AI for Escalation Risk: From Chicken to Bumper Cars

A strange thing happened in recent Middle East escalation cycles: hundreds of drones and missiles were launched, and almost nothing was physically destroyed. No cratered runways. No burned-out refineries. In some cases, no fatalities. Yet the political messaging was loud, the regional anxiety spiked, and decision-makers still had to decide—fast—whether they were watching the start of a major war or a carefully staged signal.

That gap between spectacle and damage is the point. Joshua Tallis describes a new rung on the escalation ladder where states can “act like they’re at war without paying the usual price.” He frames it as bumper cars (limited expected damage, more testing and signaling) versus the classic deterrence model of chicken (high risk, high stakes, catastrophic downside).

For the AI in Defense & National Security series, here’s why this matters: the hardest part of escalation management is not intercepting weapons—it’s interpreting intent under uncertainty. AI won’t “solve” deterrence, but it can help leaders avoid the worst failure mode: misreading bumper cars as chicken (or vice versa) and responding in a way that creates the war everyone claimed to be avoiding.

Performative aggression is real—and it changes the math

Performative aggression is the use of real military force designed primarily to signal, not to destroy. The operational fingerprint is counterintuitive: attacks that are large enough to be seen, but shaped to reduce the odds of casualties and critical infrastructure loss.

In Tallis’s examples, Iran’s strike patterns illustrate why “same actors, similar weaponry” can still produce wildly different outcomes:

• April 2024: a heterogeneous salvo where slow one-way attack drones arrived first, effectively creating time—for warning, for interception, and for political signaling.
• June 2025 (Al Udeid): a smaller ballistic missile salvo with limited effects.
• October 2024 and a later 12-day war: higher lethality and infrastructure damage, with civilian deaths and critical sites hit.

These aren’t just different attack packages; they’re different games.

Chicken vs. bumper cars: a practical definition

Chicken is about credible risk: both sides understand catastrophe is possible, and deterrence hinges on who’s willing to ride closer to the edge.

Bumper cars is about managed risk: at least one side expects defenses and signaling to keep damage low, so probing becomes more attractive.

The operational problem for defenders is blunt: you often don’t know which game you’re in until after the intercepts—and sometimes not even then. That ambiguity is exploitable.

The core danger is information asymmetry, not missile leakage

A missile that “leaks” through air defenses is scary. But strategically, the more persistent danger is information asymmetry:

• The attacker may know it’s playing bumper cars (telegraphing, pacing, targeting choices).
• The defender may fear it’s chicken (worst-case assumptions, public pressure, alliance politics).

That mismatch creates a pathway to accidental escalation.

Here’s the uncomfortable reality: modern integrated air and missile defenses can create a deterrence paradox. When defense performs extremely well, leaders may feel safe—yet adversaries may feel newly empowered to conduct symbolic strikes because they expect interception to keep the “price” low.

This matters because political leaders rarely get rewarded for restraint after absorbing a visible attack, even a low-damage one. They’re judged on:

• Perceived resolve
• Alliance credibility
• Domestic expectations
• Whether deterrence “held”

So a defender can win tactically (intercepts) and still lose strategically (coercion, narrative disadvantage, or forced overreaction).

Where AI helps: clarifying which game you’re in

AI is most valuable here as a decision support system for ambiguity—helping commanders and civilian leaders classify events, test hypotheses, and anticipate second-order effects. It’s not about replacing judgment; it’s about reducing the probability of a catastrophic misread.

1) AI-enabled intent modeling from observable “signatures”

You can’t read intent directly, but you can model behavioral signatures. Performative aggression often correlates with choices that maximize visibility while minimizing harm:

• Pacing and sequencing: slow drones first, fast missiles later; extended timelines that function as warning.
• Target selection: symbolic military sites, low-density hours, or areas where collateral damage is easier to avoid.
• Salvo composition: mixes designed to be intercepted (or at least expected to be) rather than optimized for saturation.
• Messaging behavior: coordinated narratives timed to the strike, prepared off-ramps, and statements aimed at domestic/proxy audiences.

An AI model doesn’t need to “know motives.” It can produce something operationally useful: a probability-weighted assessment of which escalation game is most consistent with the observed package.

Snippet-worthy: The key output isn’t “what do they want?” It’s “what’s the likelihood this strike was engineered to be survivable—and what response options stay stable across both interpretations?”

2) Scenario generation for miscalculation pathways

The fastest way escalation spirals is through rare but plausible branches:

• A single warhead leaks through and causes mass casualties.
• A defender hits the wrong retaliatory target, triggering alliance dynamics.
• A third party (proxy, militia, or opportunist state) acts during the confusion.

Generative AI and simulation tools can produce structured what-if trees that planners can stress-test ahead of time:

• What happens if interception is 99% vs. 90%?
• What happens if the public believes it was a major attack even when damage is minimal?
• What happens if the attacker intended bumper cars but the defender announces automatic escalation?

