In today’s world, artificial intelligence chatbots like ChatGPT and Cloud can perform many tasks, like writing work emails and planning itineraries. These chatbots are systems built around large vision-language models (VLMs): AI trained on a huge dataset that includes books, websites, code, and images.
AI algorithms are refined heavily based on human-generated feedback to follow instructions and avoid harmful or unwanted outputs, and use that “knowledge” to produce text or images based on input from the user. Although chatbots have obvious limitations, they can be very helpful for a wide variety of tasks, including some areas that traditionally require specialized skills, such as computer programming.
As part of a project for the U.S. Department of the Air Force-MIT AI Accelerator’s Fantom program, U.S. Air Force cadet Joshua Lynch – with the help of his mentor, Laura Nieves, a technical staff member in the Embedded and AI Systems group at MIT Lincoln Laboratory – wanted to determine whether, as a complete novice to coding, he could develop a fully functional program. They used a process called “vibe-coding”, in which a user relies entirely on prompts to guide a generative AI chatbot to write and refine code.
Their motivation was to empower anyone familiar with a military problem area, regardless of their technical background, to pursue their ideas for useful software applications, essentially bypassing the time and cost constraints of the traditional military software development pipeline. Lynch’s goal was to create his own application while Nieves monitored his experience with the technology.
“The Phantom student wanted to see if he could create a useful application through self-identified vibe-coding without any prior experience,” says Nieves. “Within this project, I wanted to understand how their perceptions of AI changed with use over time. We both wanted to better understand where and how AI could be used by non-technical users in the military.”
Lynch set out to see if, with no coding skills and using chatbots, he could create applications specific to his type of tactical team to help reduce collateral damage while increasing survivability in broader missions. The application will provide capabilities including AI-assisted target identification; modular intelligence, surveillance and reconnaissance; autonomous strike; and battlefield communications management.
During the project, Lynch completed several professional development courses in AI and familiarized himself with both military and non-military uses of the technology. To base their code generation, they used paid models of three AI chatbots: Anthropic’s Cloud, OpenAI’s ChatGPT, and Google’s Gemini. Much of this work was done only through the chatbots’ main chat function on a web browser, rather than as an integrated system within the development environment, as is now the standard. The final application was created using the Google AI Studio app, which can create applications that interface with the Gemini application programming interface and have the AI integrated into the development environment.
Over three months, Lynch worked with these models to create his application called Remote Operating Modular Augmentation Device (ROMAD-AI). During this time he learned many ways to improve code output. For example, they often encountered difficulties in AI chatbots lacking hierarchical focus and modifying unrelated code sections. He found that it is important to break problems into smaller parts, formulate questions clearly, and bring them back on topic when the conversation strays too far from the objective.
Recognizing the limitations of chatbots and learning to work around them effectively took up most of the project’s time. As Lynch gained more experience with chatbots, limitations in AI capabilities and lack of time for development led him to retool the project, moving it from an application that could assist on the battlefield to one that could perform basic document processing, such as analyzing tactical maps of battlefields and generating mission-planning documents through an interface with a VLM-powered chatbot. While the resulting prototype did not perform all of the capabilities originally envisioned by Lynch (and was not safe for the desired use case in its current iteration), it proved the potential and utility of such an application for service members.
“I was impressed with the final product and it showed me how powerful these systems can be in helping non-experts design prototypes,” says Nieves. “Now my opinion is that these can be powerful tools for non-technical experts to help explain problems and possible solutions to technical experts and communicate desired outcomes.”
Nieves observed a change in perspective on Lynch’s AI language model during his experience. Starting with an impressive goal, Lynch gained an understanding of the capabilities of the current technology and significantly reduced his expectations by the end of the project period. Measures of their perception of different AI systems over time and across system updates were particularly interesting to Lynch and Nieves, with Cloud showing greater stability than ChatGPT in traits such as likeability, anthropomorphism, and perceived intelligence. Lynch found the AI to be a helpful teacher, but noted its inaccuracies on subjects he knew well.
The project showed that AI chatbots can empower non-technical service members to produce viable software applications for their unique problems, although it works better as a prototyping assistant than a full production tool for handling sensitive information and critical applications. Improper checking of code can lead to security risks, as shown by an example where Lynch did not realize that the final application was sending input documents to the Gemini AI model for analysis rather than parsing the documents locally on her computer. Although AI can generate significant amounts of functional code, code review remains a hurdle in this area.
“For me, this project reinforced the breadth between experts in different fields,” says Nieves. “No matter how good AI gets, I think we will always need to collaborate to get the best solutions to the most important problems.”
The research was sponsored by the Department of the Air Force Artificial Intelligence Accelerator and was completed under Cooperative Contract Number FA8750-19-2-1000.
