Intro
L1
·
Quiz
·
Lab
L2
·
Quiz
·
Lab
L3
·
Quiz
·
Lab
L4
·
Quiz
·
Lab
Module Test
AI and National Security · Introduction

Every technology of consequence becomes a weapon. AI isn't becoming one — it's redefining the word.

This is where the stakes are highest, the actors least transparent, and the decisions most consequential.

In 1939, a handful of physicists knew that uranium fission could be scaled into a weapon. By 1945, a nation-state had used two of them. The lag between interesting physics and geopolitically decisive was six years. The history of consequential technology is the history of that lag shrinking.

For AI, it's already zero. Autonomous drones are being deployed in active conflict. Deepfake-driven disinformation is a standing feature of every national election. Cyberwarfare has been AI-augmented for nearly a decade. Intelligence agencies in every major capital are running internal AI programs whose details we can only infer.

This course is about AI as a security technology — not in the abstract but as actors (governments, militaries, intelligence services, and non-state adversaries) are actually using it in 2026. It covers autonomous weapons systems, cyber offense and defense, election integrity, the geopolitics of AI compute, and the question that sits under all of it: what responsible security policy looks like when the technology moves faster than the policy can be written.

If you finish every module, here's who you become:

  • You'll understand how autonomous weapons systems work, where international law currently draws lines, and where it doesn't.
  • You'll be able to read a news story about AI in conflict and identify which actors, doctrines, and incentives are actually in play.
  • You'll know why AI compute — chips, data centers, export controls — is now a primary instrument of geopolitical power.
  • You'll trace how a civilian AI capability becomes a military one, and why dual-use is the rule rather than the exception.
  • You'll think like someone who has to write policy for a technology that will have moved before the ink is dry.
  • You'll be able to assess AI arms control proposals against what treaty verification actually requires — and why most proposals fail that test.
  • You'll become the kind of person who doesn't need a briefing to understand what's at stake when governments talk about AI and national security.
Module 1 · Lesson 1

Autonomous Weapons & the Kill Chain

From Predator drones to Kargu-2: how AI is compressing the decision loop between target and trigger.
When a machine selects a target and fires without human approval, who bears responsibility for the outcome?

In the spring of 2020, a United Nations Panel of Experts report documented an engagement during the Libyan civil war in which a Kargu-2 loitering munition — a Turkish-built autonomous drone — hunted down and attacked retreating Haftar-affiliated forces without requiring a human operator to authorize each strike. The UN report, released in March 2021, described the system as having been programmed to attack targets autonomously, marking what many analysts identified as the first documented lethal engagement by an autonomous weapons system without direct human control over individual targeting decisions.

The event passed with little public attention. No international treaty was violated — because no such treaty yet exists.

The Autonomy Spectrum

Military AI does not arrive as science fiction. It arrives incrementally, in layers. The U.S. Department of Defense describes autonomous weapons along a spectrum defined by human involvement in the "kill chain" — the sequence of steps from target identification to weapons release.

At one end: human-in-the-loop systems require a human to authorize every engagement. The MQ-9 Reaper drone, though highly capable, still requires a human operator to press the trigger. At the other end: human-out-of-the-loop systems act entirely on their own. In between lies human-on-the-loop — systems that can engage autonomously but allow a human to intervene if they're watching and fast enough.

~900
Drone strikes conducted by the U.S. under the post-9/11 AUMF, 2001–2022, all requiring human authorization per DoD policy
2012
Year the DoD issued Directive 3000.09, the first major policy requiring "appropriate levels of human judgment over the use of force" in autonomous systems

Key Concepts

LAWSLethal Autonomous Weapons Systems — weapons that can select and engage targets without human intervention. No binding international definition or treaty yet exists.
Kill ChainThe military targeting sequence: Find, Fix, Track, Target, Engage, Assess (F2T2EA). AI is being applied at every stage, most controversially at "Engage."
Meaningful Human ControlA proposed standard in international humanitarian law debates requiring that a human must be able to understand, predict, and intervene in a weapon's targeting decisions.
Loitering MunitionA weapon that can fly to a target area, circle ("loiter"), search for targets matching pre-programmed criteria, and strike. The Kargu-2, Switchblade, and Harop are examples.

Project Maven: AI Enters the Pentagon

In April 2017, the U.S. Department of Defense launched Project Maven (formally the Algorithmic Warfare Cross-Functional Team), tasking Google's TensorFlow AI to analyze full-motion drone video and automatically identify objects — vehicles, buildings, people — to assist human analysts. At its peak, the system was processing over one million minutes of drone footage per day from operations in Iraq and Syria.

In 2018, approximately 4,000 Google employees signed an open letter opposing the project, citing ethical concerns about AI in warfare. Google declined to renew its contract. The Pentagon continued the program with other contractors, including Palantir and Anduril. Project Maven remains active today under the National Geospatial-Intelligence Agency.

Maven's controversy crystallized a tension that defines the field: AI dramatically accelerates military effectiveness, but the corporations building the most capable AI systems are staffed by engineers who may refuse to build weapons.

