AI, autonomy, and the future of naval warfare with Captain Jon Haase, United States Navy
Summary
Haase’s mine-hunting system shows that defense AI is an edge-integration problem before it is a model problem. An ensemble of deep learners runs on an NVIDIA processor aboard unmanned underwater vehicles, identifies likely mines, and triggers deterministic mission branches without surfacing; the same models accelerate laptop-based human review. Underwater, where GPS is unavailable and acoustic communications are unreliable, power efficiency, inference latency, navigation, and onboard sensing all constrain how the models can be used.
The commercial-to-military hardening gap can turn an expected 10% modification into a 90% change. Encrypting one underwater acoustic message lengthened transmission, kept a DSP operating longer, increased voltage and heat, and ultimately caused a circuit card to fail—forcing changes from software through physical architecture. Haase’s blunt warning to defense-tech vendors: products that appear to work “never actually work” until reliability, cybersecurity, sustainment, and fielding are solved.
A working prototype represents only “5% of the problem” in military deployment. Effective technology is table stakes; professionals ask whether it can survive adversarial attack, meet policy, be fielded, and be sustained. This makes acquisition knowledge and systems engineering durable sources of differentiation even as models and robotics evolve.
Haase’s team should build bespoke capabilities only where the commercial market cannot supply them. Mine recognition requires custom development, but for large language models Haase would rather change the requirement to fit a scaled commercial product than fund a government-specific substitute whose cost rises “exponentially.” Open-source models further accelerate contractors that already understand the mission, shifting value toward safe integration and deployment.
Haase supports offensive AI integration but rejects framing military effectiveness and ethical restraint as alternatives. His standard for taking life is direct, specific human involvement, explainability, oversight, and protection of civilians; after Lukas Biewald pressed the adversarial-arms-race problem, Haase answered that aggression, victory, and moral legitimacy are “an and, not an or.” AI may increase pace and decision advantage, but it should “enable and empower humans, not replace us.”
A technology lead matters, but Haase argues that integration and training matter more than a 10–20% model-performance edge. His pushback on anxiety about DeepSeek R1 and R10 was categorical: “Software has never won a war.” Over the next 5 to 10 years he expects more robots, agents, real-time data, and remote operations, while reserving “about 30% for something we just don’t see yet.”
A promising application is a hierarchy of interoperable agents that clears the fog of war. A vehicle-level agent could help distinguish a mine from a rock and replan a mission; higher-level agents could compress those missions into fleet positioning, with shared data flowing dynamically through the chain of command. Haase sees “completely outsized returns” in making exhausted humans communicate and decide more clearly—and demonstrated internally that visible leadership use can change AI adoption within “a week or two.”
Deep dive
1. Mine hunting makes edge inference operational
Haase described the Navy’s most mature AI application as automated target recognition aboard unmanned underwater vehicles. An ensemble of deep learners, assembled with multiple third parties, runs on an NVIDIA processor to produce targeting cues and bounding boxes for suspected mines while the vehicle is still executing its mission.
The same models run on operators’ laptops, making review of mission data faster and more reliable. Data from new missions returns to the development loop for fine-tuning, reweighting the ensemble, and rapid redeployment—an operating model spanning embedded inference, human analysis, and continuous improvement.
These vehicles are cylinders ranging from about the size of a normal dining-room table to roughly two tables put together, carrying batteries, downward- and forward-looking sensors, navigation software, onboard compute, one rear propeller, and control surfaces. Removing life support makes them far smaller than crewed submarines, but not mechanically simple.
Underwater autonomy begins with brutal constraints: the fluid is “about 1,000 times denser” than air, GPS is unavailable, terrain can cause collisions, and acoustic communications compete with marine noise and temperature-created sound channels. The autonomy is deterministic, but an AI detection can invoke a preprogrammed branch to rescan an area rather than surface, analyze, reacquire, and launch a second mission.
2. Cheap mines can impose billion-dollar downside
Mines may sit on the bottom, within the water column, or near the surface. They are difficult to detect, comparatively low-cost, and capable of stopping fleet operations—“an asymmetric threat” whose importance is easy to ignore because “it’s not a problem until it is, and when it is, it becomes a really big problem.”
Haase reached back to the 1864 Battle of Mobile Bay: after Tecumseh struck a mine, the formation stopped until Admiral David Farragut drove Hartford forward, producing “Damn the torpedoes, full speed ahead.” Haase’s lesson was not romantic bravery but unacceptable risk: survival through that minefield was luck, not effective countermeasures.
He said those ships cost about $17 million in today’s dollars, versus billions for modern ships. Because the ocean is too large to search indiscriminately, commanders must concentrate mine detection where geography, intelligence, known threats, and planned fleet movements make clearance valuable.
3. Lethal autonomy keeps humans at the decision
Mine clearance is ethically straightforward in Haase’s portfolio: it protects naval ships and civilian mariners without asking an AI to take life. On lethal operations, his initial boundary was stark—he “could not imagine offloading that to AI or autonomous systems in any sort of meaningful way” without humans “directly and specifically involved.”
Biewald’s pushback was the arms-race problem: adversaries may apply autonomy without comparable restraints, making an absolute prohibition strategically dangerous. Haase clarified that he was not ruling out offensive AI; pacing adversaries matters, but the design must preserve explainability, oversight, human safety, and no civilian or innocent lives lost.
Haase’s resolution was “an and, not an or”: forces can be aggressive, prevail, and still maintain the moral high ground. Ethical legitimacy also preserves allies and partners, so it is not separate from operational effectiveness but one of its strategic foundations.
4. Cyber hardening turns 10% changes into 90% changes
The defining difference from civilian autonomous vehicles, Haase argued, is an environment of deliberate, repeated adversarial attack. A compromised military fleet might malfunction, become unsafe, or unknowingly serve as a proxy for an unauthorized actor; robustness therefore has to include both visible failures and concealed control.
His candid explanation of why industry underinvests: “Nobody likes cybersecurity.” Hardening produces slower booting, harder authentication, additional hardware and software, encrypted internal messaging, and reduced convenience—“everything would work worse”—even though those burdens are essential against a professional attacker.
His team expected Navy cyber-hardening of a commercial underwater robot to require roughly a 10% change; it became a 90% change. Communications structure, the communications bus, software layers, physical architecture, and the way the vehicle communicated all moved together.
One failure captured the causal chain: encryption enlarged an acoustic packet, keeping the transmitter and DSP open longer; longer transmission applied voltage for more time, generated additional heat, and caused a circuit card to fail. “When defense technology companies or others come to my programs with things that work, they never actually work for us” until these cross-layer consequences are resolved.
5. Defense should buy scale and build only the bespoke layer
Haase sees technology-company relations improving, with startups challenging established processes, more venture capital entering defense tech, and large platforms partnering with government. He cited Google’s movement into government cloud work and Microsoft’s OpenAI-linked services as evidence that vendors increasingly treat defense as a customer.
Cultural translation remains necessary. Vendors must understand combat-grade reliability, while government teams must understand commercial incentives and make acquisition easier; the military position, in Haase’s words, is uncomplicated: it wants “the best tools and capabilities” and “we just want to win.”
For bespoke needs, the process begins with a market survey and government warfare centers: has anyone already produced the architecture, model, data repository, workflow, or even a common schema? If not, the team scopes the requirement and contracts with primes or other industry partners, as it did for mine recognition.
Commercial LLMs reverse that presumption. “We are really better off changing our requirement to meet what’s built at scale” than creating a bespoke substitute when anything close exists. Open-source models also let mission-experienced providers advance quickly, provided they can establish safety, reliability, and system fit.
6. Software enables victory but cannot supply fighting spirit
Asked whether rival militaries keep him awake, Haase distinguished fear from respect: “I’m afraid of no one,” but every adversary can resist unexpectedly, and history shows great powers struggling against smaller opponents. Technological humility means learning from everyone without treating a benchmark lead as destiny.
Biewald argued that another country holding an AI lead of several years could be frightening. Haase’s direct rebuttal—“Software has never won a war”—did not dismiss technology; it rejected the idea that a single technical dimension can replace people, human factors, training, and the ability to combine capabilities.
Haase’s John Boyd example was that aircraft with weaker specifications could prevail when hydraulics let pilots operate them more effectively. Better explosives per pound pose the same analytical trap: energetics matter, but remain only one contributor inside a much larger fighting system.
AI’s strongest role is therefore decision advantage—lifting cognitive burden, surfacing signal through noise, and helping people act at the right moment. Haase invoked Admiral Nimitz’s misunderstood message to Admiral Halsey: superior technology, intelligence, positioning, and surprise produced no attack because Halsey interpreted “attack” as “be ready to attack.”
7. The next battlefield will be connected, faster, and surprising
Over the next 5 to 10 years, Haase expects more robotic systems, remote operations, real-time information, and communication among machines, humans, and decision-support agents. The pace of operations should rise as battlefield data flows more quickly across formerly separate systems.
Responding to the Ukraine example, he noted that a nation-state with America’s full technological scope has not yet put all of those resources behind such operations. Unmanned systems are a durable trend, but how they interact with humans and one another remains unresolved.
Every major military technology—from aviation and rifling to gunpowder—has produced battlefield surprises. His forecasting rule is therefore to extend observable trends but “leave about 30% for something we don’t see yet.”
8. DoD LLM adoption is a policy and workflow problem
Haase said the Department of Defense is cautiously piloting access through offerings involving Gemini, Anthropic, and Microsoft at different impact levels. Near-term applications are deliberately low-threat: paperwork, awards, HR, document generation, review, requirements, and developer productivity, with faster cycles and better first drafts.
Core military use creates a mismatch with general-purpose safety policies. Mission planning, battlefield casualties, and assaults on positions may trigger safeguards designed to stop ordinary users; DoD needs closer partnerships so legitimate operational work can proceed without discarding information protection, ethics, or policy compliance.
Biewald raised the risk that LLMs make bioweapon or other lethal knowledge easier to obtain. Haase noted that much of the information already exists, but agents lower access friction and might divide a plan across several chat windows; he had no complete answer and put continuing responsibility on vendors’ trust-and-safety teams.
DeepSeek R1 and R10 captured attention but did not change Haase’s immediate priorities. His race analogy was to hit one’s own marks: integrating, training with, and teaching people to use available U.S. systems matters more today than a “10 or 20% competitive advantage” or a 20% improvement on an ARC benchmark.
9. Visible leadership and interoperable agents attack the human bottleneck
Haase’s team initially hesitated to use AI because employees feared it would be perceived negatively. Citing research suggesting something like a 25% improvement among consultants using AI, he began visibly querying an assistant during meetings, turning his screen toward colleagues and reading the response with them.
“Within 1 or 2 weeks,” senior leaders copied the behavior and began submitting better-organized, better-reviewed work. His emerging research question is how leaders create psychological safety for AI adoption while retaining individual accountability through demonstrations, training, enthusiasm, and integration into existing workflows.
One practical habit was transcribing every Teams call and organizing the text for possible natural-language processing and retrieval-augmented generation. Spoken video compresses dramatically into text; a controlled repository could give future agents access to institutional conversations without requiring the team to reconstruct its history from scattered program documents.
Haase’s next research frontier is multi-agent collaborative autonomy: one agent helps a 4-person boat team retask a mine-search mission, another aggregates activity across ships, and higher agents help decide where to place ships across the battlefield. Shared data would support different functions—from “mine or a rock” classification to battlefield-level decisions—with data interoperable up the chain of command.
A military-tailored LLM could also prevent another Nimitz-Halsey failure by letting the recipient query the same agent that wrote the mission. The payoff is clearing the “fog of war” when people are exhausted and messages are broken: “Technology is meant to enable and empower humans, not replace us.”
Haase closed by translating the logistics maxim into defense technology: “Amateurs get excited about working prototypes.” A functioning demo is “5% of the problem”; professionals focus on sustainment, fielding, cybersecurity, and policy compliance because “an effective system is table stakes, but it doesn’t get you all the way there.”