AI ASSISTANT: Solano used ChatGPT to help write and edit this article. Some acquisition professionals are already using language models to complete tasks in their daily personal lives. (Photo by Zaira Solano)
The opportunities and risks of advanced language models.
by Lt. Col. Robert Solano
As artificial intelligence continues to advance, large language models like ChatGPT have the potential to revolutionize the way defense acquisition and contracting are performed. With the ability to generate human-like text, language models can automate many of the repetitive and time-consuming tasks involved in procurement, such as document preparation, research and communication. As with any new technology, there are also risks associated with the adoption of large language models in the defense industry. These risks include potential security breaches, bias in decision making and unintended consequences. In this article, we will explore both the opportunities and risks associated with the use of advanced language models like ChatGPT in defense acquisitions and contracting.
ChatGPT is an artificial intelligence language model, which is a computer program designed to generate text based on user input. Although there are many language models, ChatGPT has received the most attention recently for its ability to generate human-like text with remarkable accuracy. ChatGPT was developed by OpenAI, a research organization dedicated to creating and promoting friendly artificial intelligence. OpenAI trained the model using an artificial intelligence technique called reinforcement learning from human feedback. This training process involved inputting large amounts of data into the computer program and then providing the program feedback and adjustments to improve its performance. ChatGPT was trained on a substantial amount of text estimated to be more than 300 billion words.
ChatGPT’s ability to understand and generate coherent and meaningful text has made it a popular tool, attracting more than one million users in just one week after its November 2022 release. By January, it was estimated to have reached 100 million monthly active users making it the fastest-growing consumer application in history.
HOW ADVANCED LANGUAGE MODELS WORK
Advanced language models like ChatGPT are developed using sophisticated artificial intelligence computer programing techniques. Developers feed vast amounts of text data into computer programs that analyze and process the information. The data comes from a variety of sources, such as books, articles and web pages. The output of the computer program is another type of computer program called a language model which is capable of understanding and generating human text.
Teams of data scientists and engineers then adjust the model through a process called training and fine-tuning. They adjust the data and the parameters that the model uses to make calculations, so that it can make increasingly more accurate predictions or decisions on new text data. They repeat the process millions of times until it can accurately predict the next word with high probability. Once the model is trained and fine-tuned, it can be used to perform various language tasks. For example, it can be used to generate coherent and realistic text, answer questions and summarize large amounts of text.
Most users access ChatGPT through a simple and free text prompt on a web browser, while companies like Microsoft have begun integrating AI into their software services. Microsoft recently announced a multiyear, multibillion-dollar investment in OpenAI and will deploy ChatGPT models across its consumer and enterprise products like Bing search engine and Microsoft Teams. This investment is estimated to be valued over $10 billion and highlights the growing demand for advanced language models.
OPPORTUNITIES AND USE CASES
Although ChatGPT is currently the industry leader in language models, other companies are catching up. In the near future, we can expect greater access to these language models and more integration with many of the software tools that we are already familiar with. As they become more advanced, language models also will be tailorable to specific styles, genres and domains (like DOD acquisition). Some people speculate that language models will become the “calculators of writing.” They predict that, similarly to how we rely on calculators to do most of our mathematical calculations today, in the future we will rely on language models to do most of our writing.
I recently surveyed a group of acquisition officers to gather their perspectives on how language models could be employed in defense acquisitions. Their feedback proved to be informative, with several noteworthy insights shared. For instance, some expressed interest in integrating AI or ChatGPT into the contract writing system to help draft performance statements of work or to write requirements. Others felt that ChatGPT could aid in market research to identify innovative solutions for addressing capability gaps. Another potential application mentioned was using ChatGPT’s analytical power to develop realistic and attainable requirements and criteria for source selection. Furthermore, some suggested that ChatGPT could be used to assist with writing standard forms, including single acquisition management plans and test and evaluation master plans, as well as to help foreign vendors navigate systems like SAM.gov and Procurement Integrated Enterprise Environment.