Strategic Reality

DoD Directive 3000.09, updated in 2023, continues to require that autonomous weapons be designed to allow commanders to exercise "appropriate levels of human judgment." Critics note the directive contains a national-security waiver that allows the Secretary of Defense to override these requirements — and that "appropriate" is never precisely defined.

Iron Dome and the Automation Precedent

Israel's Iron Dome air defense system, deployed since 2011, operates in an effectively automated mode during high-volume rocket attacks. The system uses radar and AI to track incoming projectiles, calculate intercept trajectories, and launch interceptors — all within seconds, far faster than any human decision loop. Human operators can technically intervene, but at the volume of attacks Iron Dome faces (it intercepted over 1,500 rockets in a single week during May 2021), meaningful human review of each intercept decision is operationally impossible.

Iron Dome is widely viewed as a defensive system and attracts little ethical controversy. But defense analysts point out that the precedent it establishes — machines autonomously engaging targets because humans cannot act fast enough — applies equally to offensive systems as adversarial speeds increase.

The Speed Problem

Modern electronic warfare, hypersonic missiles, and cyberattacks operate at speeds that make human-in-the-loop authorization increasingly impractical. The military logic for AI autonomy is partly a product of adversarial timelines: if a hypersonic missile arrives in 90 seconds, a human authorization chain may take longer than that.

Lesson 1 Quiz

Autonomous Weapons & the Kill Chain — 4 questions
1. The 2021 UN Panel of Experts report on Libya documented what historically significant event?
Correct. The Kargu-2 incident in Libya, documented by the UN, is widely cited as the first documented case of an autonomous lethal engagement without per-strike human authorization.
Not quite. The UN Panel of Experts report specifically described the Kargu-2 autonomous drone hunting and attacking targets without human authorization for each strike — a first of its kind in documented conflict.
2. In the autonomy spectrum defined by the DoD, what does "human-on-the-loop" mean?
Correct. Human-on-the-loop systems act autonomously but allow human override — the key question is whether meaningful intervention is actually possible given the speed of engagement.
That describes human-in-the-loop (button press required) or human-out-of-the-loop (no awareness). Human-on-the-loop sits in between: autonomous action, human can intervene.
3. What was Project Maven, and what controversy did it generate in 2018?
Correct. Project Maven applied commercial AI to drone surveillance. The employee revolt at Google and the company's subsequent exit became a landmark case in the ethics of AI in defense contracting.
Project Maven was the Pentagon's Algorithmic Warfare program that partnered with Google to analyze drone footage. Google's internal employee opposition — roughly 4,000 signatures — caused the company to withdraw from the contract.
4. Why do defense analysts argue that Iron Dome establishes an important precedent for offensive autonomous systems?
Correct. The operational reality that human authorization is impossible at the speed and volume Iron Dome faces creates a logical template that adversaries and militaries can apply to offensive engagements as weapons speeds increase.
The precedent is about speed and volume, not specific offensive use. When machines must act autonomously because humans cannot respond in time — as with Iron Dome — the same logic extends to any engagement scenario where adversarial timelines are shorter than human decision loops.

Lab 1: The Autonomy Threshold Debate

AI Discussion Lab — AI & National Security

Your Task

You are a policy analyst advising a Senate Armed Services subcommittee on autonomous weapons regulation. The AI assistant will play the role of a DoD doctrine expert. Engage in a substantive discussion about where to draw the line on autonomous targeting authority.

Complete at least 3 exchanges to finish this lab.

Start by asking: "Under what operational conditions should a weapon system be permitted to engage a target without real-time human authorization?" — then push back on the answers you receive.
AI Discussion Lab
DoD Doctrine Mode
Welcome, Analyst. I'm briefing from the DoD doctrine perspective on autonomous weapons policy. The committee wants to understand where human authorization requirements should begin and end. What's your first question?
Module 1 · Lesson 2

AI in Intelligence, Surveillance & Reconnaissance

SKYNET was science fiction. SKYNET was also the real name of a U.S. intelligence program that generated kill lists in Pakistan and Somalia.
When AI sorts through billions of data points to name someone a threat, what is the acceptable error rate for a system that triggers lethal action?

Documents released by Edward Snowden and reported by The Intercept in 2015 revealed that the NSA operated a machine learning program called SKYNET — its actual name — that analyzed metadata from Pakistan's mobile phone network to identify potential couriers for al-Qaeda leadership. The algorithm processed call records, travel patterns, and behavioral signatures for 55 million Pakistanis to produce a ranked list of terrorism suspects.

Data scientists who reviewed the methodology publicly assessed the algorithm's accuracy at roughly 50 percent — no better than a coin flip — when applied to the tiny fraction of the population it flagged as highest-risk. At least one person executed in a drone strike, Afghan journalist Ahmad Muaffaq Zaidan of Al Jazeera, appeared on the SKYNET list. He was never charged with any offense.

