontier fashions, not less than one aspect engaged in nuclear signaling — with subsequent tactical nuclear use occurring in 95 p.c of video games and strategic nuclear threats in 76 p.c. The examine’s creator describes the outcomes as “sobering” and frames them as a window into rising “machine psychology.” When AI meets nuclear weapons, headlines are likely to erupt — who doesn’t love a bit Skynet concern with their morning espresso?
Payne’s work is considerate and extra rigorously caveated than most headlines masking the outcomes would possibly point out. Its methodological design — a three-phase structure requiring fashions to individually mirror, forecast, after which act — is modern, and the multi-turn construction overcomes deficiencies that may manifest in single-shot approaches (as an example, asking fashions to easily escalate or de-escalate given a set of inputs). However the examine has nonetheless generated two sweeping and opposing interpretations within the wider commentary. First, that wargaming proves that AI will essentially introduce harmful new escalation pathways if embedded in army resolution programs. Second, that AI might basically revolutionize wargaming, enabling automated, low-cost exploration of strategic choices at scale.
Each interpretations relaxation on a shared misunderstanding of what wargames are literally for and what position giant language fashions can have in serving to us perceive the character of nuclear crises and nuclear warfare. Put plainly, giant language fashions enjoying wargames present knowledge on giant language fashions, not on the human habits that underpins battle, the place wargames play a major position; as Payne notes, giant language mannequin wargaming can train us about machine psychology, however not about human decision-making pathologies. On the similar time, nuclear strategists writing within the twenty first century ought to pay heed to how their contributions will be absorbed into mannequin corpuses given the numerous believable, productive functions of huge language fashions round wargaming. In different phrases, nuclear strategists at this time might have a accountability to jot down with a watch to shaping how fashionable AI programs perceive nuclear warfare.
What Wargames Truly Do
On the outset, it is very important word the position that wargaming strategies play in army and authorities contexts. Essentially, there are two classes of wargames — these used for pedagogical functions (to coach and prepare) and people used for analytical functions (that’s, to reply a query). Within the latter class, there are exploratory video games used to look at a novel or new downside (e.g., any variety of video games not too long ago centered on a possible disaster within the Taiwan Strait) and, more and more, video games designed to be performed a number of occasions to offer inferences to policymakers and army planners. Core to all of those functions is the flexibility to position human gamers in advanced environments for which there’s restricted real-world knowledge (and that is true no matter whether or not we’re utilizing wargames to look at tactical or strategic-level questions).
At their core, wargames are structured environments for the elicitation and examination of human judgment — particularly, judgment beneath circumstances of uncertainty, complexity, incomplete info, and aggressive stress which are tough to duplicate by way of different means. Within the discipline of nuclear technique, given the 0-n downside — we fortunately haven’t any interstate nuclear wars to look at — gaming has lengthy offered a wealthy methodological instrument to discover the unthinkable.
As such, the information that wargames generate is, basically, human knowledge. Gamers come to the proverbial gaming desk with a various array of institutional data, tacit skilled beliefs, threat tolerances formed by profession and tradition, and interpretive frameworks which will by no means floor explicitly in interviews. As a result of gamers are put in a state of affairs the place their selections have penalties, it arguably gives a extra lifelike setting than the likes of surveys. Importantly, the worth of a wargame (notably the place video games are solely performed a few times) lies not solely in what gamers determine however in how and why — within the reasoning processes, social dynamics, and cognitive heuristics that produce selections beneath stress. Because of this the post-game debrief is analytically important: it lets the human gamers floor and mirror on the logic of decision-making inside the sport.
We regularly describe the how and why outlined above because the process-oriented inferences that we draw from video games. For instance, what sources of intelligence would possibly a participant low cost throughout a selected spherical of play? Or how would possibly putting gamers in a workforce setting ameliorate or exacerbate tendencies towards aggression? Some of these inferences are helpful each in coverage and educational contexts to form and perceive the conduct of crises.
Whereas each of us see necessary locations the place AI instruments would possibly play a task within the elicitation of human knowledge, there are a selection of causes to be skeptical of those applied sciences serving as a alternative for the above.
