Machines at the Table: Why AI Diplomacy Needs Human Judgment, Not Faster Answers

EDITOR’S NOTE

“The Intercontinental Connection of Journalists, the global journalism platform of ICJ Global Media Group LLC, proudly presents this Diplomatic Analysis, prepared by Amro Shubair, a member of the ICJ Intelligence & Diplomacy Team, offering an independent and timely perspective on the evolving relationship between artificial intelligence and diplomacy.“

Every negotiation holds two conversations. One is spoken across the table. The other runs underneath, in the pauses, in the choice of one word over another, and in everything each side decides to leave unsaid.

Artificial intelligence is already very good at the first conversation. It can summarise a thousand pages of draft text before the coffee cools and set a single clause against decades of precedent. It struggles with the second one, and that second conversation is usually where agreements are made or lost.

Much of the public debate on AI diplomacy misses this. One camp expects machines to replace the diplomat, while the other treats the whole subject as a passing novelty. Neither view helps the people who actually have to decide how these tools enter their daily work.

The more useful question is narrower. Which parts of diplomatic judgment can safely be handed to a machine, and which parts break the moment they are?

That question matters far beyond foreign ministries. Any organisation that negotiates across difference now faces it, whether in a boardroom, a trade body or a humanitarian operation. What follows is one way to answer it.

The Oldest Craft Meets the Newest Tool

Diplomacy has lived through disruptive technology before. The telegraph turned weeks of correspondence into hours and narrowed the room envoys once had to act on their own judgment far from home. Broadcast media brought negotiations into public view, and secure digital channels later made instructions from headquarters almost instant.

Each change altered the pace of the craft while leaving its centre intact. That centre is the management of difference between parties who cannot simply impose their will on one another. It stays human because it depends on interpretation: of intent, of resolve, and of what a counterpart can realistically defend before the people behind them.

AI breaks with that pattern in one important way. The telegraph carried a message. A language model can write one, assess one and suggest how to answer it. For the first time, the tool is doing part of the judging instead of simply moving judgment from one place to another.

This is why practitioners need a working framework more than another forecast. Predictions about which roles will disappear go stale quickly. A clear sense of what can be delegated, and what cannot, will stay useful however fast the technology moves.

What Machines Do Well in Diplomatic Work

A fair assessment starts with the gains, and they are substantial. Diplomatic work carries a heavy informational load, and a good share of it is routine. Following how a phrase changes across twelve successive drafts is slow, careful work. So is scanning hundreds of public statements for a shift in tone, or assembling a briefing that pulls together economic data, legal commitments and earlier positions.

On tasks like these, AI tools save real time and widen coverage. A well configured system can notice that a counterpart has quietly dropped a qualifier it used for years. It can find the precedent an exhausted drafting team would overlook at two in the morning.

It can also give a small delegation the analytical reach that once required a large secretariat, and that point deserves more attention than it gets. Analytical capacity has never been evenly spread in multilateral settings, where small missions often follow many negotiating tracks with only a handful of officers. Used with care, AI could narrow that gap.

So the case for adoption is strong. The risk lies in misreading what these tools produce. A machine output looks like analysis, but it is more accurately a well informed first draft, and it still needs someone to own it.

Translation Is Not Interpretation

For many diplomats, machine translation is the first AI tool they use, and it is more deceptive than it looks. Moving words between languages is one task. Carrying meaning, tone and implication across them is a different and much harder one.

Diplomatic language is full of calibrated signals. Whether a text “notes,” “welcomes” or “endorses” something can mark a real shift in position. A tool that renders all three with the same warmth in another language may erase a distinction that took a delegation weeks to secure.

Register adds a second layer. Many languages express respect, distance and formality through grammar, honorifics and idiom. A literal rendering can sound curt where courtesy was intended, or overly deferential where firmness was the point.

None of this is an argument against machine translation. It is an argument for dividing the work sensibly. Machines can produce quick working drafts that open documents to more readers, while qualified interpreters and linguists remain essential wherever nuance carries weight. In diplomacy, that is nearly everywhere.

