“There is a pre-ChatGPT and a post-ChatGPT world,” said the CEO of a leading AI company for national security to your author during a conversation in Palo Alto.
Some may at first look think this dichotomy in the commercial world was a result of what ChatGPT does — it is good, but large language models (LLMs) have been around for the last decade, and, well, Google’s GEMINI is slightly better or equally worse depending on who you ask. Rather, what was innovative about ChatGPT was that it exposed “AI/ML” to the masses. For the first time, consumers experience for themselves what AI could do for them. With one application, what had been hidden, poorly misunderstood, and known by only a few went viral. With one product, people and companies globally experienced intrigue, surprise, worry and saw opportunity, all at the same time.
However, the operationalization of AI/ML across several industries began decades ago. In the defense and national security sector, AI/ML models have traditionally been used and thought about as “process enhancers.” As far back as the 1980’s the US Intelligence Community was “narrowly interested in the data mining and data processing advantages of AI, namely the ability of computers to search through vast troves of raw data, from multiple sources, and transform it into usable information for human analysts.” As the same authors put it, “AI has long been held as a panacea to the problem of data smog.” [1] That is, models were designed to aid in the cleaning, processing, and filtering of massive amounts of data to save time for traditional all-source analysts. For a relatively uncharted technology, this has been the safe application — the sustainable innovation. And big defense contract companies, like Palantir, saw the opportunity. Before the government was drying the ink, big companies were being awarded contracts to optimize collection and processing of data.
Beyond Process Enhancers
But sustainable innovation is, well, sustainable and quickly becomes the status quo.
A couple years ago, this began to change. Companies like Rhombus, an AI for national security startup in Palo Alto, and others went beyond process enhancing and data collection. These companies began to develop and deploy AI/ML models that answered the toughest geopolitical, and national security questions. At first, they were experiments, but they soon began producing powerful insights. Now, US and allied intelligence agencies have begun to operationalize this flavor of AI/ML. In June, Foreign Policy described how Rhombus AI models predicted Russia’s invasion of Ukraine months before the IC predicted it and how the same model has predicted Russia’s strategic moves during the war.[2]
This early adaptation of AI/ML models that are not just enhancing a process or processing large amounts of data raise many strategic and tactical questions. For example, where does this flavor of AI/ML belong? How can intelligence analysts use it? How can commanders or decision makers use it? What does it mean for an intelligence agency to cite an AI/ML insight? What is the process for creating an AI/ML model for intelligence analysts?
These are not simple thought experiments. Information superiority will be won by intelligence agencies that understand and adapt AI/ML — not exclusively to support or process data —but also to augment their intelligence analysis to produce better, more survivable intelligence. In 2017, Harvard researchers anticipated that “AI/ML had the potential to be a transformative national security technology, on a par with nuclear weapons, aircraft, computers, and biotech.”[3] Other authors have posited that AI/ML’s strategic advantage will be as or more powerful than the Enigma Machine was in the Second World War. [4]
The Advent of AI-INT
Whether that materializes or not depends solely on how fast the intelligence agencies of the world adapt to this new form of AI/ML.
And what is the first step to operationalizing this flavor of AI/ML: to create a common definition and nomenclature framework. AI/ML models that are designed to answer geopolitical questions and their insights are categorically and functionally different from AI/ML models that augment collection or data processes. They also produce and output “new” information (raw intelligence). This raw intelligence was informed by the data that fed and trained the model, but the outputs are unique, stand on their own, and require evaluation just like any other traditional source of intelligence (human, signals, geospatial, or open-source intelligence).
Your author coined these models, their outputs, and ensuing analytical tradecraft: AI-INT (pronounced with a pause AI — INT). AI-INT is defined as a raw analytical output derived from a specifically created AI/ML learning model that answers one key intelligence question, or a specific national security question of interest. A raw analytical output is the AI-derived insight still untouched and unevaluated by human analysts.
AI-INT is not equivalent to the many AI/ML models already in use to aid in the processing of intelligence data. AI-INT is also not a finished intelligence product. It is a raw source of intelligence that human intelligence analysts must evaluate, synthesize, and fuse with other sources of intelligence.
