Strategy in the Artificial Age: Observations From Teaching an AI to Write a U.S. National Security Strategy

210404-N-YA628-2309

Words matter to members of the U.S. defense establishment, especially if those words are found in official documents like the National Security Strategy or the National Defense Strategy. The carefully chosen words of these strategies are analyzed and debated and critiqued and praised as they are dissected in the public forum. The next National Security Strategy is central to the national strategy debate today, as the Biden-Harris administration develops its eagerly awaited addition to that pantheon of strategic crown jewels. Typical analyses of prior U.S. national security strategies focus on the implications of their content, provide sage advice for drafters of future versions, and assess the changes that each president makes to leave his or her mark on U.S. national security. Analysts are gifted at identifying the departures from precedent that appear in each new U.S. National Security Strategy.

This article attempts to do the opposite. Instead of asking what has changed across each iteration of this document, this article seeks to understand what has remained the same. To this end, I used machine learning to create the “bernardcodie” program. Bernardcodie is an artificial recurrent neural network trained on the entirety of the U.S. National Security Strategy corpus. Leveraging the natural ability of deep learning networks for pattern identification, this article complements human insight with artificial analysis. So, what patterns are revealed when we teach an artificial neural network the canon of U.S. national security strategies and ask it to write its own strategy? First, the neural network is a veritable thesaurus for language about change, consistently describing a strategic environment of constant flux and struggle. The strategy is extremely abstract, relying on broad language that speaks to many audiences and plays myriad roles. The algorithm then reveals a perpetual search for the next new concept, approach, or reform that will finally enable effective whole-of-government coordination. Lastly, the artificial strategy and complementary text analysis reveal a dependence on the language of virtue, values, and social closeness, quantifying the importance of American identity in strategy.

 

 

Taken together, these themes describe rhetorical patterns that defined prior strategies and will likely shape the next strategy. If the patterns of the past can be identified, perhaps we can gain insight into how those patterns will influence future strategies. With the recent publication of the Interim National Security Strategic Guidance, an opportunity arose to immediately test the bernardcodie language model against the strategic rhetoric of the Biden-Harris administration. As analysts scrutinize the language of the interim guidance and make their recommendations for the full strategy, recognizing the consistent abstract, transformative, aspirational language across past strategies should temper expectations for strategies to come.

Creating a Strategy Deep Writer 

Bernardcodie is an artificial recurrent neural network, or what many would call an AI. Artificial neural networks and machine learning have already been applied to geospatial intelligence analysis, drone swarms, cyber security, cryptography, and numerous other defense applications. Machine learning programs are early in their development and are not advanced enough for many defense applications. For example, modern AIs are highly dependent on rules and struggle to handle unexpected situations for which they have not been trained. An AI can win Jeopardy or convincingly mimic Shakespeare, but a hypothetical self-driving tank may be stopped dead in its tracks by unclear lane markings or poor weather. Despite current limitations, AI and machine learning are widely expected to transform future national security. In this family of national security-focused AI applications, bernardcodie is a piece of predictive text technology that has been trained to reproduce linguistic patterns and trends in U.S. national strategy.

The recurrent neural network used in this article was created through a machine learning technique called “deep learning.” To train a deep learning algorithm, the creator begins by showing the computer a training set of example text. For bernardcodie, this set included all 17 U.S. national security strategies published since the document’s creation was mandated in the Goldwater-Nichols Department of Defense Reorganization Act of 1986. By analyzing these examples, the neural network builds a web of interconnected “neurons” that generate what is essentially an exceptionally complicated probabilistic flow chart that models the examples in the training set. The more frequently a term or phrase appears in the training set, the heavier the weight assigned to that particular word choice or turn of phrase. This neuron analogy comes with a caution: Do not be deceived by the “intelligence” segment of the term “AI.” The largest publicly available neural networks of this type have billions of these interconnected “neurons” approximately the same number of neurons as are found in the brain of a giraffe. Bernardcodie has a smaller training set which results in a far less complex language model with far fewer neurons. In comparison to the giraffe-like intelligence of larger models, bernardcodie is an erudite fruit fly.

When provided with a starting word or phrase, the neural network traces through this flow chart predicting what words will come next given the examples on which it was trained. This process is deep writing. The bernardcodie network has what is called “long short-term memory,” which gives the neural network the ability to hold information from prior predictions in its “memory” as it deep writes. This is how the network is able to write a coherent paragraph on any topic without losing the logic of the argument. The output of the bernardcodie predictive text program is an entire artificial U.S. national security strategy, produced in February 2021, that is modeled on the structure and framework of prior strategies. Due to limitations of space, only relevant predictive text and revealed patterns — not the entirety of the artificial strategy — are reproduced here. All predictions and quoted text in this article, unless cited as otherwise, are unedited output from the bernardcodie recurrent neural network.

As it is the product of an exercise in prediction, the bernardcodie network is named for Bernard Brodie, the researcher whose work shaped defense theory in the nuclear age and who is credited with predicting the doctrine of massive retaliation. I make no claims to Brodie’s level of insight. Instead, this article draws its guidance from Brodie’s discussion of the limitations of prediction using scientific methods in Strategy in the Missile Age. This article takes his wisdom to heart.