The same applies to listening. AI can now follow public conversation across many languages and platforms and pick out trends no team could track by hand. It can also miss irony, humour and coded reference, which is often where public feeling is expressed most clearly. An automated summary of public mood is a starting point for analysis and should never be mistaken for the conclusion.

The Signalling Problem: Why Speed Is Not Credibility

The limits of machines become clearer through one of the most influential ideas in bargaining theory. James Fearon showed that rational parties can fail to reach agreements both would prefer, because each holds private information about its capabilities or resolve and has incentives to misrepresent it.¹

Seen this way, negotiation is less about gathering information than about revealing it credibly. That difference matters enormously for AI. Machines are excellent at assembling what is already on the public record.

What decides a negotiation, however, is usually the information that has been held back, disguised or deliberately blurred. No model trained on public sources can see what a counterpart has chosen not to show. It can only infer, and its inferences carry every blind spot in its data.

Thomas Schelling made a related point decades ago. He observed that the power to constrain an adversary may depend on the power to bind oneself, and that in bargaining “weakness is often strength.”² A commitment is credible when backing away from it would cost something visible.

A message produced in seconds by a machine costs its sender almost nothing, and that ease can quietly empty it of signalling value. As AI makes polished diplomatic language cheap to produce, counterparts are likely to give polished language less weight. Attention will probably shift toward signals that are expensive to send: time, political capital, and the decision to show up in person.

Marcus Holmes gives a persuasive account of why showing up matters. His research argues that face to face meetings allow intentions to be read with more confidence than standard bargaining models predict, because human cognition is built to read intention through direct interaction.³

If he is right, the most automated period in diplomatic history may also be the one in which personal encounter is worth the most. Far from cheapening the meeting in the room, the machine may make it more valuable than ever.

The Ambiguity Dividend: Why Precision Can Hurt Negotiation

Machines are designed to remove ambiguity. Diplomacy often depends on keeping it.

Anyone who has watched a drafting committee spend an afternoon on a single adjective understands why. Some agreements exist only because their wording lets each party read it a little differently at home. A carefully chosen verb, a definition left for later, or a sequencing clause left open can hold together a settlement that tighter language would break. That vagueness is deliberate, and it is space built for compromise.

Robert Putnam explained the importance of that space through his model of two level games. Negotiators bargain with their counterparts and, at the same time, with the domestic constituencies who must ratify the result. Any agreement has to fall within what each side can carry home.⁴

Wording that allows more than one reading can widen that range of acceptable outcomes. A tool trained to favour clarity may treat this kind of constructive vagueness as a mistake. Ask it to tighten a draft and it may very helpfully close the exact gap that was holding consensus together.

The problem grows across languages. A phrase kept elastic on purpose in one language can harden into a single meaning once a machine renders it into another. Parallel texts need careful human reading at precisely the points where the drafters wanted room to move.

In each case the machine is not really at fault. The fault lies in a workflow that never told it what the ambiguity was for. Before any drafting tool touches a sensitive text, someone with authority has to decide which imprecision is intentional. That is a political judgment, and it cannot be handed to a system that bears none of the political consequences.

Perception at Machine Speed

Diplomacy has always been exposed to misperception. Robert Jervis showed that decision makers tend to fit new information into images they already hold, and that this produces patterned rather than random errors in judging what others intend.⁵

AI can reduce that problem or make it worse. A well designed tool can test assumptions by bringing forward evidence that cuts against the prevailing view. A badly designed or carelessly used one can do the reverse, returning material that confirms what its user already believes, in such volume that it starts to feel like proof.

Researchers in human factors described this hazard long before today’s AI existed. Raja Parasuraman and Victor Riley separated appropriate use of automation from “misuse,” which they linked to overreliance, monitoring failures and decision biases.⁶ They also described “disuse,” the neglect of automation after false alarms. An institution burned once by a flawed machine assessment may drop a useful tool entirely, and neither reaction helps.

