Imagine a large organisation conducting an internal review. Staff in different offices are asked to complete a detailed questionnaire about priorities, management, internal problems, perhaps morale, and performance. The answers are collected, personal information is removed, and the remaining material is fed into an artificial intelligence system. The machine is asked to identify contradictions, recurring concerns, unusual patterns, strengths and weaknesses across dozens or perhaps hundreds of pages.
There is nothing particularly alarming about this. In fact, anyone who has spent part of a career reading institutional reports will understand the attraction immediately. Human readers become tired. They overlook repetitions, remember the more colourful examples and inevitably bring their own expectations to the exercise. A machine can compare every answer with every other answer in seconds and may notice inconsistencies that even an experienced reviewer would miss.
Now imagine the same exercise a year later. The people completing the questionnaire know how their responses will be analysed, and they too have access to artificial intelligence. Faced with an awkward question, somebody will ask a model for help. Does this answer sound defensive? Can I acknowledge the problem without making it appear that management has lost control? Is the tone confident without becoming complacent? Have I contradicted something we said earlier?
None of this requires dishonesty. Quite possibly the opposite. The person answering may simply want to express a complicated reality clearly and avoid being misunderstood by the automated system on the other side. Yet the nature of the exercise has changed. The machine conducting the assessment is no longer analysing only what a group of human beings decided to say. It is increasingly analysing what another machine helped those human beings decide how to say it.
There are already more mundane versions of the same problem. A citizen disputing a parking fine can photograph the notice and ask a model to prepare an appeal. The municipality receives a perfectly structured objection, perhaps citing the relevant regulations. Multiply that by several thousand citizens and the municipality has an equally good reason to use another model to classify the objections and help draft the replies. The parking fine remains entirely real; much of the conversation surrounding it may no longer have been written by either of the people involved.
Both sides have become more efficient, at least in a technical sense. They can produce and process far more communication than before. Whether they understand one another any better is less obvious.
I have been thinking about this because diplomacy presents an unusually consequential version of the same problem. Other professions use language to describe much of what they do; diplomats conduct a remarkable amount of their work through language itself. Reports, instructions, talking points, speeches, démarches, diplomatic notes, communiqués and negotiated texts do not merely record diplomatic activity. Very often, they are the activity.
That makes artificial intelligence unusually well-suited to diplomacy.
When Language Becomes Almost Free
For generations, foreign ministries built large bureaucracies around the production and interpretation of words. Embassies reported, headquarters read, instructions travelled back in the other direction, and officials compared formulations or searched archives. Occasionally they spent extraordinary amounts of time arguing over a single adjective because everyone understood that the adjective might matter later.
Early in my diplomatic career, while I was still trying to understand how a foreign ministry actually worked, I spent a good deal of time reading the reporting that moved between embassies and headquarters. The quantity was impressive. So was the amount of effort devoted to producing it.
I remember making a rather irreverent observation at the time: sometimes diplomacy seemed to consist of people writing reports they did not particularly want to write for other people who did not particularly want to read them.
That was unfair, of course, but not entirely wrong. Much excellent reporting was produced, and still is. Yet there were also reports whose informational content differed surprisingly little from what a competent news agency had already provided. What justified the additional diplomatic layer was not supposed to be the repetition of facts. It was access, context, institutional memory and, ultimately, judgment.
I have thought about that old observation again recently because artificial intelligence appears, at first sight, to offer an almost perfect solution to the problem. If diplomats dislike spending hours turning information into routine reporting, a machine can help draft it. If officials at headquarters lack the time to read everything that arrives, another machine can summarise it, compare it with previous reporting and extract whatever appears significant.
The reports nobody particularly wanted to write can now be written by machines, while the reports nobody had time to read can be read by machines as well. That is undeniably efficient, but it leaves an awkward question behind: if a report was not important enough for a human being to write and not important enough for another human being to read, why exactly did it need to exist?
The question is becoming less theoretical. Large language models are exceptionally good at many of the activities on which foreign ministries spend enormous amounts of time. They can read large quantities of text, summarise them, compare versions, translate between languages, retrieve previous statements, identify inconsistencies and produce polished drafts. The US State Department, for example, has already deployed StateChat across much of its organisation, allowing officials to use an internal generative-AI system for precisely this kind of work.
There are good reasons for this. The productivity gains are obvious, particularly for organisations facing ever larger flows of information without corresponding increases in staff. Anyone who has spent an evening turning a day’s worth of meetings into reporting for headquarters is unlikely to object in principle to a machine offering to do the first draft.
The less obvious consequence is that language used to have a cost. A ten-page report required somebody to spend several hours writing it. A detailed response from headquarters required somebody else to read the report, decide that it mattered and devote time to answering it. Since not everything could be written and certainly not everything could be read, somebody had to decide what deserved attention. That decision was itself information.
When an experienced ambassador sent an unusually long report about an apparently minor political development, the length told headquarters something before the first paragraph had been read. The reverse was equally true. There were occasions when three carefully chosen paragraphs from someone who normally wrote ten pages deserved considerably more attention than another embassy’s urgent six-page cable. People at headquarters understood the difference because they knew the authors.