Everything Has Changed 

According to text analysis, the world according to the U.S. National Security Strategy becomes exponentially more dangerous and more lethal, requiring greater innovation and faster decision-making every few years. Even Brodie warns of the “utterly unprecedented rate of change” that “has moved much too fast to be fully comprehended even by the most agile and fully-informed minds among us.” Bernardcodie and his namesake are of one mind on this point:

War is new and global. The nature of war is becoming increasingly complex, more lethal, with each conflict challenging our skill levels and our ability to manage the demands of the future. In the future, we must prepare for a host of threats to U.S. national security that do not fall within the traditional categories.

In every iteration of the artificial strategy, the neural network deep writes that the strategic environment will have changed “significantly,” “markedly,” or “dramatically.” Likewise, prior national security strategies have had high levels of language indicating change. Even the Biden-Harris interim guidance depicts the changing world of 2021 as an “inflection point” characterized by “accelerating global challenges.” The time period from 1987 to today spans critical geopolitical shifts, but the large amount of transformational language in U.S. national security strategies would suggest that change is occurring more rapidly than ever realized.

In addressing this dangerous fluctuating world, the National Security Strategy plays a number of important roles in shaping the foreign policy direction of the U.S. government. As mandated by the Goldwater-Nichols Act, the document is a “report” that must contain a strategy. In its role as the “strategy of strategies,” according to the Joint Chiefs of Staff, the document provides important direction for the National Defense Strategy and National Military Strategy. At the same time, the National Security Strategy is not a strategy. It is a vision document and a megaphone for the current president. As a policy document, the National Security Strategy communicates responsiveness to domestic constituents, signals the priorities of the United States to the international community, and guides and justifies resource allocations. Not only does the U.S. National Security Strategy play all of these roles, but it must remain relevant for years and do so in an unclassified manner.

How is it possible for a document to play all of these roles effectively, given different administrations, a rapidly changing environment, and various historical contexts? It must be precise enough to guide the application of national power but abstract enough to apply to a dynamic and unpredictable world — a perpetual balancing act between applicability and utility. Nebulous language in the Interim National Security Strategic Guidance is already proving to be excellent fodder for discussion. However, this is not a discussion point unique to the Biden-Harris administration. In prior efforts to achieve this balancing act, some strategies have become too abstract and have been accused of vacuousness. Text analysis supports these accusations to a degree. National security strategies are full of abstract language. Criticisms of vague language are fair critiques, but perhaps an equally important question is this: If the National Security Strategy were not abstract, could it effectively play the myriad roles that are asked of it?

Innovation, Modernization, and Revitalization 

Bernardcodie’s outputs call for new approaches and ideas and this finding is amplified when other text analytic techniques, including topic modeling, are applied. One of the most important groupings of words found in prior national security strategies is characterized by the term “transformation.” The bernardcodie program reveals the following:

The future of warfare depends on the ability of armed forces to rapidly identify and disrupt sophisticated threats to our nation’s security. Deterring and defeating them demands more than better systems or weapons. It requires new mindsets and new thinking. Its success is directly dependent on the ability of the U.S. Armed Forces to develop, integrate and deliver new concepts and ideas for engagement, networking, command and control, and integration of new intelligence and analysis, as well as new types of forces and capabilities across domains.

The Biden-Harris administration cautions against pining for times past, saying that instead, “we have to chart a new course.” This innovation can be realized through new ideas and concepts. Bernardcodie suggests that this transformation should occur in the form of government reform, particularly within the Department of Defense:

If we are unable to compete successfully in the new global environment, our adversaries will not hesitate to threaten us and our allies and partners. The Department of Defense must modernize its overall force structure to achieve a globally competitive, American military that is equally tailored to meet the requirements of today’s fight, today’s threat and tomorrow’s war.

Evolution and growth are central to a healthy organization, and the Department of Defense deserves commendation for continued focus on modernization and innovation even if there have been challenges in actualizing those reforms. The 2021 Interim National Security Strategic Guidance includes these same calls for transformation, reform, and modernization, stating that the administration will “reform and rethink our agencies, departments, interagency processes, and White House organization” to reflect a perceived degradation of boundaries between foreign and domestic policy.

In bernardcodie’s artificial strategy, the need for reform is tied to a characterization of a United States falling behind its adversaries. The program writes that “as we look toward the future, we see an opportunity to realize the great promise of America’s comeback through new and enhanced tools, new structures, and new levels of cooperation.” Even the Biden-Harris administration’s tagline “build back better” is in the same rhetorical vein as bernardcodie’s phrase “America’s comeback.” As institutions heed this call for revitalization and innovation, have historical lessons been lost in this perpetual hunt for the next new idea, process, or reorganization? What does this pattern say about the success of past efforts to restructure and reform?

Values Are Everything

Despite arguments that suggest that purported American values may not be as “normal” to U.S. foreign policy as observers may think, the language of values certainly is. In contrast to analysis that argues the U.S. National Security Strategy privileges military strength and the use of force, the bernardcodie program’s artificial analysis identifies a powerful pattern of values-based rhetoric:

Promoting democratic values and practices in every continent is an integral part of U.S. foreign policy. The United States can, and should, play a key role in fostering and sustaining our nation’s stability by working to strengthen alliances, common values and the rule of law.