What helps is calibration, meaning a clear sense in each case of how much weight an output deserves. The habits that build it are modest. Ask the tool for the strongest case against its own conclusion. Require a source for every factual claim. Put the same question several different ways and compare the answers. Keep a record of when the machine was right and when it was wrong, so that confidence rests on evidence instead of impression.

Speed makes all of this harder. When analysis arrives instantly, the pressure to reply instantly follows. Yet many consequential decisions in diplomacy are better for a deliberate pause, and the ability to let a statement settle before answering it may turn out to be a quiet advantage in an age of instant analysis.

The Accountability Gap: Who Answers for a Machine Recommendation?

Every diplomatic act has an author. A position is taken by someone authorised to take it, and that person answers for it. This chain of responsibility is what allows counterparts to hold a party to its word.

AI puts that chain under strain. Imagine a recommendation shaped by a model that contributes to a costly misjudgment. The officer relied on the tool, the tool relied on its training data, and the provider relied on its terms of service. Everyone touched the decision, and no one clearly owns it.

That diffusion of responsibility is a governance problem in its own right, and one of the more consequential questions AI raises for international relations. Ruth Grant and Robert Keohane offer a useful place to start. Their distinction between participation and delegation models of accountability, along with their account of seven accountability mechanisms, clarifies who can hold whom to account in world politics.⁷

The delegation model is particularly useful here. An institution that hands analytical work to a system has not handed over responsibility for the result. Any protocol on the use of AI in external affairs should say so plainly: every machine assisted judgment that leaves the building needs a named human owner.

Langdon Winner adds a deeper caution. In a widely cited essay, he argued that technical arrangements can embody particular forms of power and authority, and he described technologies as “ways of building order in our world.”⁸ Diplomatic AI will be no different. The defaults built into a system, such as what it ranks first, what it flags and what it leaves out, shape the options a negotiator ever gets to see. Choosing those defaults is a policy decision and should be treated as one.

Confidentiality: The Risk Hiding in Plain Sight

Governance begins well before any international forum. It begins at the desk. Diplomatic work depends on discretion, and negotiating positions, fallback lines and private assessments lose much of their value once they are exposed.

Many AI tools process information on external infrastructure, under terms that users rarely read in full. Pasting sensitive text into such a system can create exposure that nobody intended and few would notice. Because it is quieter than a breach, it is easier to overlook.

Good practice starts with clear handling rules. Institutions need to state which categories of material may go into which tools, with what safeguards and with what record of use. Those rules should be simple enough for a tired officer to apply correctly at the end of a long day.

Dependency is the longer term version of the same concern. An institution that relies on a single external provider for core analytical work also inherits that provider’s priorities, pricing and weaknesses. Building internal expertise, keeping alternatives open and preserving the skills to work without the tool are sensible forms of resilience.

Governing the Tools: Lessons From Expert Communities

If AI is to be governed internationally, practitioners do not have to start from nothing. International relations scholarship has long studied how technical complexity changes the way parties cooperate.

Peter Haas described epistemic communities as networks of professionals with recognised expertise and an authoritative claim to policy relevant knowledge. Their members share normative beliefs, causal beliefs, notions of validity and a common policy enterprise.⁹ Such communities help decision makers define problems and identify options when uncertainty is high.

AI governance fits that description closely. Few diplomats can audit how a model works, and few engineers can predict how a clause will be received in a tense plenary session. Good governance will depend on lasting connections between these two worlds, so that technical insight and diplomatic judgment meet early instead of colliding late.

Three principles follow from the argument so far.

1. Transparency of use. Parties to a negotiation have a legitimate interest in knowing whether machine tools helped shape a text. Voluntary disclosure norms could keep suspicion from undermining talks that would otherwise go well.
2. Human ownership. Every consequential output should have a responsible human author, in line with the accountability argument above.
3. Shared vocabulary. Agreements on AI will weaken if parties attach different meanings to the same core terms. Defining a concept such as human oversight with operational precision is diplomatic work in its own right.

