Nate Ranson
SECR-6914
Response 2

“Allies and Artificial Intelligence: Obstacles to Operations and Decision-Making”

In the article “Allies and Artificial Intelligence[...]”, Lin-Greenberg lays out the complicated landscape of integrating an artificial intelligence platform between formal and informal alliances. Lin-Greenberg mentions a number of potential issues ranging from competitive or mis-aligned goals for partners to outright sabotage by malicious actors. To push on the analysis further, this response attempts to extend the analysis of these challenges by providing examples of the recent past where high-trust systems have broken down because of lack of pushback – also known as “automation bias” – and how the nondeterministic outcomes of AI might impact battlefield coordination. Additionally, the rise of AI provides the possibility of new, physical, attack vectors for combatants and, finally, further press on Lin-Greenberg’s repeated warnings around compressed, escalatory timelines.

AI, as it’s understood today, has had a slow evolution through the use of increasingly automated systems. These systems range from rudimentary early-warning systems to the futuristic and – to this point – fictional, “slaughterbots.” As these systems have further been embedded in everyday routines, reliance on their accuracy grows, feeding into what science refers to as an “automation bias,” (Center for Security and Emerging Technology, 2024). This bias causes operators to trust the output of the system even when the output is lacking in context or, worst case scenario, disputing the reality of the situation.

Any military operation, offensive or defensive, is necessarily a high-trust environment. But, as these automated – or autonomous – systems are shared across allies, questions arise about how to handle this automation bias. The Patriot missile friendly-fire incidents were the culmination of tense, high-stakes environments where operators trusted the system's lethal recommendations, even when the radar alerts didn't match the actual flight signatures of tactical missiles, leading to disaster (Ricks, 2004). This reinforces Lin-Greenberg’s questions of interoperability in a potentially asymmetrically-burdened AI environment. Automated systems keep humans in the loop by controlling for a fixed set of variables. Oversimplified, it can be understood as a series of “if this, then that” routines to provide information to the human beings at the control panel. If coalition forces can’t adequately coordinate to avoid lethal deferential mistakes in automated, mechanical Patriot missile systems, what does the future look like with a nondeterministic AI-generated output?

AI systems differ from automated systems in that they can be nondeterministic. Unlike an automated system, the logic used to arrive at a conclusion may not be immediately auditable or exactly repeatable. An AI interface provided the same set of variables to solve the same problem can output mildly-to-wildly different results (Jackson, 2025). If operators are already predisposed to defer to automated systems, how will allies reconcile two different solutions? This further complicates the relationship between allies who may or may not be entrenched in equally advanced AI systems. One ally's system may recommend immediate action while another's recommends restraint; who has the final say? Does AI superiority always win out? Can allies negotiate successfully during the fog of war? These are questions to be answered by Lin-Greenberg’s proposed solutions of cooperative agreements with AI experts bridging these gaps ahead of time. Due to the United States’ geographic advantage, the burden of risk may not be evenly shared.

American allies in the Persian Gulf have certainly borne that risk in the recent Iranian conflict. AI needs datacenters to train models and datacenters need to be in relatively close geographical proximity to the battlefield to deliver these responses with limited latency. When Iran struck back against American allies and noncombatants in the region, one of the stated targets of retaliatory Iranian missile strikes was datacenters (Just Security, 2026). This hybrid infrastructure supports both civilian and military purposes, meaning in future wars, it can become a viable target for competing AI powers wanting to degrade the others’ systems (James, 2026). This could drag non-belligerent parties into a conflict just due to proximity or as a host to valuable infrastructure. The physical cost of hosting these datacenters will fall disproportionately on those allies who are geographically closest to the conflict.

The most worrying piece of Lin-Greenberg’s analysis is the speed at which AI-powered conflicts could accelerate. It’s comforting that even in the height of a conflict, two traditional forces could discuss solutions voice-to-voice before an escalatory round of strikes. In an AI-powered war, the speed at which decisions are made becomes compressed, allowing for escalatory rounds of tit-for-tat violence before humans on the loop are able to understand the breadth of the decisions being made. This hyperwar could have devastating consequences to the direct combatants as well as their allies (Big Think, 2023). Additionally, AI models are prone to hallucinations or unexpected outputs. Google’s AI used to defeat a world champion Go player made a bizarre move, dubbed “Move 37,” in one of their games (BBC News, 2016). The stakes for Go are fairly low, but with autonomous weapons and systems, how should systems react to an errant drone? Automation bias might sway an observer into believing that they are witnessing a masterclass of strategy. Nondeterministic outcomes can turn into lethal outcomes when decision making is faster than human reaction times. Is it also possible that systems would confidently assert incorrect solutions to a rapidly changing environment, much like Watson repeating an incorrect answer on Jeopardy (NBC News, 2011)?

To set up this hypothetical conflict, what if an adversarial country’s AI drones made an incursion and subsequent attack on an American ally? That ally’s US-provided AI system could intercept and/or retaliate in kind, which could spark a spiral of violence. Drone incursions are unlikely to cause disproportionate retaliatory violence in human-controlled conditions, but AI models might weigh an unknown number of other variables in its decision to respond. Given the current understanding of AI and their “thinking” process, this decision to retaliate would be opaquely arrived at, perhaps dragging other adversaries into a conflict. Under the worst conditions, a domino effect of cascading alliances could kick off a World War I scenario, but at the speed of compute.

Lin-Greenberg lays out the fraught landscape of the future of AI and alliances. The future is unwritten, but with the speed at which change happens in the AI field, countries should be wargaming out these scenarios seriously. Much like planting a tree, the best time to figure out the AI technology-sharing roadmap was 10 years ago; the second best time is today. AI world leaders and their allies alike need to delegate experts in the field to negotiate an agreed upon set of rules and operator precedence. Even loosely agreed-upon rules of engagement would help all combatants navigate the fog of war.

Bibliography

  • BBC News. Google AI Defeats Human Go Champion. March 2016. https://www.bbc.com/news/technology-35761246
  • Big Think. "Hyperwar": How AI could cause wars to spiral out of human control. February 2023. https://bigthink.com/the-future/hyperwar-ai-military-warfare/
  • Center for Security and Emerging Technology. AI Safety and Automation Bias. November 2024. https://cset.georgetown.edu/publication/ai-safety-and-automation-bias/
  • Jackson, J. The New Stack. Martin Fowler on Preparing for AI's Nondeterministic Computing. December 2025. https://thenewstack.io/martin-fowler-on-preparing-for-ais-nondeterministic-computing/
  • James, L. Tom's Hardware. Iranian drone strikes hit three AWS data centers in the UAE and Bahrain. March 2026. https://www.tomshardware.com/tech-industry/drone-strikes-hit-three-aws-data-centers-in-the-uae-and-bahrain
  • Just Security. Iranian Attacks on the Amazon Data Centers: A Legal Analysis. March 2026. https://www.justsecurity.org/133685/iranian-attacks-amazon-data-centers-legal-analysis/
  • Lin-Greenberg, E. 2020. Allies and Artificial Intelligence: Obstacles to Operations and Decision-Making. Texas National Security Review, Volume 3, Issue 2, Spring 2020, pp. 56-76.
  • NBC News. Beyond Jeopardy: Watson wins. February 2011. https://www.nbcnews.com/science/cosmic-log/beyond-jeopardy-watson-wins-flna125228
  • Ricks, T. E. The Washington Post. Investigation Finds U.S. Missiles Downed Navy Jet. December 2004. https://www.washingtonpost.com/archive/politics/2004/12/11/investigation-finds-us-missiles-downed-navy-jet/323e76f1-31d5-49c1-a27a-df9f389b0532/

Back.