To qualify as AI-INT, the raw analytical output (the raw insight) derived from an AI model must answer yes to all the following questions:
· Does the raw analytical output help answer a specific analytical key intelligence question? For example, is it a model that predicts global instability versus a model that filters and summarizes reports about global instability?
· Did it use an AI/ML model to produce that raw analytical output? In other words, is the model engineered and configured using recognized AI/ML architectures, such as long short-term memory, deep neural networks, convolutional neural networks, and other recognized model architectures? Or is it just big data processing and visualization?
· Can an intelligence analyst evaluate the raw analytical output as a new single and raw source of intelligence? In other words, does the raw output require analysts to interpret, weigh against other sources of information, and put into subject matter context? For example, imagine that an AI-INT model is predicting in August an 85 percent probability that Brazil will experience political instability in January 2024. This output is not sufficient on its own and needs an analyst to interpret. What type of political instability and why?
The Sixth Intelligence Discipline
Does the AI-INT definition and criteria suggest AI-INT is or will become the sixth intelligence discipline? It is too early to tell. But there is a persuasive argument for the Office of the Director of National Intelligence (ODNI) to consider. If there is a criterion to nominate INTs, they tend to include three things.
· First, is it a new and distinct collection source that by extension produces a new and unique set of information? Yes, AI-INT does. It is not derivative intelligence because once a data input is processed through an AI/ML model’s architecture that reasons and generates thousands of causal and inferred correlations internally that source of information has been processed (collected) and is by all measurements new and different than its input when it is outputted.
· Second, does it have specific sensors or collection mechanisms that require specialized and segmented understanding? Yes, AI-INT does. Just think of the engineers, subject matter experts, intelligence analysts, and data scientists required to create, upkeep, and improve specifically designed models for specific intelligence questions.
And third, does it require the development of a specialized analytical tradecraft? Yes, AI-INT does. This is the most complex and untapped of the three criteria. AI-INT analytical tradecraft refers to the intersection of AI-INT with the analysis part of the intelligence cycle. This is by far the most essential element to operationalize AI-INT and is the subject of an upcoming book by your author. Important, however, is that for AI-INT to deliver on its strategic advantage, it must be trusted and used by human analysts. To do that, the techniques, tactics, and procedures need to be created, taught, and enhanced so AI-INT can be fused and incorporated into finished intelligence products.
AI-INT and Information Superiority
AI-INT’s power stems not only from its ability to provide predictive intelligence but survivable intelligence. When war comes, our traditional sources of intelligence will be affected. Adversaries know what we look at and what we listen to. But it is “hard to kill” an AI/ML that predicts Russia’s next military movement. Or a model that predicts China’s next harassment activity against the Philippines in the South China Sea. The adversary can hide a tank but it can not hide or influence the plethora of unconnected and seemingly unimportant signals– food orders, fuel storage, mineral extractions etc–that AI-INT model processes and reason with to produce predictions about that country’s next tank movement. They can not limit the possibilities of that new source of information being fused with traditional sources of intelligence or informing as a sole source to an analyst’s assessment to a commander or policy maker.
Coining AI-INT is only the first step in a long road. The intelligence agencies will only materialize AI-INT’s strategic advantage when they achieve knowledge of how to understand AI-INT models, how to create them, and how to use their outputs into traditional intelligence analysis and operations (analytical tradecraft). It also equally requires policymakers to understand how AI-INT is informing or feeding the intelligence assessment they read and use to make strategic and operational decisions. But that is a separate conversation.
We may not know it yet, but we are in the first year of the AI-INT world. The question is, are we ready for it?
Learn more about the intersection of AI and national security. Reserve a free copy of your author’s upcoming book: AI as the Sixth Intelligence Discipline: A Tradecraft Guide for Analysts
Kendrick R. is the founder of MAIK and author of AI as The Sixth Intelligence Discipline: A Tradecraft Guide for Analysts
[1] https://doi.org/10.1093/jogss/ogad005
[3]https://www.belfercenter.org/sites/default/files/files/publication/AI%20NatSec%20-%20final.pdf