Artificial intelligence changes that calculation. If a system can turn notes, press reports, meeting records and open-source material into twenty competent reports before lunch, the scarcity of diplomatic prose begins to disappear. Much of that will be useful, but the ministry now faces a different problem: if producing analysis becomes almost free, attention becomes even more expensive.
The question is therefore no longer simply whether AI can produce a competent diplomatic report. Increasingly, it can. The more difficult question is what a report means once producing one no longer tells us very much about the human attention that preceded it.
The Reader Changes the Writer
Anyone who has worked in a bureaucracy knows that people adapt to the way they are assessed. If headquarters rewards short reporting, reports become shorter. If senior officials want three recommendations rather than ten pages of analysis, embassies learn to provide three recommendations. If a minister dislikes bad news, the bad news does not necessarily disappear, but its journey towards the minister tends to become more complicated.
Artificial intelligence accelerates this process. If I know that my report will be summarised by a machine, I have an incentive to write in a way that survives summarisation. If I know that an automated system will compare my answers and search for contradictions, I can ask another automated system to perform the same check before submitting them. If certain formulations are routinely classified as uncertain, negative or evasive, people will eventually learn to formulate around them.
For most of my professional life, when I wrote a political report I imagined another human being reading it. Sometimes I even knew who that person would be, and that mattered. One wrote differently for a colleague who knew the country well than for somebody encountering the issue for the first time. A cautious sentence from a diplomat known for optimism could carry more significance than an alarming paragraph from someone who predicted a crisis every fortnight. Institutional memory included memory of people.
A machine has a different kind of memory. It can compare today’s report with thousands of previous reports, identify changes in terminology and retrieve references that no human reader could reasonably remember. That is an extraordinary advantage. But once the writer knows that the machine is doing this, the writer adapts again.
At that point the subject stops being mainly about productivity. The relationship between diplomatic writing and diplomatic reading has begun to change.
The Problem of Saying Almost Nothing
Diplomatic language presents a particularly awkward case because ambiguity is not always a defect.
Consider the familiar formulation:
We have taken note of the proposal.
Anyone outside diplomacy might reasonably conclude that the sentence means exactly what it says. A diplomat would be more careful. Depending on context, it might indicate cautious interest, polite rejection, a desire to gain time or merely an unwillingness to reveal a position. Its meaning may depend less on the words themselves than on who said them, what was said previously, what was conspicuously omitted and what happened before the statement was issued.
Diplomacy has developed an extensive vocabulary for allowing several interpretations to coexist. Negotiated texts sometimes survive precisely because two governments can read the same sentence differently. Strategic ambiguity is not necessarily bad drafting. Occasionally it is the achievement.
Now give a machine the task of identifying what the sentence really means.
One recent study did something that would have sounded rather eccentric only a few years ago: it asked large language models to identify hints and implicit meanings in Chinese and Russian foreign-ministry statements. The results suggest that specialised systems can become quite capable of doing this, but the mistakes are at least as interesting as the successes. The researchers encountered semantic over-interpretation, misclassification and cases in which ordinary literal statements were treated as though they contained hidden messages.
Their preliminary tests produced a detail that deserves to be enjoyed by diplomats everywhere. Telling a general-purpose model to behave as a “senior diplomat” did not made no difference whatsoever.
Human diplomats, admittedly, are perfectly capable of discovering intentions that the supposed sender never intended. International relations offer enough examples of that without assistance from artificial intelligence. The difference is scale. A machine can misread a great deal of diplomacy before breakfast.
The possibilities nevertheless remain impressive. A system can compare years of foreign-ministry statements, notice that a familiar phrase has disappeared, identify a subtle change in terminology and do so across languages and quantities of material beyond any individual analyst’s capacity. Governments will want such capabilities, and quite reasonably so.
The Other Foreign Ministry’s Model
States have always communicated with several audiences simultaneously. A foreign minister speaking at the United Nations may be addressing another government, domestic voters, allies, financial markets and his own bureaucracy at the same time. Skilled diplomatic language allows those audiences to hear somewhat different things without making the contradictions impossible to manage.
Artificial intelligence introduces another audience: the other foreign ministry’s model.
That sounds slightly absurd until one considers the incentives. If a system can analyse years of speeches and identify changes in terminology, governments will use it. If it can compare today’s communiqué with hundreds of previous formulations and flag an unusual omission, analysts will want to know. If it can detect whether another government’s language on trade, Taiwan, Ukraine or sanctions has shifted slightly, the resulting assessment will eventually reach somebody’s briefing folder.
Once both sides understand how the other’s system is likely to read the language, that knowledge will begin to affect how the language is written. This need not involve deception. Governments already choose words knowing that journalists, markets, intelligence services and allies will analyse them. The machine simply becomes another reader whose likely interpretation must be considered.
Except that the machine may also have helped write the statement in the first place.
