AI in Defense: A 2025 Holiday Reading Roadmap

A lot of defense teams are buying AI like it’s a plug-in: procure a model, add a dashboard, declare “transformation.” Then reality shows up—messy data, unclear authorities, brittle workflows, and commanders who don’t trust outputs they can’t interrogate.

That’s why I like the annual War on the Rocks holiday reading list. It’s not an “AI list,” and that’s the point. It’s a cross-section of how serious practitioners think: about institutions, logistics, mobilization, deception, adaptation, and the human costs of force. If you’re working in AI in government & public sector, this is the kind of reading that makes your AI strategy operationally honest.

Below is a curated, AI-forward way to use the 2025 list: not as a pile of titles, but as a roadmap for building AI programs that actually work in defense and national security—especially for mission planning, intelligence analysis, and decision support.

Why AI programs in defense fail (and how reading helps)

AI in national security succeeds when it strengthens institutions and decisions—not when it replaces them. Most failures I’ve seen share the same root cause: the team optimizes the model and ignores the system.

The reading list repeatedly hits the “system” problem from different angles:

• Institutional inertia and service rivalry slow adoption more than hardware does.
• Logistics and mobilization are where wars are won and lost—and where AI value is easiest to measure.
• Misinterpretation, deception, and cognitive bias are constant, even in data-rich environments.
• Human judgment and legal constraints don’t disappear because an algorithm is confident.

If you want a modern AI governance posture for defense, you need more than model cards and safety checklists. You need historical and organizational literacy—because AI doesn’t land in a vacuum. It lands in a bureaucracy.

The five “AI problem sets” this reading list quietly trains you for

Here’s the direct value of the list for AI leaders in defense and intelligence: it builds intuition for five problem sets that determine whether AI will be trusted and used.

1) AI and the myth of perfect battlefield awareness

A recurring temptation in defense AI is to treat ISR fusion and sensor proliferation as a straight line to omniscience. Books on this list push back on that idea.

Ben Connable’s Ground Combat: Puncturing Myths of Modern War argues that the dream of near-total awareness keeps returning—and keeps failing. For AI teams, the actionable lesson is simple:

If your AI product assumes complete, clean, timely data, it’s not a defense product. It’s a demo.

Practical application for an AI decision support program:

• Design for uncertainty: show confidence intervals, missingness, and competing hypotheses.
• Treat “unknown” as a first-class output, not an error condition.
• Build workflows that help analysts and operators challenge the model, not just consume it.

2) Automation vs judgment: where humans must stay in the loop

Anthony King’s AI, Automation, and War (highlighted in the list) lands on a point many acquisition decks avoid: AI can handle data load and patterning, but it does not reliably produce judgment that is accountable under policy and the law of armed conflict.

That maps directly to how you should scope defense AI in 2026 budgeting cycles:

• Use AI for triage (prioritize leads, surface anomalies, cluster reports).
• Use AI for recommendations (rank options with assumptions visible).
• Avoid AI as the decider in uses of force, target prosecution, or escalation management.

If you’re leading AI in defense operations, align system design to real authorities: who is allowed to decide, who signs, and who can audit the decision later.

3) Disruptive innovation is mostly doctrine and organization

Andrew Krepinevich’s The Origins of Victory makes an old but painful point: technology doesn’t win by itself. Militaries win when they integrate technology into doctrine, org design, training, and culture.

This is where many public sector AI initiatives overinvest in prototypes and underinvest in adoption.

A concrete, repeatable playbook I’ve found effective:

1. Pick one operational decision (not a broad mission) where AI can measurably reduce time or error.
2. Build the “human loop” first: roles, thresholds, escalation paths, override rules.
3. Make data pipelines and labeling a standing capability, not a one-time sprint.
4. Train to failure: run exercises where the model is wrong, late, or spoofed.

That last step matters because adversaries learn too.

4) Intelligence failure is a product problem, not just an analyst problem

Steve Coll’s The Achilles Trap (on Saddam Hussein, the CIA, and the road to Iraq) is a reminder that catastrophic misreads aren’t caused by a lack of information—they’re caused by how institutions interpret, communicate, and incentivize conclusions.

