Former title: Development status of generative AI in the field of command and control of the US military
Since 2023, the generative AI technology represented by the big language model has made breakthrough progress, and new products and technologies have been unveiled, triggering a new wave of AI development. With the successful application of generative artificial intelligence systems in the commercial field, the U.S. military has realized that generative artificial intelligence has great application potential in military fields such as command and control, intelligence analysis, combat planning, and auxiliary decision-making. Therefore, it has begun to promote the military application of generative artificial intelligence technology.
The role of generative AI in command and control
Generative AI can provide users with answers to questions covering a wide range of fields, help users write articles, emails and other text content, generate concise summaries for long articles according to user input, help users quickly understand key information, carry out machine translation between multiple languages, and automatically generate code in multiple programming languages on demand, Real time online customer service, etc. The U.S. Department of Defense believes that generative AI technologies such as large-scale language models have great potential applications in military fields such as intelligence analysis, battlefield environment support, communication and network security, situational awareness and data fusion, military data management, military modeling and combat simulation, Therefore, the US military urgently needs to improve the speed and interaction ability of the generative AI model to improve the task efficiency. According to the main capabilities of the generative AI system, its application to battle command has the following significance:
First, it can make operational data support more accurate. According to the specific combat task, the commander's intention is accurately understood through the generative AI systematic service, and the data needed for the battle command planning are quickly and accurately queried in the mass of various types of combat information data.
Second, it can shorten the iteration cycle of command information and data. Through the interactive service mode provided by the generative artificial intelligence system, the warfighter can efficiently query basic information such as equipment, resources, personnel distribution, rapidly iterate all kinds of operational activity information, timely introduce related information such as operations, training and scientific research, analyze and evaluate the operational situation, and shorten the circulation and update cycle of command information in the OODA environment.
The third is to optimize the action plan. The generative AI system automatically prepares detailed action plans based on specific action tasks, on the basis of the user's input of the basic mission action framework, through learning and training of historical plan and action rules, and in combination with the current force, equipment and battlefield environment, for the reference of commanders.
The development status of generative AI in the US military
In January 2023, the US Defense Information System Agency (DISA) listed generative AI technology in the "Technology Watch List". In August of the same year, the US Department of Defense set up a "Lima" generative AI working group, led by the Chief Digital and Artificial Intelligence Office of the Department of Defense, to promote, evaluate and supervise the implementation of the generative AI plan, collect the use cases of generative AI, and tap its potential in improving intelligence analysis, combat planning and other tasks, At the same time, we will promote cooperation among the Ministry of Defense, the intelligence community and other government agencies, and strengthen resource sharing. The working group proposed a plan to jointly develop a "multimodal" generative AI model, and issued a provisional guide to generative AI.
At present, the US military is carrying out the test work of the generative AI system. The field of command and control has the potential application of generative AI in the military field. The generative AI model helps to synthesize and summarize a lot of information, answer extensive and specific questions, write a preliminary combat plan, and put forward suggestions. These can help improve the cognitive ability of military users, enable them to make more intelligent and faster decisions, and draw up unpredictable action plans of the enemy. The U.S. Army and Air Force have used the generative artificial intelligence system in the exercise to significantly reduce the commander's decision-making time, which plays a better role in assisting the commander in solving the key information needs in the decision-making process. It is known that dozens of companies, including Scale AI, Palantir and Anduril Industries, are developing AI based decision-making platforms for the Pentagon. Microsoft has also announced that users of Azure government cloud computer services can access OpenAI's AI models.
1. Development of the Army In May 2023, the U.S. Army chose to use the large language model Donovan system of Scale AI for the encryption network of the 18th Airborne Army for decision-making. Donovan system is the company's federal AI platform, which can help warfighters, analysts and decision makers accelerate the understanding, planning and action speed of battlefield situation. Donovan system uses the reinforcement learning algorithm based on human feedback to continuously fine tune the system model, so as to constantly adapt to new tasks when the task objectives change.
One of the key challenges faced by military planners is that there is a large amount of information that can support decision-making, but the format of these information is not uniform. With the help of Donovan system, decision makers can make better and faster informed decisions in the rapidly changing battlefield environment without increasing the number of staff officers. The Donovan system has acquired more than 100000 pages of actual combat data, such as mission commands, situation reports, intelligence reports, etc., to help staff officers without programming or training experience understand and manage the growing data within a few minutes. The Donovan system also supports joint global command and control, which can accelerate the planning speed and ensure the coordination between the Department of Defense's trans theater combat commands. For example, the air mission command cycle currently requires three days, and the Donovan platform can shorten the planning cycle to several hours.