How AI Transforms ISR

Intelligence, Surveillance, and Reconnaissance (ISR) was already the domain of massive data collection before AI. What AI changes is the ability to process and correlate that data at scale. A single Global Hawk drone produces roughly 1 terabyte of imagery per day — far more than human analysts can review. AI provides automated screening, object detection, pattern-of-life analysis, and anomaly flagging that allows small teams of analysts to effectively supervise surveillance of entire regions.

The NSA's XKeyscore program — also disclosed in the Snowden documents — used rule-based and statistical classifiers to triage internet traffic from over 150 global collection points. More recent systems like the Pentagon's MAVEN Smart System and the intelligence community's Sentient program (operated by the National Reconnaissance Office) apply deep learning to satellite and aerial imagery to track moving objects, detect construction changes, and identify military vehicles across millions of square miles.

1 TB
Data generated per day by a single RQ-4 Global Hawk surveillance drone — equivalent to roughly 250,000 digital photos
55M
Pakistani mobile subscribers whose metadata was processed by NSA's SKYNET machine learning program to identify terrorism suspects

Key Concepts

Pattern of LifeAnalysis of a target's routine behaviors — movement, communications, associations — over time. AI enables pattern-of-life analysis at scales impossible for human analysts.
Activity-Based Intelligence (ABI)An intelligence methodology focused on detecting anomalies in patterns of human behavior, rather than tracking known individuals. AI dramatically expands ABI's reach.
MetadataData about communications — who called whom, when, how long, from where — rather than content. Courts have ruled metadata collection requires less legal protection than content; SKYNET used only metadata.
Algorithmic AccountabilityThe principle that the designers and deployers of AI systems bear responsibility for the outcomes those systems produce. In ISR, this question is largely unresolved in law.

Palantir and the Targeting Pipeline

Palantir Technologies has held contracts with the U.S. Army, Special Operations Command, and intelligence agencies since 2005. Its Gotham platform ingests data from disparate military and intelligence databases and presents fused, correlated intelligence pictures — effectively converting raw signals into actionable targeting information. In 2021, the Army awarded Palantir a $823 million contract for its Tactical Intelligence Targeting Access Node (TITAN) system, designed to use AI to connect ground-based sensors to targeting data for long-range precision fires.

TITAN represents a shift: rather than AI helping analysts understand a battlefield, AI helps compress the time between data collection and weapons employment. The human analyst remains in the loop, but the loop gets shorter — and the AI's role in shaping what the analyst sees, and therefore what they decide, grows correspondingly larger.

The Accuracy Problem

When applied to populations where the true base rate of the thing being detected (terrorism) is extremely low — perhaps 1 in 1,000,000 — even a 99.9%-accurate classifier produces thousands of false positives for every true positive. This statistical reality, known as the base rate fallacy, makes AI targeting systems inherently dangerous when applied to large, low-prevalence threat populations. The SKYNET program's reported 50% accuracy on its highest-confidence outputs illustrates how far short of acceptable current systems can fall.

China's Military ISR AI

China's People's Liberation Army (PLA) has invested heavily in AI-enabled ISR, including the Sharp Eyes (雪亮工程) program, which links hundreds of millions of surveillance cameras with AI-powered facial recognition across Chinese territory. Publicly available PLA documents describe plans for "intelligentized warfare" in which AI fuses data from space, air, sea, ground, and cyber domains in real time to present commanding officers with a continuously updated operational picture. The U.S. Department of Defense's 2023 China Military Power Report assessed that China views AI-enabled ISR as central to achieving information dominance — the ability to see and act faster than an adversary — in any future conflict.

Lesson 2 Quiz

AI in ISR — 4 questions
1. What did independent data scientists conclude about the NSA's SKYNET algorithm?
Correct. Data scientists who reviewed the Snowden-disclosed methodology assessed SKYNET's accuracy at roughly 50% for its highest-confidence outputs — a deeply problematic result for a system contributing to lethal targeting decisions.
The documented assessment by outside data scientists was far more damning: approximately 50% accuracy — no better than a coin flip — on the individuals the algorithm rated as most suspicious.
2. What is "Activity-Based Intelligence" and how does AI expand its utility?
Correct. ABI looks for anomalies in patterns of behavior rather than tracking known individuals — AI makes it possible to apply this methodology to entire populations across large geographic areas.
ABI is about behavioral pattern analysis and anomaly detection. AI's contribution is scale: systems like SKYNET and Palantir's Gotham can apply ABI logic to millions of people simultaneously, something human analysts could never do.
3. What is the "base rate fallacy" and why does it matter for AI-enabled targeting?
Correct. The base rate fallacy is a critical statistical reality for national security AI: if only 1 in 100,000 people is a genuine threat, a 99%-accurate classifier still generates roughly 1,000 false positives for every true positive in a population of 100 million.
The base rate fallacy refers to a statistical phenomenon: when the true prevalence of a condition (terrorism) is extremely low in a large population, even accurate classifiers produce far more false positives than true positives — a problem that directly affected SKYNET's reliability.
4. What is the Pentagon's TITAN system designed to do?
Correct. TITAN (Tactical Intelligence Targeting Access Node) is a Palantir-built system that uses AI to fuse sensor data into targeting solutions for long-range artillery and missiles — compressing the kill chain.
TITAN is the Army's AI-powered targeting system contracted to Palantir for $823 million in 2021. Its purpose is to fuse sensor data with AI analysis to generate targeting solutions for long-range precision weapons — shortening the kill chain.