This is where AI improves speed and coverage. Humans are great at a few scenarios; AI can help teams examine dozens, then focus human judgment on the ones that matter.

3) Decision advantage through proportionate response planning

Performative aggression thrives when defenders feel trapped between two bad options:

• Do nothing and look weak
• Hit hard and appear escalatory or disproportionate

AI-assisted planning can help identify response bundles that preserve deterrence without creating unnecessary escalation pressure, such as:

• Time-bounded, reversible actions (temporary posture shifts, controlled deployments)
• Non-kinetic responses tied to attribution confidence
• Narrowly scoped kinetic options with explicit deconfliction pathways

The point isn’t “restraint at any cost.” It’s maintaining control of the interaction—and avoiding the attacker’s goal of using your own proportionality norms against you.

Deterrence in a repeated game: AI can track the pattern, not just the event

A single performative strike is one thing. A series of them becomes a pattern—and the deterrence logic shifts.

Tallis flags a crucial framing: is this a one-shot or a repeated game? In repeated games, the attacker may:

• Gain a low-cost signaling tool that doesn’t risk regime stability
• Exploit narrative control at home
• Accept interceptions as “proof” of facing a coalition, not as failure

Meanwhile, the defender may face the opposite:

• Higher public expectations (“Why didn’t we prevent this entirely?”)
• Less narrative control
• More visible restraint costs

AI is well-suited to repeated-game analysis because it can maintain a living model of behavior across time, updating estimates of:

• Thresholds that trigger retaliation
• The attacker’s tolerance for escalation
• Whether “bumper cars” is turning into staged chicken—or genuine chicken

A practical metric set leaders can demand

If you’re advising a command, a ministry, or an alliance cell, ask for metrics that tie air defense outcomes to escalation risk—not just intercept rates.

A useful dashboard often includes:

1. Time-to-impact distribution (how much warning the attack design creates)
2. Target class risk score (civilian density, critical infrastructure proximity)
3. Salvo saturation intent estimate (does the package look optimized to overwhelm or to be seen?)
4. Narrative synchronization index (timing of official statements, proxy messaging, and media amplification)
5. Escalation elasticity (how much response force shifts adversary behavior in the next iteration)

These aren’t magic numbers. They’re a way to keep teams from debating vibes when minutes matter.

Guardrails: AI can reduce miscalculation, but it can also amplify it

AI introduces its own risks in escalation environments. Most organizations underinvest in these until something breaks.

Where AI can go wrong

• Model overconfidence: probabilistic outputs get treated as certainty.
• Training data bias: the model learns yesterday’s escalation patterns, not tomorrow’s adaptations.
• Adversarial manipulation: attackers shape signals—timing, targets, decoys—to push the model toward a desired classification.
• Automation pressure: leadership relies on AI because decision tempo is high, not because the AI is correct.

Three safeguards that actually work

1. Calibrated uncertainty reporting: require confidence intervals and “what would change my mind” indicators.
2. Red-team inputs: dedicated cells that try to spoof your classifier and break assumptions.
3. Human-in-the-loop thresholds: AI can propose; humans must authorize actions that cross political or kinetic escalation lines.

Snippet-worthy: In deterrence, the costliest bug isn’t a wrong forecast. It’s a forecast that makes leaders feel certain when they shouldn’t.

What policy teams should do now (before the next “bumper cars” strike)

If you’re building capability—whether in a defense ministry, a combatant command, or an allied planning staff—focus on preparations that are robust across both games.

1. Pre-approve response options that are proportionate to intent, not just impact. Waiting for physical damage as the only trigger invites performative coercion.
2. Invest in AI-enabled fusion for intent signals. Blend air defense telemetry with messaging intelligence, diplomatic signals, and proxy activity.
3. Practice “ambiguity drills.” Run exercises where the same strike can plausibly be chicken or bumper cars, and measure decision stability under uncertainty.
4. Treat integrated air and missile defense as a political system, not only a technical one. Intercepts shape narratives, alliance perceptions, and escalation incentives.

These steps don’t remove ambiguity. They keep ambiguity from owning you.

Where this goes in 2026: symbolic fighting will scale faster than defenses

Cheap drones, improving guidance, and proliferating strike packages mean the performance of air defenses will keep being tested in public. That’s not a tactical footnote; it’s a strategic environment.

For the AI in Defense & National Security series, the lesson is clear: AI’s most immediate value is helping leaders interpret and respond to attacks designed to be intercepted. That’s the bumper-cars rung—real violence used as theater, with just enough risk to be dangerous.

If your organization is exploring AI for intelligence analysis, mission planning, or escalation modeling, build around the central question modern deterrence keeps asking: Are we seeing capability, intent, or performance? The next crisis will punish teams that can’t tell the difference.