It is worth noting that some acquisition professionals are already leveraging ChatGPT to increase their productivity, albeit not for official purposes. They are using it for personal projects, research and idea generation. One officer used ChatGPT to find recent Government Accountability Office reports and to simplify technical language into more accessible terms. Another officer used ChatGPT to create a notional statement of objectives for a potential prototype project, which they found to be highly useful. One officer’s spouse even used ChatGPT to write her resume, which turned out better than if her partner had written it. Finally, many appreciated ChatGPT’s simplicity and efficiency in providing quick, accurate answers.
I have personally used ChatGPT to assist in writing articles like this one, holiday messages to my workforce, promotion speeches, social media posts, sympathy cards, love notes to my wife and even a book. This saved me hundreds of hours and resulted in better products than I could have ever done myself. My personal experiences and the previous comments highlight that ChatGPT and similar language models have huge potential to transform the nature of writing in defense acquisition.
Despite the potential opportunities and use cases, there are also risks associated with the adoption of large language models in defense acquisition and contracting. The first major risk is the handling of confidential or sensitive information, as language models are not specifically designed or tested for safeguarding controlled or classified data. This can result in security breaches or the dissemination of inaccurate information.
If language models are used to generate government documents, it will also be important to review existing policies regarding records retention laws. These policies dictate how long certain types of records should be retained. The use of language models may change the way documents are created and stored, potentially impacting compliance with these policies.
Language models also can generate text that contains biases, inaccuracies or other errors that could compromise the government’s credibility or integrity. This is why any text generated by language models must be carefully reviewed and validated before use. As a result, network security policies often block access to ChatGPT and similar language models, slowing its adoption in the defense industry. However, some companies are looking into customizing language models specifically for government teams by fine-tuning their models with vast amounts of government-related data. Despite still being in early development, the integration of ChatGPT into Microsoft products like Office 365 and Microsoft Teams will likely be the first time many government users encounter language models.
Outside of government offices, the widespread adoption of language models by industry partners is expected to improve their efficiency in program management and contracting operations. The use of language models can significantly improve the speed and efficiency of administrative tasks, such as drafting meeting minutes and submitting contract proposals. The automation of information gathering and proposal writing can help level the playing field for small businesses and nontraditional contractors, who may not have the same resources as larger contractors.
Language models could make it easier for vendors to respond to government solicitations, leading to a greater volume of bids and more desirable competition. The increase of solicitations received by the government may overwhelm the already short-staffed contracting workforce. As such, it is important for the government to start considering how to effectively manage the potential influx of proposals and ensure the evaluation process remains fair and thorough.
The use of language models by offerors also raises some risks in the contracting process. Offerors can use language models to optimize their responses to government requirements and increase their chances of winning contracts, giving an advantage to vendors who use language models. This may lead to contract awards going to offerors who have the best artificial intelligence models, rather than those who provide the best value to the government.
To address these risks, our contracting professionals will need to have a firm understanding of how language models compile and present information. They also will need to use the same language model technology to assist in market research, summarizing lengthy proposals or identifying risk areas. As offerors become more efficient with the response process, our contracting professionals will need to leverage resources to maintain an equal level of efficiency.
Large language models have the potential to improve efficiency and productivity across various industries, including the government sector. While the widespread adoption of these models in the government may lag behind the commercial sector, industry partners are likely to adopt the technology first and use it to assist with proposal writing and contracting processes, making it easier for offerors to respond to government solicitations and leveling the playing field for small businesses and nontraditional vendors. It is important for acquisition professionals to consider both the opportunities and risks involved and to use the technology responsibly and securely.
For more information, follow Lt. Col. Robert Solano on LinkedIn at https://www.linkedin.com/in/therobertsolano.
LT. COL. ROBERT SOLANO is the commander of the Defense Contract Management Agency at Boeing in Mesa, Arizona. His prior roles include program manager at the Army Artificial Intelligence Integration Center and Training With Industry Fellow at Palantir Technologies, where he developed artificial intelligence systems. He holds an M.S. in aerospace engineering from the Georgia Institute of Technology and a B.S. in mechanical engineering from the United States Military Academy at West Point. He holds the DAWIA Advanced certification in program management.