The "ChatGPT Moment" and National Security
How Will Large Language Models Impact U.S. National Security?
Hello, I’m Ylli Bajraktari, CEO of the Special Competitive Studies Project. In this edition of 2-2-2, Abigail Kukura and Asher Ellis of SCSP’s Platforms Panel discuss what is at stake for national security with the proliferation of Large Language Models (LLMs) like ChatGPT.
ChatGPT has taken the world by storm. Since its release in November 2022, OpenAI’s chatbot has scored a 1020 on the SATs, passed a medical licensing exam, successfully negotiated down a user’s cable bill, and even drafted legislation for how to regulate AI. It has been the topic du jour from classrooms to podcasts to dinner tables, and even celebrity-studded television commercials.
There has been a lot of commentary on the impact of ChatGPT on areas like education and the workforce. But what does this “ChatGPT moment” — and LLMs more broadly — mean for U.S. national security?
At a strategic level, nations will seek to harness AI models for economic, military, and national security advantages. As OpenAI CEO Sam Altman told Eric Schmidt during the National Security Commission on Artificial Intelligence’s (NSCAI) 2021 Global Emerging Technology Summit, “there will be multiple global efforts to build these powerful AI systems, and the best way to get the outcomes that we want in the world, AI that aligns with liberal democratic systems, would be for us and our allies to build them first.”
The rest of this newsletter will outline some strategic considerations for LLMs, first taking stock of where we are today in a crescendo of AI progress, then turning to specific implications of LLMs for the Defense Department, Intelligence Community, and the wider U.S. national security apparatus. We will close with key considerations for policymakers.
The “ChatGPT Moment” Amidst a Crescendo of AI Progress
ChatGPT did not come out of nowhere. Rather, it is the latest data point in a broader trendline of rapid AI progress. There has been a rapid acceleration in recent years in Natural Language Processing (NLP), the field of AI in which computers analyze and generate human language, as well as in other Machine Learning (ML) capabilities that have improved task performance by machines. Between the two bookends of OpenAI releasing GPT-3 in 2020 and ChatGPT in late 2022, the LLM ecosystem alone has grown to include more than a dozen entities from around the world that have released models and/or built fit-for-purpose tools on top of open-source LLMs.
Actors in the LLM Ecosystem:
Many of the key actors driving LLMs forward today are U.S. companies. However, Chinese entities and companies, as well as organizations in the UK, Israel, South Korea, and other countries, have become players on the board as well. While the U.S. private sector has a competitive advantage, the PRC is determined to surpass the United States in AI, where leadership across the stack — computing power, algorithms, data, applications, integration, and talent — is up for grabs. China is also a fast follower in LLMs, and rumors have begun to swirl that Chinese search engine company Baidu will release its own ChatGPT-like chatbot in March.
What’s more, we can expect the AI ecosystem to expand further as the barriers to training AI models, like cost, decrease with new techniques like “no-code AI.” The United States must prepare for more and more actors across the world – from states to companies to individuals – to be able to create and leverage customized algorithms for myriad purposes.
Indeed, tools like LLMs are becoming increasingly practical and accessible. In the two months it has been online, ChatGPT has reportedly reached a record-breaking 100 million monthly active users. It took TikTok nine months to hit that milestone. Google has announced it will soon release its own chat model, Bard, built on top of its in-house LaMDA model. Like many other organizations, SCSP has used LLMs to help generate content in recent months. In our first report, Mid-Decade Challenges to National Competitiveness, we asked three LLMs — OpenAI’s GPT-3, Hugging Face’s BLOOM, and a model developed by Anthropic — to weigh in on key questions the report aimed to address (the LLMs’ answers were reasonable, if generic). We also used an LLM to help write a previous edition of this newsletter, and our 2022 Global Emerging Technology Summit featured an Anthropic LLM (alongside Anthropic co-founder Jack Clark) answering live questions from the audience.
Yet LLMs can do more than simply talk about the national security landscape. They are actually starting to shape it. Below are a selection of potential use cases for LLMs in the U.S. national security enterprise.
Opportunities for the U.S. national security enterprise:
LLMs as the interface for "intelligentized" tools: A ChatGPT-like LLM could be the interface through which humans access a wide range of AI-powered tools, including computer vision and robotics. Human-machine teams (HMTs) have already proven to be more effective than either a human or AI model on its own for tasks such as cancer screening. In a national security context, such intelligentized tools could transform the Indicator & Warning landscape through AI-generated predictive insights (see more on “decision support” below), enable semi-autonomous systems like using LLMs to train and interact with helpful robots, and support tasks ranging from logistics and supply chain management to the vetting of export control applications.