For AI teams building tools for intelligence analysis, that should change your interface priorities:

• Don’t just output an assessment—output the chain of evidence, dissent, and what would falsify the claim.
• Track assumption drift over time (what you believed last month vs now).
• Capture analyst feedback as structured data so the system learns institutional nuance.

If your AI tool can’t support dissent and auditability, it will either be ignored or—worse—used as a stamp of false certainty.

5) Logistics is the best near-term ROI for defense AI

If you want AI value that survives scrutiny, put it where outcomes are legible: readiness, sustainment, maintenance, supply chains, and mobilization.

Ecyk Freymann and Harry Halem’s The Arsenal of Democracy (noted for its logistics chapter) and Phillips Payson O’Brien’s War and Power both reinforce the “systems of production and sustainment” view of war.

For AI in government procurement leaders, that suggests a near-term portfolio tilt:

• Predictive maintenance with clear metrics (mean time between failure, parts availability)
• Demand forecasting for spares under contested logistics assumptions
• Workforce and training pipeline analytics
• Fraud, waste, and abuse detection in contracting data

These aren’t glamorous. They also don’t require pretending AI can replace strategy.

A curated “AI in national security” reading path from the 2025 list

If you only read five titles from the list with an AI-in-defense lens, start here. (These are all on the War on the Rocks holiday roundup; I’m grouping them by the AI questions they help you answer.)

For AI policy, law, and human control

• AI, Automation, and War (Anthony King)
• Ground Combat: Puncturing Myths of Modern War (Ben Connable)

Use these to set boundaries: what AI should do, what it must not do, and how to design for contested information.

For modernization that actually sticks

• The Origins of Victory (Andrew F. Krepinevich Jr.)
• Adaptation Under Fire (David Barno and Nora Bensahel)

Use these to pressure-test your adoption plan: training, org charts, authorities, and wartime adaptation.

For strategy and escalation realism

• How the United States Would Fight China (Franz-Stefan Gady)

Use this to keep AI claims grounded in operational realities—especially where decision speed collides with escalation risk.

If you have time for two more, I’d add:

• The Magic Bullet? Understanding the Revolution in Military Affairs (Tim Benbow) for anti-hype perspective.
• The Achilles Trap (Steve Coll) to internalize how misperception becomes policy.

How to turn “holiday reading” into a 30-day AI capability upgrade

Reading is only useful if it changes what you build on Monday. Here’s a lightweight 30-day plan that defense and public sector teams can run without a big reorg.

Week 1: Write your “AI job description”

Pick one mission process (collection management, watchfloor triage, maintenance scheduling). Write a one-page statement:

• Decision being made
• Who owns it (authority)
• Inputs available (and missing)
• Time constraints
• Failure modes that would be unacceptable

Week 2: Map the uncertainty and deception surface

Assume adversarial pressure. Document:

• What data can be spoofed or saturated
• What sensors go dark first
• Where the model might overfit historical patterns

Then build a requirement: the system must degrade gracefully.

Week 3: Build evaluation like a mission rehearsal

Stop benchmarking only on static datasets. Create test events:

• Late data
• Conflicting reports
• Label noise
• “Looks plausible” deception

Measure not just accuracy, but:

• Time saved
• Analyst overrides
• False confidence rate
• Audit completeness

Week 4: Put governance where operators will feel it

Good AI governance is operational, not ceremonial. Implement:

• A clear override policy
• Logging for accountability
• Thresholds tied to authorities
• A cadence for model updates that matches operational tempo

If your governance can’t survive an IG inquiry or a post-incident review, it’s not governance.

The reading list’s quiet message: AI doesn’t remove friction—it redistributes it

A theme running through the 2025 recommendations—whether the book is about pirates, Polynesian navigation, airpower failures, or Cold War covert networks—is that complex systems fail in predictable ways: incentives misalign, information distorts, and confidence outruns competence.

AI will not remove those dynamics from defense and intelligence. It will concentrate them in new places: training data, interfaces, evaluation, acquisition pathways, and policy control points.

If you’re building AI for defense operations, the professional move this holiday season is to read like a builder, not a spectator. Take the titles that challenge your assumptions, then use them to tighten your requirements, governance, and evaluation.

The next budget cycle will reward programs that can prove two things: they improve decisions, and they stay trustworthy under pressure. Which AI capability in your organization is most at risk of failing that test—and what are you going to change before it’s deployed?