In the "Red Dragon" exercise held by the US military, the 18th Airborne Army used the Donovan system in the following battlefield scenarios:
(1) Real time operations of friendly and enemy forces can be understood in real time by receiving real-time data, including commands, situation reports and intelligence reports. Donovan system can help commanders better play the role of staff officers, more easily identify new solutions, and assess rapidly changing situations.
(2) By receiving battlefield reports and force observation information, the background understanding of intelligence fusion can be realized, thus reducing the cycle of operational planning.
(3) By receiving open source data about potential conflict areas to support crisis planning, provide situational insight about the battlefield, support battlefield intelligence preparation, and provide decision-making advantages for commanders.
2. Development of the Navy In September 2023, the U.S. Department of the Navy issued a Guide to the Use of Generative Artificial Intelligence and Large Language Models to guide the development and application of such technologies within the Navy.
At present, the U.S. Navy has positioned generative AI as a virtual assistant to reduce the work pressure of personnel and improve efficiency. In August 2023, the Navy cooperated with General Dynamics Information Technology Company (GDIT) to launch the artificial intelligence assistant Amelia, which aims to improve the response efficiency of the Navy system service desk and liberate human operators for more complex and important matters. Amelia belongs to the Navy Enterprise Service Desk (NESD) project. It adopts the same generative AI technology as ChatGPT, integrates more than 90 naval network services into a single platform, and uses government authorized data resources to provide one-stop consulting services for more than 1 million naval, marine and civilian personnel, According to Navy officials, Amelia can respond to thousands of service requests within 45 seconds.
At present, Amelia mainly uses the databases of non military information systems such as human resources, training and education. However, the Navy and GDIT are working hard to expand Amelia's training and database scale. It is expected that in the future, the system can have higher confidentiality authority, handle more complex command requirements, and become the right assistant of the sergeant.
3. Development of the Air Force The U.S. Air Force is relatively cautious about the application prospect of generative AI in the military field. Frank Kendall, the Secretary of the U.S. Air Force, publicly stated at the event held by the U.S. Security Center that although generative AI has made significant breakthroughs in some fields, he believes that due to reliability and other reasons, The content generated by generative AI may contain false information, so there is still a long way to go in terms of functions and application scenarios. At present, the application prospect in the military field is relatively limited. To this end, he has instructed the Scientific Advisory Committee to carry out relevant investigations and assessments to find out how to reasonably introduce generative AI as soon as possible.
Even so, the Air Force has carried out a series of tests for generative AI. In July 2023, the US Air Force first tested the use of big language models to execute military tasks in the 6th Global Information Advantage Exercise (GIDE), and tested five big model systems, hoping to use the data generated by artificial intelligence systems to assist decision-making, obtain target information and support fire strikes. In the test, the US Air Force inquired about information from a US military unit by telephone. The AI tools used during the test only took 10 minutes to complete the query, which may take hours or even days in the traditional way. During the test, the US Air Force provided secret level operational information to the model, and the Ministry of Defense did not disclose the language model model being tested, but Scale AI said that its Donovan system was one of the platforms being tested.
In July of the same year, the US Air Force Rapid Combat Service Support Office also introduced a big language model intelligent application called "combat readiness" to improve combat service support capability. "Combat readiness" can integrate multiple program processing results, including forms, texts, scripts, codes, sensor data, log files and other multi-source information, and realize knowledge embedding through in-depth learning model. Through this application program, the US Air Force can conduct human-computer dialogue, data retrieval, sensitive data access and other operations, thereby optimizing the air force fleet combat service support process, significantly reducing aircraft downtime and maintenance time, and providing support for improving the completeness and attendance of combat aircraft.
4. Development of the Marine Corps In April 2023, the U.S. Scale AI Company and professors from the U.S. Marine Corps University jointly developed the Hermes big language model to test the model's ability in campaign level combat planning. Students use the "Hermes Military Planning Large Language Model" to understand all aspects of the opponent's national strategy, so as to shape the battle strategy. The large language model helps military planners observe the battlefield in multiple dimensions.
Since the design team loaded the opponent's theory and doctrine data into the corpus, students could ask questions such as "What is a joint blockade?" and "How does Country X use diesel submarines?". Hermes' big language model is excellent at helping students answer questions related to doctrine, which help to formulate the enemy's action plan. The student team also uses this model to understand the economic links between regions, the political timetable of specific countries (such as elections), and the major infrastructure investment of specific countries (such as China's "Belt and Road" initiative). Through analysis at different levels, student groups can become familiar with various links in the combat environment to better develop the concept of adversarial warfare. In addition to the factual problems, the students also used the "Hermes Military Planning Large Language Model" in the confrontation environment to generate imaginary situations about time advantages and location advantages. By analyzing these imaginary situations, the student groups improved their action plans.