The third principle contains a useful irony. Constructive ambiguity can save a settlement on many subjects, yet in the governance of AI it becomes a liability. On questions of safety, vague commitments lead to divergent practice, and divergent practice wears away the cooperation those commitments were meant to protect. Knowing when ambiguity helps and when it harms may be the clearest test of judgment in this field.

The New Diplomatic Skill Set

What does this mean for the people who practise diplomacy and for the institutions that recruit them? It does not mean every diplomat must become an engineer. It means the established skills of the craft now need a technical counterpart.

A practical place to begin is a simple delegation map, separating what machines can carry from what must stay with people.

The map is meant as a starting point. As the tools improve, some tasks will move toward a larger machine role. The last three rows are unlikely to change, because they depend on things that cannot be computed: knowledge that has been withheld, commitments that carry real cost, and personal accountability.

Working along that map calls for a recognisable set of abilities. The first is asking precise questions, framed so that a system exposes its uncertainty instead of hiding it. The second is interrogating outputs, treating each one as a claim to be tested and asking what it rests on and what it may have missed.

The third is knowing when imprecision is a defect and when it is the deal itself. The fourth is reading signals well, with the understanding that as polished language becomes cheap, value moves toward presence, patience and political capital. The fifth is ownership, which means accepting personal responsibility for decisions informed by tools whose workings remain partly opaque.

None of this is limited to diplomatic service. Corporate negotiators, trade specialists, humanitarian coordinators and policy advisers face the same questions in their own work. Any organisation that bargains across difference will need people who can pair machine speed with human judgment, and that pairing deserves to be treated as a core professional skill.

Institutions can act on this now. Training can combine negotiation simulations with exercises in auditing machine outputs. Recruitment can favour people who move easily between policy and technology. Performance reviews can reward good judgment about when not to use a tool, which is often harder than knowing how to use one.

Conclusion: Where the Machine Ends

AI will not end diplomacy. It will show which parts of the craft were always routine and which were never replaceable. The routine work will be automated, and that is welcome, because it frees skilled people from clerical burden.

What remains is the hardest and most human part of the job: reading intention, signalling resolve, protecting the useful ambiguity that lets rivals agree, and standing personally behind the outcome. None of that was ever about processing speed. All of it is judgment exercised under uncertainty, with consequences attached.

The practitioners who shape the coming decade will neither resist these tools nor surrender to them. They will know exactly where the machine’s answer stops and their own responsibility starts.

In an age of intelligent machines, the decisive advantage at the table still belongs to whoever knows what not to say.

Notes

1. James D. Fearon, “Rationalist Explanations for War,” International Organization 49, no. 3 (1995): 379 to 414.
2. Thomas C. Schelling, “An Essay on Bargaining,” The American Economic Review 46, no. 3 (1956): 281 to 306.
3. Marcus Holmes, “The Force of Face-to-Face Diplomacy: Mirror Neurons and the Problem of Intentions,” International Organization 67, no. 4 (2013): 829 to 861.
4. Robert D. Putnam, “Diplomacy and Domestic Politics: The Logic of Two-Level Games,” International Organization 42, no. 3 (1988): 427 to 460.
5. Robert Jervis, “Hypotheses on Misperception,” World Politics 20, no. 3 (1968): 454 to 479.
6. Raja Parasuraman and Victor Riley, “Humans and Automation: Use, Misuse, Disuse, Abuse,” Human Factors 39, no. 2 (1997): 230 to 253.
7. Ruth W. Grant and Robert O. Keohane, “Accountability and Abuses of Power in World Politics,” American Political Science Review 99, no. 1 (2005): 29 to 43.
8. Langdon Winner, “Do Artifacts Have Politics?,” Daedalus 109, no. 1 (1980): 121 to 136.
9. Peter M. Haas, “Introduction: Epistemic Communities and International Policy Coordination,” International Organization 46, no. 1 (1992): 1 to 35.