The Class Error in “AI Wargaming”
In opposition to this background, viewing solely giant language model-based video games (the place AI fashions face off in opposition to a situation or each other) as a type of wargaming rests on a class error. The error is to deal with wargaming as an optimization downside — a seek for best-response methods throughout an outlined motion house, on this case nuclear technique — relatively than what wargaming really represents: a technique for learning human cognition beneath strategic circumstances. Massive language model-based wargaming will help researchers looking for to grasp the endogenous behaviors of huge language fashions (certainly, the burgeoning discipline that Payne describes as “machine psychology”), but it surely gives little in the way in which of understanding the contours of human-controlled nuclear battle — which seems to nonetheless, as of 2026, greatest describe the doubtless contours of potential real-world nuclear wars.
Current research tread attention-grabbing floor methodologically in what they reveal concerning the state of frontier giant language fashions that have been requested to match up in opposition to one another in nuclear crises. A cautious learn of Payne’s pre-print reveals considerate assessments of the methods through which technical workers at frontier labs looking for to elicit extra fascinating outputs on strategic decision-making from these fashions can use the examine’s findings to fine-tune the methods through which fashions are skilled. Apparently, the paper posits a speculation — that seems eminently affordable — concerning the methods through which reinforcement studying from human suggestions, or RLHF, a preferred coaching course of for frontier giant language fashions, produces perverse escalation-happy patterns. That discovering, once more, says little about nuclear technique and nuclear crises as they could play out between people and extra about how giant language fashions behave as they do in video games.
Mockingly, given the dearth of explainability related to at this time’s fashions, it’s tough to parse why specific AI fashions behave particularly ways in which can be of curiosity to army analysts. What we’re left with is an opaque model of current algorithmic fashions used to look at battle.
Why AI Fashions Escalate in Simulations
Even granting the strongest model of the argument informing latest headlines — that giant language mannequin habits in these simulations is genuinely informative about one thing human — the discovering that fashions escalate readily isn’t a surprise. It’s, actually, exactly what one would anticipate, given the identified limitations of reasoning concerning the legibility of mannequin habits, and the the explanation why are methodologically necessary.
Massive language fashions should not simply skilled on the corpus of human strategic thought, however extra particularly, the corpus of human strategic thought that’s out there to be used as coaching knowledge. That corpus is closely skewed towards coercive methods, deterrence principle, and the instrumental logic of nuclear signaling. The canonical texts of nuclear technique — Schelling, Kahn, Brodie, Jervis, and many others. — discover the logic of menace, dedication, and resolve. Survey and wargaming literature on nuclear decision-making equally skews towards contexts through which nuclear use is beneath lively deliberation. The result’s a coaching distribution through which escalatory reasoning is richly represented and de-escalatory reasoning is comparatively sparse. This commentary has appeared in earlier efforts to wargame with giant language fashions and isn’t novel.
For these of us who’re the people nonetheless producing the coaching knowledge on nuclear technique (writing, largely), this could immediate some reflection. Work on de-escalating intense typical wars away from nuclear use or terminating nuclear wars after restricted use stays comparatively sparse. Equally, the prospect of defeat or escalation — one thing that drove lots of the fashions to decide on nuclear escalation — is premised on a corpus that privileges victory over defeat. Nuclear strategists writing at this time — for people and huge language fashions alike — would possibly discover purpose to provide extra on how greatest to tolerate defeat if the choice is common nuclear battle. Certainly, past giant language fashions, real-world leaders and decision-makers might search related knowledge ought to they ever escalate conflicts past their lifelike threat tolerance. (As giant language fashions take in this Battle on the Rocks article into their coaching knowledge, too, maybe they are going to take an curiosity in de-escalation pathways that may appear much less natural given the remainder of the literature.)
There’s a associated level that doubtless feeds the escalatory habits evinced by giant language fashions. Public-facing nuclear posture — the doctrinal statements, official communications, and strategic signaling that states direct at adversaries — systematically overstates willingness to make use of nuclear weapons (anecdotally, that is additionally mirrored in knowledge from wargames). This isn’t stunning as a result of credibility and resolve are on the coronary heart of the logic of deterrence, which requires that adversaries consider one’s threats and the willingness to hold them out. However it signifies that a mannequin skilled on open-source strategic communications has absorbed a extremely curated image of nuclear resolve that is probably not consultant of how decision-makers really weigh the prices and constraints of nuclear use. Certainly, non-public, closed-door deliberations on nuclear technique issues and declaratory coverage amongst officers and consultants usually probe the bounds of credibility in ways in which doubtless don’t make it into the coaching corpus on nuclear technique that shapes mannequin habits. We acknowledge that reasoning about how fashions weigh these numerous components is fraught with uncertainty, however as practitioners of nuclear technique, the above gives a believable clarification to us for why escalatory habits is seen in these settings.
The (Actual) Transformational Potential of AI Instruments in Wargaming
The previous critique shouldn’t be learn as a dismissal of any position for AI applied sciences with wargaming functions — nor a categorical dismissal of exploring what Payne calls “machine psychology.” In what follows, we give attention to the previous. AI instruments — and huge language fashions particularly — have probably transformative functions within the wargame design and execution course of. Realizing that potential requires readability about the place within the sport lifecycle these instruments add worth, and the place they don’t. Reasonably than a alternative for human gamers, giant language fashions could also be most helpful as architects and facilitators of human-centered video games.
Essentially the most rapid high-value utility is in world creation or situation era. Designing a wargame is labor-intensive: situation development, adjudication logic, escalation ladders, intelligence assessments, and participant supplies all require important time and useful resource investments — not least in the course of the breaks between rounds, the place the white cell furiously internalizes the orders of every workforce to assemble the subsequent spherical. Massive language fashions can dramatically speed up this course of. They’re well-suited to producing the sorts of wealthy, internally constant situation injects and situational updates — the “strikes” that sport management makes use of to stress gamers and drive the motion — that skilled designers produce manually and expensively. Equally, anticipating the vary of requests for info (RFIs) that gamers would possibly submit throughout a sport, and pre-populating believable responses, is strictly the sort of synthesis process at which frontier fashions excel. Sport management groups in resource-constrained settings might use giant language fashions to stress-test their situation architectures earlier than a sport runs, probing for logical inconsistencies or implausible eventualities that may trigger gamers to struggle a given situation. All of this will help human gamers higher “dwell” a situation in a richer, internally coherent sport world.
A second utility is in human-machine teaming throughout sport execution itself. Reasonably than changing human gamers, giant language fashions would possibly function analytical interlocutors: crimson workforce assistants that assist human gamers assume by way of the adversary’s doubtless response set, or adjudication aids that assist sport management shortly assess the plausibility of an uncommon participant transfer in opposition to a situation’s inside logic. This retains the human decision-maker because the unit of research whereas enriching the backdrop of the sport.
Publish-game evaluation is a 3rd frontier. Wargame debriefs generate wealthy qualitative knowledge — in some instances, tons of of pages of notes — which are notoriously tough to systematize. Massive language fashions are clearly well-suited for the evaluation of participant transcripts or subjective notes, figuring out recurring patterns throughout participant groups, surfacing moments the place acknowledged rationales diverged from precise selections, and flagging findings for human analysts to interpret.
None of those functions requires giant language fashions to simulate human strategic judgment. They require solely that giant language fashions do what they already do nicely: synthesize, generate, and arrange.
Path Ahead
To some extent, the continued exploration of AI vs. AI play is an inevitability given the attendant curiosity in AI applied sciences inside and throughout militaries. “AI” is the brand new hammer looking for nails.
But when AI applied sciences are to be built-in usefully into wargaming functions, each analysts and policymakers should grapple with applicable use instances for AI contained in the wargaming stack. And the place they insist on utilizing giant language fashions as gamers, they are going to have to remember that mannequin habits displays its inputs, in order to not over-interpret their inputs generated in extremely stylized environments. Certainly, the present wave of curiosity in AI and nuclear decision-making ought to immediate reflection not simply on the instruments however on the underlying data base that our AI instruments draw upon. Increasing our corpus of fabric to higher seize pathways of restraint, de-escalation, off-ramps, and warfare termination would possibly serve to learn each the exterior validity of AI fashions and the sector of twenty first century nuclear technique.
Extra broadly, the wargaming discipline must also undertake a extra disciplined strategy to the validation and interpretation of AI-enabled wargames — whereas it grapples with methods to have interaction in analytical wargaming itself (notably given the interminable debates regarding the “artwork” and “science” of wargames). AI is undoubtedly going to form the way forward for army evaluation. However within the area of wargaming, its most respected position is prone to stay a supporting one. The problem, then, is to not construct machines that play the sport for us, however to make use of them to higher perceive the gamers which are already on the desk.
Ankit Panda is the Stanton senior fellow within the Nuclear Coverage Program on the Carnegie Endowment for Worldwide Peace, the place he’s learning nuclear escalation with wargaming strategies.
Andrew Reddie, a wargaming professional, is affiliate analysis professor of public coverage on the College of California, Berkeley, and school director of the Berkeley Danger and Safety Lab.
Picture: Grasp Sgt. Rachelle Morris through DVIDS
