UNCLASSIFIED // FOR TRAINING USE ONLY // ALL SCENARIO DATA IS FICTIONAL
USSOCOM // AI TRAINING

THE OPERATOR'S EDGE

AI Utilization & Prompt Engineering Workshop
PHASEPHASE 00
PHASE 00

AI Crash Course

The New Digital Battlespace — 30 Minutes

"The digital battlespace is as critical as the physical one. Our adversaries are already leveraging AI, and we must not only keep pace but dominate this new domain. Today is about giving you a new weapon system — a cognitive force multiplier."

AIArtificial Intelligence

The broad category. Any computer system designed to perform tasks requiring human intelligence — pattern recognition, decision-making, language understanding. AI is the category, not the product.

MLMachine Learning

The engine. Instead of explicit programming, ML systems train on enormous datasets and learn to recognize patterns. Feed it thousands of tank images — it learns to identify tanks in new imagery. This is how Project Maven works.

LLMLarge Language Models

The tool you'll use today. Trained on a massive corpus of text — essentially the written internet. It can summarize a 200-page intel report in 30 seconds, translate documents in real time, draft CONOPs, and wargame COAs. ChatGPT, Gemini, Claude are all LLMs.

PRIMARY MILITARY USE-CASES
DomainAI ApplicationSOF Relevance
Intelligence & ISRAutomated object detection in FMV and satellite imagery; pattern-of-life analysisReplaces hundreds of analyst hours per day. Real-time threat ID from drone feeds.
Targeting & SSECross-referencing targets against databases; generating target packages; biometric matchingCollapses the find-fix-finish cycle. Enables same-night follow-on raids.
Mission PlanningTerrain analysis; logistics optimization; wargaming enemy responses; CONOP draftsGives a small team the analytical horsepower of a full staff.
Info Environment OpsAnalyzing adversary propaganda; identifying key influencers; gauging population sentimentEnables faster, more targeted PSYOP and counter-narrative efforts.
Training & ReadinessGenerating realistic training scenarios; AAR analysis; language trainingPersonalizes and accelerates individual and unit training.
WHAT AI IS NOT GOOD FOR — THE GUARDRAILS
Not a Decision-Maker

AI provides recommendations, not orders. The final ethical and command responsibility for any action — especially lethal action — rests with the human operator. Always.

No Operational Intuition

AI lacks genuine understanding and the nuanced context that comes from years on the ground. It cannot read your mind or infer intent. It is a tool to augment your judgment, not replace it.

Confidently Wrong

LLMs hallucinate — generating factually incorrect information with complete confidence. Always verify AI-generated factual claims against authoritative sources before acting on them.

AGENTIC AI — THE NEXT EVOLUTION

Standard LLMs answer questions. Agentic AI completes missions. An AI agent is a system that can autonomously plan, make decisions, use tools, and execute multi-step tasks — without a human approving every action. It is the difference between asking a junior analyst to summarize a report and tasking a senior analyst to independently research a target, build a network map, draft a product, and brief the commander. The agent does all of it, in sequence, on its own.

LLM— Standard Language Model
Responds to a single prompt
Requires human to chain tasks together
No memory between sessions by default
Cannot access external tools or data
You are the orchestrator
AGENT— Autonomous AI System
Plans and executes multi-step task sequences
Self-directs: decides what to do next
Maintains memory and context across tasks
Can use tools: web search, code, APIs, files
The AI is the orchestrator
HOW AN AGENT WORKS — THE OODA LOOP ANALOGY
01
OBSERVE

The agent receives a goal and gathers available information — reading files, searching the web, querying databases.

02
ORIENT

It reasons about the problem: what do I know, what do I need, what tools do I have, what is the best sequence of actions?

03
DECIDE

It selects the next action — run a web search, write code, call an API, summarize a document, ask a clarifying question.

04
ACT

It executes the action, observes the result, and loops back — continuing until the goal is achieved or it needs human input.

AGENTIC AI — SOF APPLICATIONS
Mission TypeWhat the Agent Does AutonomouslyHuman Role
Target DevelopmentSearches open-source databases, cross-references biometrics, builds a target package, drafts a CONOP — all from a single taskingReview, approve, and execute
SSE ProcessingIngests captured media, translates documents, identifies names and locations, flags actionable intelligence, produces a finished reportValidate and disseminate
Pattern-of-Life AnalysisMonitors multiple data streams over time, identifies anomalies, generates alerts when pattern changes, updates the threat pictureSet parameters, review alerts
PSYOP Campaign PlanningAnalyzes target audience, identifies key influencers, drafts message variants, assesses likely reception, recommends channelsApprove messaging, execute
Logistics & PlanningCalculates routes, identifies resupply points, wargames contingencies, generates risk assessments for multiple COAsSelect COA, provide constraints
CURRENT AGENTIC PLATFORMS — WHAT'S AVAILABLE NOW
ChatGPT (GPT-5)Commercial
Custom GPTs with tool access
Code interpreter (runs Python)
Web browsing and file analysis
Memory across sessions
Free tier available
Data goes to OpenAI servers — UNCLASSIFIED use only
Claude (Anthropic)Commercial
200K - 1 Mil token context window
Computer use (controls a desktop)
Strong reasoning and analysis
Projects with persistent memory
Free tier available
Data goes to Anthropic servers — UNCLASSIFIED use only
DoD / SOCOM PlatformsGovernment
CDAO Task Force Lima initiatives
SOCOM AI-enabled SSE tools
NIPR/SIPR-hosted LLM deployments
Palantir AIP (operational)
Shield AI autonomous systems
Approved for classified use — check your unit's current authorizations
CRITICAL DISTINCTION — AGENTS AND HUMAN CONTROL

Agentic AI introduces a new risk category: autonomous action at machine speed. An agent tasked poorly can take dozens of actions — sending messages, modifying files, querying databases — before a human realizes the task has gone off-course. In military contexts, the principle of meaningful human control is non-negotiable. Agents are force multipliers for planning and analysis. They are not authorized to make targeting decisions, transmit operational information, or take actions with irreversible consequences without explicit human approval at each step. Treat every agent output as a recommendation, not an order.

THREE REAL MILITARY AI EXAMPLES
Project Maven — Automating the ISR Burden
EXAMPLE 01

Project Maven — Automating the ISR Burden

Project Maven, launched by the DoD in 2017, was a pathfinder initiative to operationalize AI across military operations. The DoD was drowning in full-motion video — drone operations across Iraq, Syria, and Afghanistan were generating more footage than human analysts could ever watch. The backlog was measured in years. Maven's solution: apply machine learning to automatically detect, classify, and track objects of interest — vehicles, personnel, weapons systems — in drone footage in near-real time. What previously required a team of analysts working around the clock could now be processed in seconds. For SOF, Task Force 714 in Iraq had already demonstrated the power of intelligence-driven operations, running 300+ raids per month. Maven represented the next evolution — not just processing intelligence faster, but at a scale no human team could match.
SOCOM Sensitive Site Exploitation — The 20-Minute Intelligence Cycle
EXAMPLE 02

SOCOM Sensitive Site Exploitation — The 20-Minute Intelligence Cycle

The 2011 Abbottabad raid produced what intelligence professionals called the most significant haul since the fall of the Soviet Union. SEAL Team Six operators left carrying trash bags full of hard drives, thumb drives, and documents. Processing that material took months. Today, SOCOM is actively working to collapse that timeline from months to minutes. The Command has issued requirements for AI-powered tools capable of performing facial recognition, speaker identification, and DNA profiling directly at point of capture. An operator on a raid uses a ruggedized tablet to run biometrics against military, CIA, and FBI databases in real time. Simultaneously, captured documents and hard drives feed into an AI exploitation system that translates, summarizes, and identifies actionable intelligence — names, locations, planned operations — within minutes. The intelligence from that raid can drive a follow-on operation the same night.
Ukraine — AI on the Peer-Competitor Battlefield
EXAMPLE 03

Ukraine — AI on the Peer-Competitor Battlefield

The conflict in Ukraine has become the world's most intensive real-world laboratory for AI-enabled warfare. Ukrainian forces, fighting a larger and better-equipped adversary, have leveraged AI as an asymmetric equalizer. AI-powered target recognition software integrated into FPV drone systems allows the drone to autonomously identify and track Russian armored vehicles even when GPS is jammed or the communications link to the operator is disrupted. The AI guides the drone to the target autonomously. Results: AI integration boosted Ukrainian FPV drone strike accuracy from the 30–50% range to approximately 80% — a dramatic improvement that fundamentally altered the cost-exchange ratio of armored warfare. A single operator, using AI-assisted systems, can manage multiple drone platforms simultaneously — effectively multiplying individual combat power.
1 / 10
The Technology