Imagine a ministry asking its internal system to draft language that expresses serious concern without signalling an intention to escalate. Officials review the draft and release it. Another ministry’s system compares the statement with previous language and concludes that the risk of escalation remains limited. That assessment is summarised for senior officials, perhaps by yet another model.
No human being has disappeared from the process. Ministers still decide, officials still approve, ambassadors still explain and governments remain responsible for what they say. Yet a growing part of the linguistic exchange has been produced, filtered and interpreted by machines. The bureaucratic example with which we began has become diplomatic: one system helps the sender formulate a message that another system has been designed to interpret.
Diplomacy has always contained feedback of this kind. What is new is the speed and scale with which it can now occur.
What Diplomats Actually Know
The conventional answer to all this is that machines will perform routine tasks while diplomats provide judgment. I think that answer is probably correct, but it is too comfortable.
Every profession confronted with automation eventually discovers that “judgment” is the word used for whatever remains after the machine has taken over everything that can easily be described. Diplomacy therefore needs a more demanding account of what human judgment actually consists of.
A machine will almost certainly become better than any individual diplomat at recalling previous statements, comparing documents, detecting textual changes and processing information across languages. It may become better at producing routine diplomatic prose as well. Anyone who has read enough bureaucratic writing should perhaps greet that prospect with a certain amount of optimism.
What remains harder to capture is the human history surrounding the text.
A diplomat may know that a minister sounded confident in public but worried in private. He may remember that a supposedly insignificant official has the president’s trust. She may notice that a counterpart who normally talks for an hour ended today’s meeting after twenty minutes. A sentence in a communiqué may look routine to anyone reading the document later but matter greatly to the people in the room because one delegation fought for two hours to keep it there. None of this necessarily appears in the meeting record.
Much of diplomatic knowledge has never been contained entirely in diplomatic documents. It resides in relationships, accumulated impressions, personal credibility, memory and the ability to understand why a particular person chose a particular formulation at a particular moment.
Perhaps machines will learn much of this as well. It would be foolish to draw permanent boundaries around a technology changing this quickly. There is, however, a more immediate problem. As AI improves the language itself, some of the human evidence contained in language may disappear.
If every report becomes better written, it may become harder to know who wrote it. If every argument is balanced, every weakness properly acknowledged and every conclusion carefully calibrated, the peculiarities through which experienced readers recognise individual judgment begin to fade. A mediocre writer with excellent political instincts and an elegant writer with poor judgment may start to sound remarkably similar.
Diplomacy could gain clarity while losing something less easily measured: texture.
There is also a more prosaic problem, and bureaucracies tend to discover prosaic problems eventually: responsibility. Suppose an AI-generated summary concludes that another government has softened its position. The assessment reaches senior officials, influences a decision and later proves wrong. Who made the judgment? The diplomat who approved the summary, the analyst who framed the question, the institution that selected the system, or the model that produced the inference?
The formal answer will presumably be that a human official remained responsible, and that is probably how it should be. Yet bureaucracies have always possessed considerable ingenuity when it comes to diffusing responsibility. Artificial intelligence introduces an unusually convenient additional participant.
More Communication, Less Meaning?
None of this is an argument for keeping artificial intelligence out of foreign ministries. That would be both unrealistic and probably undesirable. The volume of information is too large, the analytical possibilities too valuable and the administrative savings too obvious. A ministry that refused to use these tools would likely become less capable, not more principled.
The more serious question is whether institutions understand what changes once AI moves from assisting diplomatic work to mediating diplomatic language itself.
The bureaucratic examples are instructive precisely because nobody needs to behave badly for the feedback loop to emerge. The institution uses AI to analyse human responses; humans learn how those responses are analysed and use AI to improve them. Citizens use AI to communicate more effectively with government; government uses AI to cope with the resulting volume. Each individual decision is perfectly rational. The cumulative result is less straightforward.
For centuries, bureaucracies were constrained partly by the cost of producing and processing language. Generative AI is removing that constraint at both ends. We may therefore be approaching an odd institutional condition in which the supply of competent language becomes effectively unlimited while human attention remains stubbornly finite.
Perhaps the result will be excellent. Artificial intelligence may remove an enormous amount of bureaucratic labour, allowing diplomats to spend less time producing documents and more time talking to actual people. If that happens, foreign ministries may rediscover activities that many diplomats entered the profession to perform in the first place.
There is another possibility. The more efficiently machines produce, summarise and interpret diplomatic language, the easier it becomes to confuse the processing of diplomatic communication with diplomacy itself.
For all its cables, notes, instructions and communiqués, diplomacy exists because states cannot understand one another through documents alone. They send people because intentions are uncertain, language is incomplete, trust is personal and political relationships cannot be reduced entirely to information. Artificial intelligence may become exceptionally good at reading the cable, and before long it may also have written most of it.
The question then is no longer simply whether the machine understood what the diplomat meant.
It is whether there was still a diplomat who meant something in the first place.
Christopher Angel — currently posted, officially silent, and biting anyway.