The LLM-Intelligence flywheel: ChatGPT and other emerging technologies have the capacity to enable a stronger and faster Intelligence Community across all intelligence disciplines (INTs) and phases of the intelligence cycle, and for the military across all levels of war. HMT capabilities using ML models have already demonstrated the ability to increase the efficiency of intelligence processing and analysis by filtering through massive quantities of data and flagging important information, and enabling analysts to focus on more in-depth analysis. LLMs can also improve productivity by helping to generate drafts of estimative analysis, accurately translate foreign texts, etc. For open-source intelligence (OSINT) in particular, LLMs with external search capabilities and expert-informed training data will supercharge our collection and analysis capabilities. This underscores the need for an open-source intelligence center where these tools can be trained and developed for intelligence purposes.
Decision support: Decision advantage will accrue to the country(ies) whose leaders are best able to harness AI models as decision support tools. With the right combination of AI, data, modeling and simulation capabilities, and predictive insights, models will be able to game out various policy options and provide that analysis to human decision-makers. Recent achievements like an AI model outperforming humans in the strategy game Diplomacy, which involves interacting and bargaining with opponents, and in war games demonstrate considerable progress along this tech vector, even if more work remains. There are obvious limitations for how much of a role LLMs should play in semi-autonomous systems that make operational decisions. However, these “savants” could be an additional voice in the room, providing a real-time, data-informed perspective on policy or war options.
Other key national security considerations:
The information landscape: ChatGPT heightens the risks described in Chapter 1 of the NSCAI final report of our adversaries turning the dial on AI-enabled info operations before we develop adequate countermeasures. LLMs like ChatGPT could be used to generate unique texts at speed and scale that evade existing filtering systems. Perhaps most importantly, even though these technical capabilities existed pre-ChatGPT, now that the public is paying attention, more people may question whether what they read on social media, etc. was generated by AI. We may indeed be entering a world where most of the content we consume is AI-generated, creating new vectors for state and non-state actors alike to shape the information landscape. On a promising note, the recent proliferation of synthetic content detection capabilities – including a text classifier released by OpenAI and C2PA’s specification standards – raise the prospect that national security professionals and informed citizens alike will have tools at their disposal to help identify synthetic media.
It all comes down to data and training: LLMs have made great strides in recent years, but they are not oracles. They are ultimately only as good as the data upon which they were trained, and can “hallucinate” information that has no factual basis. Their use will always carry associated risks and must be carefully considered by humans. Nonetheless, LLMs will continue to proliferate, as nations, companies, and other actors advance the state of the art and people integrate tools like ChatGPT into their daily lives. Any nation with global technology ambitions would be unwise to turn away as the ecosystem lurches forward.
AI vs. AI: As AI capabilities are integrated into national security apparata, our LLMs will likely be put to the test against those of our adversaries, whether in specific contexts or for overall information advantage. The United States will need to leverage the competitive advantages of our private sector to ensure that our LLMs are the best on earth. Additionally, we will need to ensure that, where appropriate, the U.S. military and intelligence community’s AI tools are interoperable with those of our allies and partners.
The “black box” problem: As we increasingly rely on AI systems, we need a way to explain and understand them. Advances in “attention modeling” and “chain-of-thought reasoning” are promising steps, but we still cannot “query” an AI application to “discuss” its thought process. Breakthroughs in this space would help establish trust in AI systems by enabling human users to understand why an AI system made a certain decision.
Cybersecurity and code: LLMs will affect national security at a fundamental computer science level via coding, with implications for the national security apparatus, businesses, and consumers. Copilot, a version of GPT-3 fine-tuned for coding, has already vastly improved coding accuracy and productivity. Similar techniques are beginning to be used to detect and prevent Zero Day attacks and can generally help to harden code. LLMs will also create new risk vectors for cybersecurity, such as lowering the barrier to entry for malicious cyber actors. Those risks will make it all the more essential to deploy AI tools at scale that can help keep our cyber systems secure.
Moving at the speed of technology: Whether in developing AI regulations or integrating AI capabilities into the workforce, governments will struggle to keep pace with this rapidly developing technology. The European Commission, for instance, is grappling with how the EU’s forthcoming AI Act will handle systems like ChatGPT, highlighting the difficulty of setting in place governance models that account for the latest capabilities. In a similar vein, the rapid churn of tools and techniques could dissuade government departments and agencies from taking the leap to adopt systems that may be antiquated by the time they are in use. The only way forward will be for procurement processes and governance frameworks alike to be made sufficiently flexible that they can evolve at the pace of AI.
At the end of the day, LLMs and AI more broadly will reshape the national security landscape whether our leaders and organizations lean into the “ChatGPT Moment” or not. The United States has key competitive advantages at the cutting edge of LLM R&D today. But long-term advantage will be shaped by the adoption of such tools, to the risk of those who fail to adapt and fall behind.