Analysis | Which AI tools should diplomats use today?
And how practitioners can leverage the technology
This article is the final part of two pieces on the relationship between advances in artificial intelligence and diplomatic practice. Check out part one here.
It is part of ISD’s series, “A better diplomacy,” which highlights innovators and their big ideas for how to make diplomacy more effective, resilient, and adaptive in the twenty-first century.
Hype surrounding OpenAI’s ChatGPT seems to steam ahead unabated (perhaps even prompting Chinese tech firm Baidu to announce its own rival). While chatbots are indeed wondrous in their ability to produce human-sounding language, their actual reliability in business-use cases is still an open question. In the meantime, there are a handful of startups deploying AI in novel ways that could save effort and time across the spectrum of white-collar functions. Here are three compelling use cases that could enhance the business of diplomacy.
Using AI to mine internal knowledge
Large language models (LLMs) are great at sifting through large datasets and providing insights — that’s what Microsoft’s new Bing update does, and Perplexity.ai’s rival as well (though both are prone to hallucinations). Blending an LLM with semantic search would let policy wonks quickly query thousands of documents. Independent think tank fp21 recently came to a similar conclusion and created a prototype tool outlined in their paper, The Bayes Brief. Using these sorts of tools could save countless hours spent on clicking through folders scattered across computer systems.
I asked Dave Shapiro, an AI expert based in the Bay Area, how policy offices could leverage LLMs today.
“Imagine being able to summarize large volumes of text,” he told me. “You could have a bot that, for example, scans the internet for parliamentary bills passed by countries all over the world, that then takes those bills and summarizes them for you.” And that could be just the start. “You could even have two chatbots debate each other and spark ideas.”
Shapiro’s startup focuses on fine-tuning LLMs for specific applications.
Imagine a team charged with creating a strategy to negotiate a recurring UN Security Council Resolution. It could use a fine-tuned LLM to query internal documentation and see what strategies previous teams used. Instead of reading through hundreds of pages, the team could use AI to summarize its findings and reference its source documents. That means the time-consuming work of manually compiling research could be outsourced to a machine, freeing up the team to generate new approaches. Here again, the negotiating team could ask an LLM like GPT3 for ideas to gain consensus on a resolution, blending ideas in novel ways.
Automating routine tasks
As bureaucracies scale, back-office tasks scale with them, sometimes quadratically. Every civil servant knows there is no escape from the paperwork, box-ticking, and lengthy procedures that inevitably mushroom as legislatures plug holes in processes that might be abused by nefarious employees.
That’s why I asked Yan Zhang, the chief operating officer of UK-based startup PolyAI, how the public sector could use cutting-edge tools, like LLMs. “Back-office tasks are perfect for AI automation,” he told me via video call in February. Wherever accuracy is of utmost importance — for example, a quarterly board report — you probably wouldn’t want AI in the driver’s seat, he said. But in routine tasks, AI could save time and resources. When asked how recalcitrant and famously risk-averse bureaucrats could be convinced to deploy machine learning inside offices, Zhang said, “You need to show that the technology makes fewer mistakes than a person, statistically.”
PolyAI’s product is an example of the use case: It mines internal resources and information and delivers them through a voice assistant. Imagine creating a helpdesk voice bot that used a bureaucratic rulebook, like the State Department’s Foreign Affairs Manual, to give users quick answers and guidance to scenarios. You could ask a question in natural language, such as, “Can I contract a local supplier in cryptocurrency?” The voice assistant would mine its internal information for the answer and then direct the user to the right channel. Similarly, consular sections could use the voice assistant to help visa applicants track their cases, even in multiple languages.
Everything from expense reports to navigating complex rulebooks, to generating meeting summaries could be streamlined with machine learning tools. For example, using off-the-shelf tools, an AI assistant can log-in to a zoom meeting, transcribe the call, and create a summary of what was discussed. That would let participants keep track of tasks and free up work hours otherwise spent on a mundane item. Similarly, an in-person meeting could be recorded, fed to OpenAI’s Whisper transcriber — with its unbelievable accuracy — and then passed to GPT3 to create an executive summary. Much of this work could take place on proprietary servers disconnected from the Internet, meaning the information would stay secure.
Improving the mundane
Generative AI lowers the marginal cost of certain applications, transforming the way we approach work that might be mundane, yet could still benefit from enhancements. When it comes to high-end material — whether an academic producing a textbook, a government writing a public-facing policy document, or a TV news studio making a broadcast video — current AI tools make a small difference in workflows and final products. Yet, at the lower end of the spectrum — whether looking at a blog post, an internal memo, or a training video — generative AI could turn something dull and commonplace into a semi-interesting final production.
Synthesia operates under this premise. When the company’s founder and CEO, Victor Riparbelli, saw early examples of so-called “deep fakes,” he and a large team of AI experts polished the technology into a simple, yet compelling tool. Sythesia’s platform lets anyone turn boring text into a less boring video, aiming for internal corporate training and HR applications. It might not be ready for full-broadcast-quality applications just yet. But with more refinement, we could see a real-life version of the eponymous protagonist in the 2002 film Simone grace our screens. When I asked Riparbelli what new applications we could expect out of AI developments, he said, “Look out for the companies that use AI in ways that are not obvious.” That’s where the excitement will be.
Just as text-to-image services are lowering the barrier on custom visuals, new editing software makes higher-quality writing more accessible. Machine-assisted writing tools aren’t new, but their scale and usefulness are. These new tools go beyond autocorrect and can help create better-written material in complex use cases. Width.ai, for example, produced a system that scans contracts and prompts users to improve the quality of clauses. By training the model on examples of well-written and poorly written legal clauses, Width’s AI nudges users when it detects clauses that need improvement. This is no small feat: policymakers understand that much of the public sector’s machinery operates on written material, from internal rulebooks, to policy statements, to public speeches. Improving the clarity and usefulness of written texts would improve the quality of government services.
Similarly, the inexpensive and lightweight tool from Grammarly helps you write in real time. Rather than an improved spellchecker, Grammarly operates more like an editor and copilot. The company’s AI lets enterprise users create a unified style guide and tone, even incorporating acronyms and jargon. And using Grammarly on a personal basis means less time spent editing (speaking from first-hand experience).
Limitations
New machine learning tools, like any other, have their use cases and their drawbacks. I spoke to a friend and IBM AI researcher, Nathan Herr, to understand what those look like. “Look, the truth is that a lot of what we do in offices could be automated with AI,” he told me. “I think of the corporate lawyer whose job is to apply precedent to new situations — it’s basically, ‘what is the situation, what rules apply, and what is the outcome?’” he said. “That could 100 percent be automated using AI, since it’s rules-based. But you see, the lawyer is charged with providing an opinion and is liable for it. Who would we blame if the AI gave the wrong answer?” When I asked Herr if white-collar jobs were at risk in the five-year horizon, he was skeptical. It could certainly save people time and resources, like most information technology, but until organizations grapple with issues of responsibility, completely outsourcing knowledge work to a machine is unlikely.
That means humans are firmly in the driving seat, at least for now.
Zed Tarar is completing an MBA at London Business School, where he specializes in the intersection of technology and policy. He has worked in five countries as a U.S. diplomat.
Disclaimer: While Zed Tarar is a U.S. diplomat, the views expressed here are his own and do not necessarily reflect those of the Department of State or the U.S. government.
Read Tarar’s previous miniseries on the relationship between technological innovation and diplomatic practice:
