As the U.S. Air Force begins to integrate artificial intelligence into its operational planning, service leaders are encountering significant practical limitations. While AI can generate military strategies in seconds, recent experiments show these plans often lack viability, highlighting an ongoing need for human oversight and exposing deep-seated challenges in infrastructure, data management, and personnel.
Key Takeaways
- An AI system created military plans in 10 seconds, compared to 16 minutes for a human, but the suggestions were often impractical.
- All forms of AI, not just generative models, are prone to issues of accuracy and bias, requiring careful management.
- The Air Force identifies fragmented technology infrastructure, data silos, and a shortage of skilled personnel as major barriers to effective AI deployment.
- Current strategies involve using AI for narrow tasks and segmenting systems to limit potential failures, ensuring human operators can verify the results.
AI Speed vs. Practical Reality
During a recent experiment conducted by the Advanced Battle Management System (ABMS) Cross-Functional Team, an AI algorithm was tasked with generating Courses of Action (COAs), which involves deciding how to use weapons and platforms against a target. The AI produced a COA in just 10 seconds, a task that took a human operator 16 minutes.
Despite this remarkable speed, the results were not always usable. Maj. Gen. Robert Claude, the Space Force representative to the team, noted that the AI-generated plans were not “necessarily completely viable.” He explained that while the AI produced more options in less time, it sometimes failed to consider critical factors.
For example, the system proposed using infrared-guided weapons during cloudy weather, a scenario where such systems would be ineffective. This illustrates a critical gap between processing speed and real-world situational awareness.
“We’re getting faster results and we’re getting more results, but there’s still going to have to be a human in the loop for the foreseeable future to make sure that, yes, it’s a viable COA,” Claude stated.
Inherent Flaws in AI Models
The challenges extend beyond specific operational errors. Industry experts at the AFA’s Air, Space & Cyber Conference emphasized that all AI models have inherent limitations. While generative AI is known for its tendency to “hallucinate” or invent information, other types of AI face different but equally serious problems.
David Ware, a partner at the consulting firm McKinsey, clarified that accuracy and bias are universal concerns. He warned that training an AI on historical data can embed outdated or flawed decision-making into its logic.
The Problem of AI Bias
According to Ware, if a targeting model is trained exclusively on past information, its recommendations will be biased toward those prior conditions. “It’ll still have an accuracy issue that I have to overcome,” he said, highlighting the need for continuous validation and updated training data.
To manage these risks, experts recommend carefully controlling the environment where AI operates. Ryan Tseng, president of Shield AI, advocated for a segmented technical architecture, especially in critical systems like aircraft.
“Flight and other critical systems... should be segmented off, with the LLMs and very complex autonomy in another part of the system,” Tseng explained. This approach effectively “helps keep the AI in a box,” limiting the potential damage from an erroneous AI decision.
Infrastructure and Scalability Hurdles
Beyond the models themselves, the Air Force is wrestling with foundational infrastructure issues that prevent the widespread and effective use of AI. Maj. Gen. Luke Cropsey, who leads the command, control, communications, and battle management (C3BM) program, described this as one of his greatest challenges.
Cropsey’s program aims to unify approximately 50 different systems into a coherent “battle network.” He pointed to the severe fragmentation of the current technological landscape as a primary obstacle.
“I have literally independently owned and operated [technology] stacks all over the place,” Cropsey said, noting that incompatible data formats and lack of interoperability are among his most difficult problems to solve.
Maj. Gen. Michele Edmondson, Deputy Chief of Staff for warfighter communications and cyber systems, echoed these concerns, identifying scalability as a fundamental issue. The problem is not just a lack of hardware but a complex web of interconnected systems.
What is 'Edge AI'?
Edge AI refers to running artificial intelligence algorithms on devices at the front lines, away from centralized data centers. This is crucial for military operations where connectivity may be limited. However, it requires powerful, compact computer chips and robust data systems that can function in harsh environments, presenting a significant technical challenge.
Ware added that infrastructure encompasses hardware, software, and the systems needed to manage data and tools, particularly in classified environments where vendor equipment and proprietary data formats create additional barriers.
The Human Element: Skills and Data Integrity
Ultimately, the success of AI in the military depends on people and data. According to Maj. Gen. Edmondson, there is a significant shortage of personnel with the right expertise. “We need more people with the right skill sets,” she stated.
The Air Force has established a new data analytics career field, but Edmondson called it a “very small core subset” of what is needed. She stressed the importance of leveraging the innate digital literacy of new recruits.
“We have got to capitalize on that, and we have got to upscale them throughout their career, so that we continue to build on the skills they will bring with them,” she argued.
Data itself remains a critical chokepoint. Edmondson highlighted ongoing struggles with data integrity and the ability to integrate and share information effectively, especially at the operational edge. Without reliable, accessible, and high-quality data, even the most advanced AI algorithms cannot perform effectively.