Lab 2: Interrogating AI Targeting Accuracy

AI Discussion Lab — AI & National Security

Your Task

You are a civil liberties attorney presenting to the UN Special Rapporteur on extrajudicial executions. The AI assistant represents a military ISR program director defending algorithmic targeting. Interrogate the accuracy standards, error accountability, and legal frameworks governing AI-generated kill lists.

Complete at least 3 exchanges to finish this lab.

Open with: "What accuracy threshold must an AI targeting algorithm meet before its outputs can ethically inform a lethal strike decision, and who verifies compliance?" Then press for specifics on accountability.
AI Discussion Lab
ISR Program Director Mode
I'm briefing as ISR Program Director. Our algorithms undergo rigorous validation before operational deployment — but I understand the civil liberties community has concerns. What specific aspects of targeting accuracy would you like to examine?
Module 1 · Lesson 3

AI-Enabled Cyber Operations & Electronic Warfare

Stuxnet was hand-coded. The next generation of cyberweapons may write, deploy, and mutate themselves.
If an AI-written cyberweapon causes a nuclear plant to malfunction and people die, is it an act of war — and who is responsible?

Stuxnet, the joint U.S.-Israeli cyberweapon that destroyed roughly 1,000 Iranian uranium centrifuges at the Natanz enrichment facility between 2007 and 2010, was meticulously hand-crafted. Analysts at Symantec and Kaspersky estimated it contained over 15,000 lines of code, required years of development by a team with deep knowledge of Siemens industrial control systems, and was tested on replica centrifuge arrays in the United States. Its precise, surgical design was a feature: the weapon was engineered to affect only specific Siemens PLCs running at specific speeds — to destroy centrifuges while leaving surrounding infrastructure intact.

What took a nation-state team years to build, AI-assisted code generation can now approximate in weeks. The implications for cyber proliferation are severe.

How AI Changes Cyber Offense

Traditional cyberweapons require specialized human expertise: finding zero-day vulnerabilities, writing exploit code, building command-and-control infrastructure, and testing against target systems. Each step is time-intensive and requires rare skills. AI is compressing all of these phases simultaneously.

Vulnerability discovery: Large language models fine-tuned on software codebases can identify security vulnerabilities in code at scale. In 2023, researchers demonstrated that GPT-4 could identify and suggest exploits for 87% of "one-day" vulnerabilities — publicly disclosed but unpatched flaws — when given their CVE descriptions. DARPA's AI Cyber Challenge (AIxCC), launched in 2023 with $18.5 million in prizes, explicitly tasks AI teams with autonomously finding and patching vulnerabilities in critical infrastructure software.

Malware generation: Security researchers have demonstrated that AI tools including WormGPT (a jailbroken LLM) and purpose-built tools can produce functional malware code. In 2023, CrowdStrike reported a 71% increase in cloud intrusions, with AI-assisted attack tools cited as a contributing factor to the acceleration of attack timelines.

2010
Stuxnet discovered — first publicly documented cyberweapon to cause physical destruction of industrial equipment. Human-coded, years in development.
2016
DARPA Cyber Grand Challenge — AI systems competed to autonomously find, exploit, and patch vulnerabilities in real-time without human guidance. First AI-only hacking competition.
2022
Russia-Ukraine cyberconflict — Wiper malware (HermeticWiper, CaddyWiper) deployed against Ukrainian infrastructure; researchers noted AI-assisted obfuscation techniques in several samples.
2023
DARPA AIxCC launched — $18.5M competition explicitly tasking AI with autonomously securing critical infrastructure software. Signals U.S. government acceptance of AI as a primary cyber tool.

Key Concepts

Zero-DayA software vulnerability unknown to the vendor, for which no patch exists. AI systems that can find zero-days autonomously represent a potential step-change in offensive cyber capability.
Wiper MalwareMalicious software designed to destroy data rather than steal it — used as a weapon against critical infrastructure. Russia deployed multiple wiper variants against Ukraine before and during the 2022 invasion.
Cyber SovereigntyThe contested principle that states have the right to control internet infrastructure within their borders — often invoked to justify offensive cyber operations as responses to foreign interference.
AI Red TeamA team that uses AI tools to probe systems for vulnerabilities before adversaries do. Both NSA's Cybersecurity Directorate and CISA have established AI red-teaming programs.

Electronic Warfare and AI Signal Processing

Electronic warfare (EW) — jamming, spoofing, and intercepting radio-frequency signals — has always been a contest of speed and signal processing. AI is reshaping this domain by enabling systems to adaptively learn an adversary's radio signatures in real time and adjust jamming strategies faster than any human operator could.