With the help of large language models, student groups can better understand the relationship between the operational environment, time, space and combat forces by asking and obtaining answers to questions. This experiment shows that it is necessary to start integrating large language models into the military planning process.
Existing problems
At present, the military application of generative AI technology is still in the exploration and initial stage, mainly due to the following three risks.
1. Machine hallucination problem The Navy Department's Guide to the Use of Generative Artificial Intelligence and Large Language Models points out that such AI tools currently lack sufficient reliability, because they will have machine hallucination problems, that is, they provide users with feedback content that appears to be real but is actually completely wrong and fictitious. In addition, cognitive biases and machine hallucinations generated by the model, and if the data for training artificial intelligence models are artificially affected, the training results will be biased, which is also the problem that the US military is currently working to solve when establishing the "Lima" working group.
Margie Palmieri, chief digital and AI officer of the US Department of Defense, also said publicly that the evaluation criteria of generative AI are still unclear, and the application effect of generative AI technology can only be judged by using cases, and generative AI technology has the disadvantage of providing wrong information, This is a major defect for the battlefield application of the US military, so it is necessary to test different models through military exercises, and constantly adjust and modify the system on this basis.
Maynard Holliday, deputy chief technology officer in charge of key technologies of the US Department of Defense, said that the Department of Defense would not deploy commercial off the shelf technologies such as generative artificial intelligence systems in the current instance. In the future, it might cooperate with the industry and academia to tailor corresponding technologies to meet military needs. DoD also seeks multimodal generative AI algorithms that integrate language, vision, and signal information to respond to the requirements for joint operations in the Joint Global Command and Control (JADC2) concept.
2. Insufficient training data is the underlying driving force of generative AI applications. Generative AI mainly relies on a large amount of data to train a huge neural network converter to generate content interacting with human beings. Only through sufficient data "feeding" and relying on massive data "training" can various functions of AI be realized. Taking ChatGPT as an example, from the release of version 1.0 in 2018 to version 4.0 at the end of 2022, the number of model parameters of ChatGPT has increased from 100 billion to 100 billion, and the number of training data has increased from GB to 100 TB.
However, there are information barriers and data extraction problems in the current military field. The "chimney" phenomenon of various information systems in the army is still quite prominent. Most of the basic data, such as scattered "depressions", are scattered in various information platforms of the army. It is difficult for the generative AI system to call "big data". In most cases, it depends on manual collection and summary level by level, which has a large workload, slow data update, and low accuracy. At the same time, ChatGPT and other major generative AI products originate from American technology companies, whose training texts are mainly in English, and lack of high-quality military data in other languages as learning materials, resulting in the lack of credibility of the output provided by such products when it comes to international or foreign issues.
3. Disclosure risk In terms of data privacy security, the application of generative artificial intelligence needs to input a large amount of sensitive information and data for continuous training and learning, so there is a risk of disclosure. During the storage and flow of data between the government, military departments and third-party institutions, network attacks, equipment destruction, improper personnel operations, etc. are likely to cause data leakage, which will bring serious security risks to the military, and objectively put forward higher requirements for data security.
In terms of military personnel privacy and security, the big model can analyze and parse personal data and photos to obtain a large amount of sensitive information, including personal identity, location and movement trajectory. This information can be used to track, track and monitor military personnel, resulting in privacy violations and personal security threats. In addition, advanced generative AI technology can deceive and extract sensitive data of senior officers and senior officials by generating real dummies and false information, which will also increase military security risks.
In view of such data security risks, the US Space Force has suspended the use of web-based AI tools such as ChatGPT on government computers by its staff members since September 29, 2023, unless officially approved by the Chief Technology and Innovation Officer of the Space Force.
Viewpoints
Generative artificial intelligence has shown great potential in the field of national defense. Although the technology is still not perfect, the U.S. military has begun to actively explore and promote the application of generative AI in information control, military strategy, intelligence analysis, unmanned technology and other aspects to improve military capabilities and mission efficiency. In the long run, the United States hopes that AI can enhance the capabilities of military planning, sensor analysis and fire decision-making. Therefore, the dialogue with the big language model is only the first step of their broader AI goals in the next decade. In the face of the impact and impact of the generative AI system on changing the world and changing the future, we must stand on the height of winning the future war, recognize its significance of the times, and promote and guide the development of battle command.
