The Quiet Collapse of the War Threshold

For most of modern history, war was understood as an event. It required declaration, mobilization, and intent. Even when those formalities weakened, the underlying assumption remained intact: war began when human actors crossed a visible line and accepted responsibility for what followed.

That assumption no longer reliably holds.

The most consequential shift of the last decade is not the speed of weapons, the reach of surveillance, or even the sophistication of artificial intelligence. It is the erosion of the lines that once separated preparation from action, influence from force, and automation from authority. Conflict has not become louder. It has become easier to miss.

This change is rarely announced. It settles in.

---

Thresholds as Governance, Not Ceremony

Thresholds were never symbolic gestures. They functioned as governance tools. Their purpose was to slow action, concentrate accountability, and force deliberation at moments when consequences escalated faster than correction.

A declaration of war, for example, went beyond rhetoric. It triggered alignment across institutions, required public acknowledgment of risk, and bound leaders to decisions that could not be quietly undone. Even covert action operated within escalation ladders that, while imperfect, created pauses where responsibility had to surface.¹

Those ladders relied on friction. Time delays. Human review. Institutional hesitation.

Over time, many of these frictions were reframed as inefficiencies. Restraint came to be treated as delay, and delay as failure. This shift did not happen all at once, which is why it proved so difficult to contest.

---

Action Without Intent

AI-enabled systems are exceptionally good at pattern completion, optimization, and response. They now sit upstream of many domains that shape perception, prioritization, and reaction well before any explicit human decision occurs. Recommendation engines influence sentiment. Automated systems triage information. Predictive tools rank risks, targets, and responses.²

None of these systems intend escalation. That absence of intent is central to the risk.

When systems act without intent, responsibility spreads thin. When responsibility spreads thin, boundaries blur. Decisions still happen, but they no longer feel like decisions. They appear instead as defaults, outputs, or background conditions that are difficult to challenge in real time.

This pattern is already visible in areas such as automated trading, predictive policing tools, large-scale content moderation systems, and cyber operations, where response speeds routinely outpace escalation review. Action moves earlier in the chain, while responsibility remains anchored to later moments that may never fully arrive.³

By the time humans intervene, the terrain has already shifted.

---

Habituation as the Principal Risk Vector

Public discussion of AI risk often focuses on misuse, bad actors, or dramatic failure. Those scenarios matter, but they are not the most common way systems fail. The more persistent risk is habituation.

When automated influence becomes routine, it stops being debated. When surveillance becomes ambient, it fades from notice. When escalation unfolds through a sequence of small optimizations, no single step looks disruptive enough to halt.

This is how boundaries fade: not through violation, but through repetition.

Institutions struggle here. Most are designed to respond to discrete events, not to slow, distributed change. Governance mechanisms look for spikes or breaches. AI systems tend to produce gradients. The result is a gap in which systems quietly shape outcomes that institutions can only explain afterward.⁴

---

War Without a Clear Line

The term “gray-zone conflict” is often used to describe this space. It is useful, but incomplete.

What is emerging is not conflict below a defined threshold. It is the weakening of the threshold itself. Influence operations, economic pressure, cyber intrusion, automated targeting, and narrative manipulation now operate continuously, mediated by systems optimized for engagement, efficiency, or advantage rather than restraint.⁵

In this environment, escalation does not feel like escalation. It feels like normal operation.

Restraint begins to look anomalous. Pressure becomes ambient. Systems that pause appear inefficient, and actors who hesitate appear uncompetitive. The war threshold did not vanish because it was dismantled. It faded because no one was explicitly responsible for sustaining it.

---

How AI Accelerates the Drift

AI does not create this condition on its own, but it accelerates it.

First, it compresses time. Feedback loops tighten, and responses propagate faster than meaningful human review can reliably keep pace.

Second, it abstracts agency. Outcomes are framed as system behavior rather than human choice, even when humans defined the objectives and constraints.

Third, it scales influence unevenly. A small number of actors can now shape large populations indirectly, without triggering traditional markers of escalation or conflict.⁶

Taken together, these effects push power upstream while leaving oversight downstream. The central social risk is not that AI will decide to wage war. It is that it will quietly remove the moments when humans realize escalation is already underway.

Alignment is not a property of a model. It is a property of a system.

---

Frontier AI and Responsibility Upstream

For organizations developing and deploying frontier AI systems, this shift places responsibility earlier than many governance models anticipate. As models increasingly mediate perception, prioritization, and response, social risk emerges not only from explicit misuse, but from how capability reshapes what feels normal, acceptable, or inevitable.

The challenge is no longer limited to preventing obvious harm. It is preserving the lines that signal when influence becomes authority and optimization becomes escalation. That work cannot rely solely on model performance metrics. It demands attention to system behavior, deployment context, and how repeated exposure changes expectations over time.

In practice, this means treating habituation itself as something to be monitored and stress-tested, especially where long-term trust and institutional credibility are at stake.⁷

---

Why Traditional Safety Framing Misses the Point

Much AI safety work focuses, appropriately, on alignment, robustness, and misuse prevention. These approaches are necessary, but they assume failures that arrive as discrete events.

Boundary erosion does not work that way. It accumulates.

Controls designed to catch misuse will not surface habituation. Logs will not flag drift. Accountability applied after the fact cannot restore a line that was never crossed in a way anyone recognized.⁸

What is needed is a posture that treats boundaries themselves as design objects, not as background assumptions.

---

Reintroducing Friction Deliberately

The answer is not to halt progress or to idealize human decision-making. Humans are inconsistent and biased. They are also capable of hesitation, dissent, and second thought.

In high-consequence systems, friction is not a defect. Delay creates space for review. Review concentrates responsibility. Explicit checkpoints force intent to surface before action propagates.

Building systems that preserve boundaries means intentionally creating moments where action slows, responsibility becomes visible, and escalation can be recognized while it is still reversible. This is not about control. It is about legibility.⁹

---

Governance Before the Breaking Point

The most consequential failures do not announce themselves. They arrive quietly, wrapped in efficiency gains and normalized practice. By the time outcomes look catastrophic, institutions have already been trained to live with them.

The erosion of the war threshold is not a future problem. It is already in motion. The open question is whether institutions will recognize that boundaries require maintenance before they are gone.

Social risk analysis must therefore move earlier. It must ask not only what systems can do, but what they make routine. Not only who is accountable, but when accountability becomes difficult to locate.

The work ahead is not dramatic. It is careful and often uncelebrated. It consists of noticing what no longer feels uncomfortable, and asking why.

That is where governance begins.

---

Endnotes

1. Schelling, T. C., Arms and Influence. Yale University Press.
2. O’Neil, C., Weapons of Math Destruction. Crown Publishing.
3. Zuboff, S., The Age of Surveillance Capitalism. PublicAffairs.
4. Perrow, C., Normal Accidents. Princeton University Press.
5. Mazarr, M. J. et al., Understanding the Emerging Era of Gray Zone Conflict. RAND Corporation.
6. Brundage, M. et al., “The Malicious Use of Artificial Intelligence.” Oxford / CSER.
7. Floridi, L. et al., “AI4People—An Ethical Framework.” Minds and Machines.
8. Nissenbaum, H., Privacy in Context. Stanford University Press.
9. Weick, K. E. & Sutcliffe, K. M., Managing the Unexpected. Wiley.