The U.S. Air Force's SPAR (Self-Protected Approach and Landing Capability for Rockets and Missiles) program and the Navy's Next Generation Jammer both incorporate machine learning for adaptive spectrum management. DARPA's Spectrum Collaboration Challenge (2019) demonstrated that AI radio systems could negotiate spectrum sharing with greater efficiency than human-managed systems — a foundational capability for AI-controlled battlefield communications and jamming.

China's PLA Electronic Systems Department has published research on AI-enabled cognitive electronic warfare — systems that learn adversary radar waveforms from intercepted signals and generate optimal countermeasures autonomously. Russian military doctrine documents obtained and published by the European Council on Foreign Relations describe "reflexive control" — manipulating adversary decision-making through information operations — as a domain where AI provides decisive advantage.

Attribution Problem

AI-generated cyberattacks are harder to attribute than human-crafted ones. Distinctive coding styles, tooling fingerprints, and infrastructure reuse — the clues forensic analysts use to attribute attacks to specific nation-states — can be deliberately randomized by AI. The 2021 Microsoft Exchange Server breach (attributed to Chinese group HAFNIUM) and the 2020 SolarWinds supply chain attack (attributed to Russian SVR) both demonstrated sophisticated supply chain and evasion techniques that AI could further automate and obscure.

Lesson 3 Quiz

AI in Cyber & Electronic Warfare — 4 questions
1. What made Stuxnet historically significant as a cyberweapon?
Correct. Stuxnet crossed a threshold: it used code to produce kinetic physical effects — spinning centrifuges to destruction — establishing that cyberweapons can cause physical damage comparable to conventional attacks.
Stuxnet's significance was that it was the first cyberweapon documented to cause physical destruction — the centrifuges spinning themselves apart at Natanz. It was human-coded, not AI, and targeted Iranian nuclear infrastructure specifically.
2. In 2023, researchers demonstrated that GPT-4 could identify exploits for what percentage of disclosed vulnerabilities it was tested on?
Correct. Researchers found GPT-4 could identify and suggest exploits for 87% of "one-day" vulnerabilities when given their CVE descriptions — a finding that alarmed cybersecurity professionals about AI-accelerated attack timelines.
The documented research finding was 87% — GPT-4 could suggest working exploits for 87% of tested disclosed-but-unpatched vulnerabilities, demonstrating how AI can radically lower the skill barrier for cyberattacks.
3. What is the significance of DARPA's AI Cyber Challenge (AIxCC) launched in 2023?
Correct. AIxCC represents an explicit government endorsement of autonomous AI for cyber operations — agencies are no longer merely studying AI cyber tools but actively funding their development through competition.
AIxCC is DARPA's competition explicitly designed for AI systems to autonomously handle cybersecurity tasks on critical infrastructure — a clear signal that the U.S. government views AI as a central tool for both offensive and defensive cyber operations.
4. Why does AI make cyber attack attribution harder for forensic investigators?
Correct. Attribution forensics depend on distinctive patterns — a group's tools, coding habits, infrastructure. AI can generate deliberately variable code and randomize these signatures, making it harder to identify who launched an attack.
The attribution problem stems from AI's ability to randomize the forensic fingerprints analysts rely on: distinctive coding styles, reused tools and infrastructure, behavioral patterns. AI can deliberately vary these to obscure the attacker's identity.

Lab 3: AI Cyberweapons Proliferation Debate

AI Discussion Lab — AI & National Security

Your Task

You are a cybersecurity policy director at the State Department. The AI assistant represents a national security hawk arguing that the U.S. must develop and deploy AI-enabled offensive cyber capabilities at scale before adversaries do. Engage critically on the question of AI cyberweapon proliferation, escalation risk, and arms control.

Complete at least 3 exchanges to finish this lab.

Begin with: "If we develop AI systems that can autonomously write and deploy cyberweapons, what prevents our adversaries from doing the same — and what happens when both sides have them?" Challenge the strategic logic carefully.
AI Discussion Lab
Cyber Hawk Mode
The strategic calculus is straightforward: adversaries are already building AI cyber capabilities. The question isn't whether these tools will exist — it's whether we lead or follow. What's your concern with maintaining U.S. offensive cyber superiority?
Module 1 · Lesson 4

Governance, Arms Control & the International Debate

Nuclear weapons took 30 years to develop partial international controls. Autonomous weapons have been deployed for a decade, and the world still has no binding rules.
Can treaties meaningfully constrain AI weapons when the technology is dual-use, the definitions are contested, and verification is nearly impossible?

In May 2014, the United Nations' Convention on Certain Conventional Weapons (CCW) hosted its first formal discussion on Lethal Autonomous Weapons Systems. A decade later, nations are still meeting, still discussing, and still producing no binding agreement. The United States, Russia, China, Israel, South Korea, and Australia have all formally opposed a binding treaty ban on autonomous weapons. The International Committee of the Red Cross called in 2021 for new rules establishing mandatory human control over lethal force — its most explicit demand to date. As of 2024, the CCW process has produced a set of non-binding guiding principles that carry no legal enforcement mechanism.

