What is "AI-Powered Diplomacy" anyway?

I just finished Dan Spokojny’s great article “Should Diplomats Trust AI?” The piece is anchored by two new items: the recent Foreign Service Journal issue that focused heavily on AI (which I contributed to), as well as the December 2025 AI-Powered Diplomacy paper from the Pavel Slunkin writing at the Belfer Center. Dan expresses skepticism about the claims in the Belfer study, as well as the broad claims that AI has the potential to render diplomats irrelevant or obsolete. 

I generally agree with Dan on this, though I want to offer a bit of nuance to that, particularly on why I think that AI will probably reinforce the need for human diplomacy, rather than curtail it. 

The core of diplomatic work is engagement with people and institutions to build relationships, inform and shape perspectives, and reach agreements* with those people/institutions in order to advance foreign policy objectives. 

That core work is unlikely to ever get automated, because it’s probable that organizations that diplomats interface with are generally still going to be run by humans, and civil society (comprised of actual, real people!) is likely to remain central in shaping how governments and large institutions behave. Even the most wildly optimistic AI proponents generally believe that humans will remain central to decisionmaking and strategy for large organizations, albeit given new powers of automation by AI and other tools. You can just look at the leading signal among the new class of emerging, AI-driven organizations that automate a huge number of tasks. The humans in charge are actually more powerful individually – they control more economic direction in their organizations if automation is enabling higher productivity per person. That suggests to me that as long as people and institutions make decisions around the world–in governments, businesses, and civil society–the human aspect of diplomacy is likely to matter. And an optimistic AI scenario where individuals work across greater spans of control may increase the opportunity for human relationships to make a big difference.   

Now here’s the flip side of that. While the front-line diplomatic work isn’t readily replaceable with technology, there are a lot of individual actions that are part of making diplomacy work–both actions for front-line diplomats and for the support functions that make it possible. Those individual functions can often benefit from automation or AI. And here’s where I’d introduce some nuance: not all automation is AI, not all AI is large language models (LLMs), and not all LLM use is equal. Just to break this down: 

• Automation is the use of technology to perform tasks and processes with minimal human intervention, often through forms or patterns that use deterministic logic (e.g., If/then). Automated approval routing on forms is an example of this.  
• AI is a broad category of technology that typically uses pattern-matching. It is broadly probabilistic (i.e., how similar is this to other things it has seen, and how confident is the machine in a match). Predictive analytics, like forecasting whether stocks will go up or down given other macro trends, is a good example of general AI – as is image detection. 
• LLMs are specialized subset of AIs trained on large amount of text, which gives you the ability to match text (language) patterns, again in a probabilistic way (the best fit for the next collection of words, given the previous ones). ChatGPT’s chat function is (mostly) LLM. 

Why do I say mostly? Well, ChatGPT uses an LLM to respond to your text inquiries. But it uses a different type of AI (diffusion) to generate images. And it uses a yet different type of AI to read an image. If you give it a PDF, it may be using a non-AI Optical Character Recognition tool (OCR), which is just garden-variety automation. 

Almost all the applications we talk about saving time for diplomats are automations of some form, while only some of them are AI. Often people use AI tools when a more conventional automation might actually be a better and more reliable fit for the problem they’re working on–largely because AI tools are available to them, and quality process automation tools are not. 

This is where the nuance and slight divergence from Dan’s perspective rests. I absolutely agree that LLMs have limitations and they may not pan out as the technology, and I’m also skeptical about the claims of AI super intelligence. I also think there are substantial risks to deskilling by adopting AI the way some people advocate, including Pavel’s paper. 

At the same time, there is so much about the work of diplomacy that would be improved, not weakened through automation. I don’t think anyone benefits from highly paid FSOs fixing the margins on memos, or in manually pulling together Permanent Change of Station (PCS) packages. I don’t think there’s a substantial loss if drafters have tools to retrieve and summarize the key points from a previous meeting. This isn’t fantastical stuff–in contrast with the dreams of a fully automated policy analysis engine that tells you what the talking points or the policy should be. LLMs, out of the box, don’t (and can’t) do this well. Those are complex processes, and I completely agree that LLMs don’t tend to model complexity well on their own–for reasons that could fill another entire post. 

In some ways, this is where Dan and I converge. While I think automation and AI can solve a lot of problems, it can’t replace the core work of diplomacy. It also is unlikely to replace strategic thinking and human engagement in diplomacy, just as it’s unlikely to replace those in other organizations (including almost all the AI labs, which continue to hire in marketing/storytelling, sales, and other related roles). It can and should change what the day-to-day roles look like, and I hope continue to push diplomats' toward high-value and high-expertise uses of time. Continued automation will almost certainly reduce the need for people to do some of the administrative support in EX offices and support functions (e.g., manual travel voucher reviews)--but that is just an extension of a trend that started in all organizations 30+ years ago.

One of the questions I am still chewing on is how the early career development will change for junior diplomats. That parallels the good point from Miles Williams cited by Dan that “The people I see using AI effectively in their work today are people who developed expertise and skills before AI tools were readily available… they know how to double-check and refine what their AI agents produce because they already know how to do the tasks they outsourced.” Figuring out how to cultivate people with critical minds, familiarity with the tools, and deep expertise is an interesting, but important puzzle.