Generative AI is starting to change who can participate in software design. A project involving U.S. Air Force cadet Joshua Lynch and MIT Lincoln Laboratory mentor Laura Niss tested a practical question: can someone with no coding background use chatbots to build a useful military AI application?
The answer was not a simple yes or no. Lynch did create a working prototype, called the Remote Operating Modular Augmentation Device (ROMAD-AI), but the project also showed why chatbot-generated code still needs technical oversight, especially in sensitive or critical settings.
What the project set out to test
The work was part of the U.S. Department of the Air Force–MIT AI Accelerator's Phantom Program. Lynch wanted to see whether a complete novice could develop a functional program through “vibe-coding,” a process where the user relies on prompts to guide a generative AI chatbot as it writes and revises code.
The larger motivation was practical. People who understand a military problem may not know how to code, while the traditional software development pipeline can involve time and cost constraints. Lynch wanted to explore whether chatbots could help nontechnical users turn domain knowledge into early software prototypes.
Niss, a technical staff member in the Embedded and AI Systems Group at MIT Lincoln Laboratory, monitored the experience and studied how Lynch’s view of AI changed as he worked with the tools.
"The Phantom student wanted to see if he could create a useful application through self-identified vibe-coding, without any previous experience," Niss says. "Within this project, I wanted to understand how his perception of AI changed over time with use. We both wanted to understand better where and how AI could be used by nontechnical users in the military."
From battlefield ambition to document processing
Lynch’s initial goal was highly ambitious. He wanted to build an application for his type of tactical team that could help reduce collateral damage while enhancing survivability in the broader mission. The planned capabilities included AI-assisted target recognition; modular intelligence, surveillance, and reconnaissance; autonomous striking; and communication management on the battlefield.
During the project, he completed several professional development courses in AI and studied both military and nonmilitary uses of the technology. For code generation, he used the paid models of three AI chatbots: Anthropic's Claude, OpenAI's ChatGPT, and Google's Gemini.
Most of the work happened through the chatbots’ main chat function in a web browser, rather than inside an integrated development environment. The final application was produced using Google AI Studio App, which can create applications that interface with the Gemini application programming interface and includes AI in the development environment.
Over three months, the original scope changed. Lynch found that current AI capabilities, development time and practical limitations made the initial battlefield application unrealistic. The project was narrowed to a tool that could perform basic document processing, including analyzing tactical maps of battlefields and generating mission-planning documents through an interface with a VLM-powered chatbot.
That narrower version did not deliver all the capabilities Lynch originally imagined. It also was not secure for the desired use case in its current form. Still, the prototype showed how a service member could use AI to turn a problem idea into a functional software concept.
What Lynch learned about coding with chatbots
The project’s most important work was not only the code. Much of the timeline was spent learning how to manage the chatbots’ limitations.
Lynch found that the systems could lose hierarchical focus and modify unrelated code sections. He learned that the best results came from treating the process as guided iteration rather than a single request. Smaller tasks, clear framing and repeated course correction helped keep the chatbot aligned with the objective.
Several lessons emerged from the process:
- Break software problems into smaller parts before asking the model to solve them.
- Ask clear, bounded questions instead of broad requests.
- Watch for changes to code sections unrelated to the immediate task.
- Redirect the conversation when the chatbot drifts away from the goal.
- Expect the user to learn the tool’s failure modes over time.
Lynch also found AI useful as a tutor, while noticing inaccuracies on subjects he knew well. That combination matters: a chatbot can help a novice move faster, but it can also produce output that feels plausible while still needing review.
Why prototyping is different from production
The project showed real promise for military AI prototyping. Niss said the final product changed how she thought about nonexperts using these systems to communicate software ideas.
"I was quite impressed with this final product, and it showed me how powerful these systems can be at prototyping designs from nonexperts," Niss says. "I'm now of the opinion that these can be powerful tools for nontechnical experts to convey problems and possible solutions to technical experts, and aid in communicating desired outcomes."
That is a narrower claim than saying chatbots can replace software teams. The source project points in the opposite direction for critical applications. AI chatbots can generate significant amounts of functional code, but code review remains a bottleneck.
Security was one of the clearest concerns. In one instance, Lynch did not realize that the final application was sending input documents to a Gemini AI model for analysis rather than parsing them locally on his computer. For sensitive information, that distinction is not a minor implementation detail. It changes how the system handles data and what risks may exist.
The project therefore frames vibe-coding as a bridge, not an endpoint. It can help nontechnical service members describe problems, explore interfaces and produce working demonstrations. But when the application involves sensitive information or critical use, technical review and collaboration remain essential.
The deeper lesson for military AI
Lynch’s view of AI shifted as he moved from a large initial goal to a more limited prototype. Niss observed that his expectations became more grounded as he encountered the capabilities and limits of current tools.
The project also compared perceptions of different AI systems over time and across system updates. Claude showed more stability than ChatGPT across traits such as likeability, anthropomorphism, and perceived intelligence.
The lasting implication is straightforward. AI chatbots can give nontechnical experts a new way to express ideas in software form, including in military contexts. But the strongest result comes when domain experts and technical experts work together, using prototypes to clarify the problem rather than treating generated code as finished infrastructure.
"For me, this project reinforced the expanse between experts in different fields," Niss says. "No matter how good AI gets, I think we'll always need to collaborate to get to the best solutions for the most important problems."