Meanwhile, the Kargu-2 was sold. The Switchblade was deployed. Loitering munitions proliferated to over 40 nations.

Why AI Arms Control Is Uniquely Difficult

Traditional arms control — from the Nuclear Non-Proliferation Treaty to the Chemical Weapons Convention — relies on three foundations: a clear definition of the prohibited item, a verification mechanism (inspections, satellites, forensics), and a small number of states possessing the technology. AI weapons satisfy none of these conditions.

Definition problem: Where does "autonomous targeting" begin? Is an air defense system that automatically intercepts incoming missiles autonomous? Is a loitering munition that searches for radar emissions? Different states answer these questions differently, and no consensus definition of a LAWS exists in international law.

Verification problem: Unlike nuclear facilities (which are large, expensive, and detectable by satellite) or chemical precursors (which have distinctive signatures), autonomous weapons software can run on commercially available hardware and be updated without physical evidence of change. There is no inspection regime that could reliably verify compliance.

Proliferation problem: The foundational technologies — computer vision, machine learning, edge computing — are commercially available globally. A nation that signs a LAWS treaty faces adversaries who can develop equivalent capability from commercial drones and open-source AI frameworks.

40+
Nations that have acquired loitering munition capability as of 2024, up from fewer than 5 in 2010 — illustrating how rapidly autonomous weapon technology proliferates
10 yrs
Duration of CCW discussions on LAWS with no binding agreement produced — the longest running major weapons control debate without a treaty in the post-Cold War era

Key Concepts

CCWConvention on Certain Conventional Weapons — the UN forum where LAWS governance is debated. Operates by consensus, meaning any major power can block action.
International Humanitarian Law (IHL)The body of law governing armed conflict, including the principles of distinction (combatant vs. civilian), proportionality, and precaution. Autonomous weapons must theoretically comply with IHL.
Dual-Use TechnologyTechnology with both civilian and military applications. Computer vision that identifies products in a warehouse can also identify vehicles on a battlefield. Dual-use nature makes export controls and treaties ineffective.
Executive Order 13859Signed by President Trump in 2019, directing U.S. agencies to prioritize AI investment to "maintain American leadership in AI" — framing AI superiority explicitly as a national security priority.

The Great Power Competition Framework

The U.S. National Security Strategy (2022) identifies China as the "most consequential geopolitical challenge" and lists AI as a primary domain of competition. China's "New Generation AI Development Plan" (2017) explicitly set the goal of global AI leadership by 2030, with defense applications listed as a priority. Russia's 2030 AI strategy, published in 2019, similarly frames AI leadership as essential to strategic parity with the United States.

This competitive framing creates a powerful dynamic against arms control: any nation that agrees to limit AI weapons development risks falling behind adversaries who may defect from or simply ignore such limits. The structure mirrors the early nuclear era — before Hiroshima — when each power reasoned it could not afford to be second.

The Campaign to Stop Killer Robots

A coalition of more than 270 NGOs from 70 countries, the Campaign to Stop Killer Robots has lobbied since 2013 for a binding international treaty banning fully autonomous weapons. The campaign counts support from 70 UN member states as of 2024. Among the major military powers — the U.S., China, Russia, U.K., France, and Israel — none supports a binding prohibition. The gap between civil society consensus and state behavior reflects the perceived strategic value of autonomous weapons to those who currently lead in their development.

The U.S. AI Safety Approach: A Middle Path?

The Biden Administration's October 2023 Executive Order on AI included a classified annex on military AI applications. The publicly visible elements of U.S. policy — including the DoD's Responsible AI Strategy (2022), the Political Declaration on Responsible Military Use of AI (2023, endorsed by 50+ nations), and Pentagon Directive 3000.09 — attempt a middle path: permitting AI in defense applications while requiring "appropriate" human control and ethical use standards.

Critics argue this approach avoids the hard questions. "Appropriate human control" over a system engaging targets at machine speed is not defined with operational precision. The Political Declaration is explicitly non-binding. And the nations whose practices raise the most concern — Russia, China, Iran, North Korea — did not sign it.

The governance gap is not primarily a technical problem. The technology to implement meaningful human control exists. It is a problem of political will: nations that believe autonomous weapons provide decisive military advantage have strong incentives to preserve that advantage by resisting binding constraints.

Looking Ahead

The most likely near-term governance outcome is not a comprehensive LAWS treaty but rather fragmented national regulations, informal norms, and escalating incidents that gradually — and perhaps catastrophically — pressure states toward more binding agreements. The parallel to chemical weapons is instructive: the Chemical Weapons Convention was negotiated after widespread use in World War I had established the horror of these weapons in international consciousness. For autonomous weapons, that formative event has not yet occurred at sufficient scale.

