Army kicks off generative AI pilot to tackle drudge work, hallucinations

The Army is starting a new pilot to support a broader push toward the service’s adoption of generative artificial intelligence capabilities, officials announced Friday.

The effort, dubbed #CalibrateAI, comes on the heels of other efforts — such as #DefendAI, #BreakAI and #CounterAI — that have been launched as the service pursues a 500-day implementation plan for adopting these types of technologies and overcoming some of the challenges associated with them.

“Gen AI models present unique and exciting opportunities for the Army. These models have the potential to transform mission processes by automating and executing certain tasks with unprecedented speed and efficiency. Commanders and senior leaders should encourage the use of Gen AI tools for their appropriate use cases,” Army CIO Leonel Garciga said in a directive issued to the department in June.

Generative AI systems can create new content — such as text, images, audio and video — based on the data they’ve been trained on and human prompts. ChatGPT, a popular large language model, is one prominent example.

While a lot of focus in the defense community is on how artificial intelligence could aid and enable U.S. military operations on the battlefield, officials are also keen on employing the technology for back-office functions and to handle some of the drudge work that humans have previously had to perform.

On Friday, Jennifer Swanson, deputy assistant secretary of the Army for data, engineering and software, announced the launch of #CalibrateAI, which will “explore innovative applications” of the tech for Army acquisition activities.

The initiative will use a “cutting-edge” capability developed by the non-profit LMI at no cost to the Army, according to officials.

The tool, which is intended to “simplify repetitive and time-consuming tasks,” will use data analytics, machine learning and natural language processing to “deliver tailored responses that are highly relevant to specific topics, improving the efficiency and effectiveness of information retrieval and analysis,” the Army stated in a press release.

However, officials are also concerned about ensuring the security and accuracy of the Defense Department’s data and the outputs of gen AI tools.

In his June memo, Garciga noted that these technologies present unique challenges with regard to data privacy, security and control over the generated content. “Therefore, their use should be carefully evaluated and monitored,” he wrote.

The #CalibrateAI capability will be deployed in an Impact Level 5 (IL5) secure cloud environment and be able to handle controlled unclassified information (CUI) data. It includes customizable user-access controls to protect “need-to-know” info, per the release.

The Army is worried about potential hallucinations, a trend seen in other generative artificial intelligence systems that have been launched in the commercial sector. For example, large language models sometimes yield responses to user inputs that are factually inaccurate or otherwise problematic.

Recognizing these potential pitfalls, the new pilot will also focus on “identifying and calling out potential ‘hallucinations’ or erroneous outputs, thereby increasing the reliability of AI-generated content,” according to the release.

“By using off-the-shelf AI tools and leveraging cross-service authority-to-operate reciprocity granted by DoD CIO, #CalibrateAI will explore how we increase productivity while enhancing the accuracy of information,” Swanson said in a statement. “The ability to query curated document sets for generating new content, along with providing citations, will ensure that our outputs are not only accurate but also easily fact-checked.”

The broad goals of the pilot for the acquisition community include increasing productivity, improving accuracy, promoting innovation, and identifying a cost-effective route for wider adoption of the technology. According to officials, that will be achieved by using AI tools to collate, curate and generate critical information relevant to acquisition activities, implementing mechanisms to provide citations, and promoting the exploration of novel applications of artificial intelligence in acquisition.