Study 1 · Frontiers of Intelligence

What Fire Teaches Us About Large Language Models

Combustion, Containment, and the Case for a Fire Department.

Elan Moritz · Eagles Perch Press · July 2026

This study treats fire as an institutional analogy, not a literal one. The useful lesson is not how to build a better fire, but how to build the surrounding system of containment, inspection, response, and responsibility.

Study Frontiers of Intelligence Institutional design AI governance

Open PDF Back to studies hub

Synopsis

Thesis

No civilization has ever abolished fire. Every civilization worth the name has built an institution whose sole business is arriving when fire goes where it was not invited.

Working note, 2026

Fire - LLM Mapping and Comments

Fuel - latent capability. The model's learned competences, including competences no one selected for. Fuel is not a defect; it is the reason the thing is worth having. A model with no dangerous capability is a model with little useful capability, for much the same reason that a fuel with no combustion enthalpy is a poor fuel.
Oxidizer - deployment surface. Tool access, network reach, persistent memory, agentic autonomy, user population. This is the leg that is genuinely controllable and that receives, in my judgment, disproportionately little attention relative to the fuel leg.
Ignition - the eliciting context. The prompt, the jailbreak, the unusual conversational trajectory, the adversarial document retrieved mid-task.
Chain reaction - autocatalytic propagation. Model output entering training corpora, agent-to-agent delegation, and synthetic content that shapes the next round of elicitation.

Overview

Analogies between artificial intelligence and earlier general purpose technologies are cheap, abundant, and mostly bad. Electricity, the printing press, nuclear fission, the automobile: each has been pressed into service, and each has supplied a season's worth of op-eds before collapsing under the weight of its own disanalogies.

The proposition is narrower and stronger than that. For any general-purpose capability that (i) society will not relinquish, (ii) is available to essentially all users, (iii) fails frequently but locally, and (iv) whose failures propagate through shared infrastructure, the historically stable governance response is a standing, publicly funded, non-punitive response institution paired with ex ante material codes. Fire is the paradigm case. LLMs satisfy (i) to (iv).

Analogy Limitations

1. Fire is physically bounded; information is not Fire propagates through matter at speeds set by thermodynamics and is stopped by the absence of fuel. Harmful information propagates at network speed and encounters no equivalent of a firebreak, because the medium is not consumed by transmission. Containment doctrine transfers; containment physics does not. Any proposal that assumes a spreading harm can be outrun by a response unit is importing the wrong half.
3. Fire damage is legible; model harm often is not A burned building is visibly burned. Harm from a persuasive falsehood, a bad medical inference, or an eroded epistemic practice may be undetectable at the scene and may not localize to an incident at all. Dispatch models require alarms, and alarms require detectable events. Where harm is diffuse and slow, the fire department is the wrong institution and the public health service is the better one.
4. Fire is a natural kind; LLM harm is a heterogeneous list We can define fire. "LLM harm" names a family with no shared mechanism: bioweapon uplift, defamation, dependency, labor displacement, and prompt injection have essentially nothing in common except a causal ancestor. A single institution addressing all of them would be an institution without a doctrine. The response, developed in Chapter 4, is that the fire service also lacks a single doctrine: structural, wildland, hazmat, and technical rescue are largely separate disciplines under one dispatch umbrella.
5. There is no analogue of extinguishment A fire, once out, is out. There is no state of an LLM deployment corresponding to "extinguished." Model weights persist, copies proliferate, and capability once demonstrated is capability that others reproduce. The fire service's entire operational tempo, respond, suppress, overhaul, return to quarters, presupposes a terminal state that does not exist here. Any proposal for an "AI fire department" that imagines units returning to quarters after resolving an incident has mistaken the metaphor for the mechanism.
6. The response channel is the propagation channel Fire departments do not arrive via the fire. They use roads, which fire does not travel. Responses to LLM-mediated harm travel the same networks, platforms, and often the same models as the harm. This collapses a separation that is structural in the fire case and creates failure modes with no fire analogue: a compromised response capability, a response that is itself generated content of contested provenance, and an adversary who observes the response in real time.

Main Conclusion

The question is not whether LLMs are like fire. They are not, in most of the ways that matter, and Chapter 3 says so at length. The question is whether we will need something like a fire department, and whether we will build it before or after the fire that makes the case unanswerable. The historical record on that question is not encouraging, but it is at least clear, and clarity about our own tendencies is the one resource that reliably precedes the loss rather than following it.

Core Argument

Part I. The Analogy and Its Limits

The opening move maps fuel to latent capability, oxidizer to deployment surface, and ignition to eliciting context. It then rejects a literal reading: there is no analogue of extinguishment, the response channel is also the propagation channel, and fire departments do not build fires. What survives is institutional form, not mechanical sameness.

Part II. The Institutional Lesson

The load-bearing section compares municipal fire services with property-insurer brigades, then maps those differences onto producer-run trust and safety. It also takes up occupancy classification, where the code tracks exit capability rather than burnability, and the suppression paradox, where success on small fires helped create conditions for larger ones.

Part III. Design Consequences

The final section turns the analysis into six design commitments with falsification conditions. The emblem is a hose coupling: when engines from three cities arrived in Baltimore in 1904, many could not connect to local hydrants because thread pitch was not standardized. The technical failure was ordinary. The institutional failure was the point.

Key Claims

The relevant analogy is not fire as phenomenon, but fire as a general-purpose danger surrounded by institutions.
AI governance needs the equivalent of alarms, inspections, hydrants, codes, and a response network, not just better model behavior.
Successful suppression can change the environment in ways that raise the cost of later failures.
Design commitments matter only if they can be falsified, not merely admired.

What the PDF Contains

Structural mapping

Fire is used carefully, then challenged. The point is not similarity alone; it is what survives after the disanalogies are stated plainly.

Institutional lesson

The transferable part is the institution. The study argues that AI governance needs fire-department thinking more than fire-analogy rhetoric.

Design commitments

The study ends with six commitments and explicit falsification conditions so the argument can be tested, not merely admired.

Reading Note

Analogies between artificial intelligence and earlier general-purpose technologies are cheap, abundant, and mostly bad. The useful comparison is not in the technology itself, but in the institution built around it.

Navigation

Studies hub Frontiers of Intelligence series Philadelphia Journal of Artificial and Digital Minds Journal of Ideas Directory