Lesson 4 Quiz

Governance & Arms Control — 4 questions
1. Why does AI weapons control face a harder "definition problem" than nuclear or chemical weapons control?
Correct. The definitional ambiguity is real and exploited: Is Iron Dome a LAWS? Is a Switchblade? Different militaries answer differently, and without a shared definition, a treaty cannot specify what it prohibits.
The CCW has been studying this since 2014 — the problem isn't awareness, it's definition. Autonomous weapons exist on a continuum from semi-autonomous aids to fully autonomous systems, and states disagree fundamentally about where prohibition should begin.
2. What is the central verification problem facing any LAWS arms control treaty?
Correct. Unlike nuclear facilities or chemical precursor stockpiles, autonomous weapons capability is defined by software that can be invisibly updated on commercially available hardware — making any meaningful inspection regime essentially impossible with current technology.
The verification problem is about software, not hardware. An autonomous weapons software package can be updated remotely, runs on commercial drones, and leaves no physical evidence that traditional arms control inspections could detect.
3. As of 2024, what has the CCW's decade of LAWS discussions produced?
Correct. Ten years of discussion have yielded only non-binding principles. The gap between the pace of debate and the pace of proliferation represents a defining governance failure of the current era.
A decade of CCW discussions has produced only non-binding guiding principles — no binding treaty, no moratorium, no enforcement mechanism. Meanwhile, autonomous munitions have proliferated to more than 40 nations.
4. Why does the "great power competition" dynamic work against AI arms control agreements?
Correct. The strategic logic is a classic security dilemma: any binding restraint you accept, an adversary might not — so restraint risks disadvantage. This same logic drove the nuclear arms race until the costs became catastrophic enough to force agreements.
The structural problem is strategic, not industrial. If the U.S. agrees to limit autonomous weapons and China does not, the U.S. has sacrificed a military advantage. This reasoning — which all major powers apply symmetrically — creates a collective action failure that resists resolution short of a formative catastrophe.

Lab 4: Designing LAWS Governance

AI Discussion Lab — AI & National Security

Your Task

You are a diplomat leading a new UN working group tasked with drafting a framework for governing lethal autonomous weapons. The AI assistant will play a skeptical Chinese diplomatic counterpart who supports developing AI weapons but is open to limited norms. Your challenge: find workable governance mechanisms that might actually attract state consent.

Complete at least 3 exchanges to finish this lab.

Open with: "Rather than a comprehensive ban, could we agree on a set of minimum operational requirements — a floor of human control that all systems must meet — without defining the ceiling? Walk me through why that might or might not be acceptable to your government."
AI Discussion Lab
Diplomatic Negotiation Mode
China's position is clear: we will not accept any framework that constrains our sovereign right to develop defense technologies necessary for national security. But we are prepared to discuss norms around responsible use. What specific "floor" are you proposing, and who would verify compliance?