Linguistic analysis does not find that the language of military power is as central to these documents as human analysis of recent strategies might suggest. In a quantitative analysis of the language of prior national security strategies, the most significant finding is a disproportionately high amount of abstract, value-based language such as democracy, human rights, fairness, etc. Furthermore, U.S. national security strategies contain a great deal of language referring to a common identity, by using “we” and “us,” and explicitly naming virtuous ideals and values and connecting them to the American way of life. Bernardcodie writes:

The United States of America is the foremost guarantor of peace, freedom, human rights, and dignity. We are the leader of the free world in mobilizing for a universal peace effort, using diplomatic, economic, military, cultural, and moral tools.

Given the divisiveness of the current political climate, it is even more likely that the next National Security Strategy will call on an aspirational American identity. This should not be taken as new. The National Security Strategy has always presented an idealized version of the United States and its role in the world. The Biden-Harris administration’s Interim National Security Strategic Guidance does just this, stating that an “undaunted” America will lead through the power of its example and its “fundamental advantage: our democracy.” Bernardcodie sees U.S. national security strategies as defining an identity, describing a country guided by values, rule of law, integrity, democracy, and morality, and featuring a call to remember those values. The next National Security Strategy will likely not only provide a vision for the national defense, but also provide a vision for the example that the United States aspires to be — just as prior strategies have done before.

Limitations of an Artificial Approach

In Strategy in the Missile Age, Bernard Brodie warns those of us who would attempt to apply novel scientific methods to military decision-making to “appreciate how imperfect is even the best we can do.” Bernarcodie is imperfect, as all artificial neural networks are, but there are a few specific limitations that qualify this analysis.

First, human language is complex and AI has no common sense. AIs excel at narrowly defined tasks but have no transferrable general intelligence, a challenge known as Moravec’s Paradox. Where human strategists can draw upon their prior experience and intuition to easily identify subtext and subtle cues, the neural network merely sees strings of letters and punctuation.

Second, deep learning networks are only as good as the source material on which they are trained. Two important features of the training set used to create bernardcodie warrant emphasis. First, the Clinton administration produced seven of the 17 published national security strategies. As a result, bernardcodie has learned more about President Clinton’s approach than that of any other president. Second, the most recent training material that the bernardcodie network has analyzed is from 2017 — the date of the last U.S. National Security Strategy. Much has happened since 2017, but the bernardcodie program knows nothing about it.

Third, there is always a danger of “overfitting,” or the network removing variation in the data to fit more abstract patterns in the training set. The manner in which neural networks learn means that the more frequently a pattern appears, the heavier the assigned weight and the more likely the deep writer will produce that pattern. Important variance may be discarded because the network identifies and reproduces patterns rather than deviations. Bernardcodie is not designed to recognize or appreciate deviations.

Lastly, neural networks make mistakes. At one point, bernardcodie was convinced that the United Nations had laid the international date line at the 38th parallel. In another strange comment, bernardcodie suggested that the war in Afghanistan was “more than just a war of force; it is, in fact, a journey of self-discovery and self-reflection.”

Conclusion

Critics of national security strategy analysis might question whether these documents warrant such analysis. If the assumption is that “words matter,” then it would benefit everyone to understand not only the substantive shifts of the past decade, but also the linguistic consistencies. AI and natural language processing make word choice central to analysis, offering a unique opportunity to assess the national security strategy genre in its entirety and identify patterns and consistencies in strategic language. The artificial approach is imperfect and cannot replace the intuition of the human mind. However, when considered in concert with human analysis, AI provides a novel perspective and expands our understanding of American strategy.

If the signals of each subsequent president represent surface waves, then decades of rhetorical patterns are the deep currents of U.S. national defense strategy. As readers await the publication of the next full strategy, perhaps it would be beneficial to consider whether a document that has always been nebulous, forward-thinking, and aspirational can truly be everything to everyone, as it is currently expected to be, or whether readers will always be left wanting.

As the debates are already ongoing as to what should be included or changed in the next National Security Strategy, I offer one last rhetorical pattern. Ironically, there is a clear historical pattern of claiming that the security environment and its challenges and opportunities are unprecedented. As Biden himself notes in the preface to the interim guidance, “under the Biden-Harris Administration, America is back. Diplomacy is back. … But we are not looking back.” History holds a great number of valuable lessons, but the currents of strategic language appear to be forever looking to the future. As the debates continue about today’s rapidly fluctuating environment, the necessity of innovation and reform, and the aspirational vision of America as it aspires to be, it is worthwhile to remember that there are decades’ worth of prior attempts to meet these particular challenges from which strategists can draw insight.

 

 

Elena Wicker is a Ph.D. candidate in international relations at Georgetown University. She researches innovation; military jargon; and the past, present, and future of strategy and the domains of war.

Image: U.S. Navy (Photo by Mass Communication Spc. 3rd Class Arthur Rosen)