Module 1 Test

AI in Defense Systems — 15 questions · 80% to pass
1. What was the name of the Turkish-built autonomous drone weapon involved in the 2020 Libya engagement documented by the UN?
Correct. The Kargu-2 loitering munition, built by Turkish defense firm STM, was documented in the 2021 UN Panel of Experts report as having autonomously engaged targets in Libya.
The UN documented the Kargu-2 — a Turkish STM-built loitering munition — in the Libya engagement. The Bayraktar TB2, while also Turkish, is operated with a human pilot in the loop.
2. DoD Directive 3000.09, first issued in 2012, requires what regarding autonomous weapons?
Correct. The directive requires "appropriate levels of human judgment" — deliberately ambiguous language that critics note lacks operational precision, particularly for high-speed engagements.
Directive 3000.09 requires "appropriate levels of human judgment over the use of force" — but contains a national-security waiver and does not precisely define what "appropriate" means operationally.
3. Which company was contracted by the Pentagon for Project Maven and subsequently withdrew after internal employee opposition?
Correct. Google withdrew from Project Maven in 2018 after approximately 4,000 employees signed a petition. Palantir subsequently took on much of the AI targeting analysis work.
Google held the original Project Maven AI contract but withdrew in 2018 following the employee petition. The project continued with other contractors including Palantir and Anduril.
4. What does the F2T2EA model describe in military targeting?
Correct. F2T2EA — Find, Fix, Track, Target, Engage, Assess — describes the military kill chain. AI is being applied at every stage, with the "Engage" phase being the most ethically contested.
F2T2EA is the military kill chain: Find, Fix, Track, Target, Engage, Assess. The debate around autonomous weapons focuses particularly on whether AI should control the "Engage" step without human authorization.
5. What Afghan journalist appeared on the NSA's SKYNET target list despite never being charged with any offense?
Correct. Ahmad Muaffaq Zaidan, Al Jazeera's bureau chief in Islamabad, appeared on the SKYNET list — an illustration of how an algorithmic system with a reported ~50% accuracy rate can misidentify journalists as threats.
Ahmad Muaffaq Zaidan — Al Jazeera's Pakistan bureau chief — appeared in the SKYNET outputs. His case became emblematic of the danger of low-accuracy AI systems contributing to lethal targeting decisions.
6. What does "Activity-Based Intelligence" focus on, and how does AI change its effectiveness?
Correct. ABI looks for anomalies in behavioral patterns across large populations. AI makes it possible to apply this methodology to millions of people simultaneously — a qualitative expansion of surveillance reach.
ABI is about behavioral pattern analysis across populations. AI's contribution is scale — what once required identifying specific known individuals can now be applied to entire regions, flagging behavioral anomalies without prior identification of suspects.
7. Stuxnet destroyed centrifuges at which Iranian facility?
Correct. Stuxnet targeted Siemens PLCs controlling centrifuges specifically at the Natanz uranium enrichment facility, destroying approximately 1,000 centrifuges between 2007 and 2010.
Stuxnet targeted Natanz, Iran's primary uranium enrichment site. The weapon was specifically coded to affect Siemens PLCs running at the speeds used in Natanz's centrifuge cascades.
8. In 2023, what percentage of "one-day" vulnerabilities could GPT-4 suggest exploits for when given their CVE descriptions?
Correct. The 87% figure from 2023 research alarmed cybersecurity professionals: an off-the-shelf LLM could suggest working exploits for the vast majority of tested disclosed vulnerabilities, dramatically lowering the expertise barrier for cyberattacks.
Research demonstrated GPT-4 could suggest working exploits for 87% of tested one-day vulnerabilities — a finding that demonstrated how AI tools can radically democratize sophisticated cyberattack capabilities.
9. What is the "base rate fallacy" and why does it specifically threaten AI targeting systems?
Correct. This is the essential statistical problem for AI targeting: in a population of 55 million where perhaps 100 people are genuine threats, even a 99.99%-accurate classifier would still produce roughly 5,500 false positives for every 100 true positives.
The base rate fallacy is about mathematical inevitability: when genuine threats are extremely rare in a large population, classifiers produce vastly more false positives than true positives. This makes SKYNET-style population screening statistically treacherous regardless of the algorithm's technical accuracy.
10. What was the value of the Army's 2021 contract with Palantir for the TITAN targeting system?
Correct. The $823 million TITAN contract (Tactical Intelligence Targeting Access Node) reflects the scale of U.S. investment in AI-enabled kill chain compression — connecting ground sensors to precision fires through AI fusion.
The TITAN contract was $823 million — a substantial investment signaling that AI-enabled targeting fusion is no longer experimental but a core component of Army modernization.
11. What three fundamental requirements of traditional arms control do AI weapons fail to satisfy?
Correct. AI weapons lack a consensus definition (where does autonomy begin?), a verification mechanism (software is invisible to inspectors), and concentrated technology (commercial AI hardware is globally available).
Arms control requires: a clear prohibited item definition, verification capability, and limited state possession of the technology. AI weapons satisfy none of these — the definition is contested, software is invisible to inspectors, and the underlying technology is commercially available worldwide.
12. The CCW's LAWS discussions have been ongoing since 2014. What enforcement mechanism do the resulting "guiding principles" carry?
Correct. After a decade of discussion, the CCW has produced only non-binding principles — a significant governance failure given the pace of autonomous weapons proliferation during the same period.
The CCW's guiding principles are non-binding — they carry no legal force and include no enforcement mechanism. Nations that violate them face no formal consequence under the CCW framework.
13. Israel's Iron Dome intercepts incoming rockets in what mode that raises autonomy concerns?
Correct. During high-volume rocket salvos, Iron Dome engages autonomously because human decision timelines are far longer than the engagement window — establishing a practical precedent for autonomous engagement driven by adversarial speed.
Iron Dome can technically be supervised by humans, but during high-volume attacks (1,500+ rockets in a week in May 2021), per-intercept human authorization is practically impossible. The system operates autonomously out of operational necessity — a precedent defense analysts find significant.
14. China's "Sharp Eyes" program connects what infrastructure to AI-powered facial recognition?
Correct. Sharp Eyes (雪亮工程) links China's vast civilian camera network — hundreds of millions of devices — to centralized AI facial recognition, creating a comprehensive domestic surveillance infrastructure with clear military ISR applications.
Sharp Eyes connects hundreds of millions of Chinese surveillance cameras to AI facial recognition — a domestic surveillance program that also represents a foundational ISR capability applicable to military intelligence operations.
15. What historical parallel does the lesson draw to explain why major powers resist binding LAWS agreements despite clear proliferation risks?
Correct. The pre-Hiroshima nuclear analogy is apt: meaningful arms control arrived only after the weapon's catastrophic potential was demonstrated in use. The implication for LAWS governance is sobering — a formative catastrophic event may be required before binding agreements become politically achievable.
The lesson draws the pre-Hiroshima nuclear parallel: before 1945, each power reasoned it could not unilaterally restrain its nuclear program while adversaries continued. Binding nuclear limits only emerged after catastrophic use. The same dynamic may govern LAWS — making a catastrophic autonomous weapons incident the likely prerequisite for meaningful governance.