Six capabilities underlying the potential of language models for political analysis and negotiation support. In different combinations, they power the Tools & Methods →.
Much of conflict analysis runs on open sources: mapping stakeholders, tracking positions, following how narratives evolve across news and social media. Traditionally, analysts have navigated this material through keyword search - specify a set of terms, and the system retrieves the documents that contain them.
This approach carries risks - relevant material is missed when expressed in vocabulary the analyst did not anticipate. Irrelevant material is returned when a keyword appears only incidentally.
Delegations arrived in Geneva today for a third round of ceasefire talks.
foundfoundBoth sides have agreed to meet in Muscat to discuss a truce.
missedfoundNegotiators are working towards a halt to hostilities before the rains.
missedfoundThe council committed itself once more to silencing the guns by year’s end.
missedfoundمذاکرات آتشبس در ژنو آغاز شد.
missedfoundOfficials called for a ceasefire in the stands after the derby.
founddroppedSemantic search retrieves material by what it means rather than which words it contains. An analyst can describe a position, narrative or line of argument, and the system finds passages that express the same underlying idea, even when the wording differs - and across languages.
In practice this means that news and social media, but also academic research and institutional knowledge bases, become searchable by what they say rather than the words they use - a way of cutting through very large volumes of material to what is relevant. The searching can also run unattended: alert agents scan incoming material at regular intervals and flag significant developments as they appear.
Where it breaks down
Semantic search retrieves only from the sources it has been given. If a relevant outlet, channel or language is missing from that collection, its material cannot be found - and the results give no sign of the gap: a search over a narrow set of sources looks just as complete as one across the whole spectrum. The decisive judgement therefore comes before any question is asked: choosing sources that honestly cover the actors, languages and positions involved, and being able to defend that choice. What is retrieved, in turn, must stay traceable - each finding linked back to the original passage, so that a claim can be checked against what was actually written rather than resting on a model's rendering of it, which can be confidently wrong.
In the tools
Two of the tools presented here are built on this craft: Media Monitoring, which reads a country's press across its political spectrum, and Telegram Monitoring, which follows public channels as the conversation unfolds. Several others draw on it more quietly - retrieving precedent from past peace agreements, or finding related positions across the records of past dialogues.
AI-assisted sensemaking makes qualitative analysis possible at a scale no team could reach manually. A language model can work through thousands of contributions and surface what people are saying: what they want, which concerns come up repeatedly, and on what issues different groups agree or diverge.
Six read closely
“Salaries have not been paid in months.”
“Companies come from outside to rebuild the road and leave - our engineers watch from the sidewalk.”
“They invite us to the workshop, photograph us, and decide without us.”
“Checkpoints decide who works and who does not.”
“Aid feeds us for a week; work would feed us for life.”
“Give us a role in monitoring the truce.”
The rest
Too many messages to read and sort manually.
What the material says
“Companies come from outside to rebuild the road and leave - our engineers watch from the sidewalk.”
“Reconstruction contracts never reach local firms.”
“Local contracts go to the same families every time.”
“Outside firms are the only ones that finish the work on time.”
The dissent is not about outsiders. It is about which local hands the money reaches.
“They invite us to the workshop, photograph us, and decide without us.”
“Youth are asked for opinions, never for decisions.”
“Reserved seats would make us a quota, not a constituency.”
“A seat means nothing while the parties pick who sits in it.”
Both sides want the same thing. They disagree on whether a reserved seat delivers it.
“You cannot vote freely where every neighbourhood has its own gunmen.”
“Weapons decide the result before the ballot does.”
“Disarmament first means waiting forever - nobody moves first.”
“Only an elected authority has the standing to ask anyone to disarm.”
The room is split almost evenly, and the split is about sequence, not about weapons.
This changes a longstanding trade-off in qualitative research: breadth versus depth. Faced with hundreds of responses, a team once had to choose between reading a small selection closely and treating it as representative, or reviewing everything by sorting responses into categories and counting them - losing much of the detail along the way. In either case, some of what people said would disappear from the analysis. Sensemaking at scale makes it possible to consider every contribution in its original form, however large the collection.
Where it breaks down
The limit is interpretation. A model can find patterns and suggest what they might mean, but it cannot decide which interpretation best fits the situation - or explain why that judgement is sound. Because its answers are so fluent, a tentative inference can easily sound as convincing as a well-supported finding. Checking the evidence, understanding the context and deciding what should happen next remain the responsibility of the analyst.
In the tools
Sensemaking forms the analytical core of many of the tools presented here, including AI Consultation and the two monitoring tools. Retrieval finds the relevant material; sensemaking examines what it says.
A language model begins with broad, general knowledge. It does not know what an organisation has learned from earlier processes, which approaches have already been tried, why certain decisions were made or which words and ideas matter to different parties. But it will still produce an answer.
Should vision building come before or after scenarios?
Vision first. The 2025 notes record scenario work grew sharper once groups had a shared destination to test against.
Context engineering gives the model access to that missing knowledge. It connects it with institutional reports, methodologies, process notes and past examples, while setting out how those sources should shape the task. The goal is to bring in the most relevant knowledge at the moment it is needed.
This turns institutional memory into something people can actively engage with, and the model can ground its recommendations in documented experience and point to the sources behind them.
In peace and conflict work, this might take the form of a specialised advisor on process design, or dialogue facilitation. When presented with a live question, it draws on the organisation's documented experience, follows its methodology and structures its response accordingly - building on accumulated practice rather than relying on general knowledge alone.
Where it breaks down
The main limitation is what gets selected. Institutional memory contains an organisation's blind spots as well as its lessons. Experience from one process may not apply to another, and anything left out will be invisible in the answer. Context engineering therefore requires careful judgement: what should be included, what should be left out, and how much weight should a past example carry?
In the tools
Among the tools presented here, context engineering is the main approach behind the Process Advisors. It is also central to Sensemaking in Dialogue and to adapting AI Consultation to a particular institution and context.
A language model can hold a conversation: it can ask a question, follow up on an answer and return to something said earlier. Stakeholder engagement makes use of this ability to create a two-way exchange. The system asks questions, listens and responds to what each person says - gathering their perspectives and priorities while also sharing relevant information.
This combines the reach of a survey with some of the responsiveness of an interview. A conversational system can engage many people remotely while adapting its questions to each person's answers. The exchange can take place on participants' own devices, through familiar services such as WhatsApp. They can respond by text or voice message, in their own language and at a time that works for them. This makes it possible to engage large stakeholder groups, across nearly all regions and contexts.
Because the exchange is conversational, participants are not limited to a fixed set of questions and answers. The system can ask for clarification, explore an issue in more detail or provide information in response to a participant's concerns. This can produce richer contributions while making the process more accessible and relevant to each person.
Where it breaks down
Remote consultation can widen participation, but it does not automatically make it inclusive or representative. Access to a device, reliable connectivity and digital confidence - as well as the choice of platform - shapes who takes part and whose voice is missing. In sensitive settings, privacy, surveillance and data security also affect what people feel safe enough to say. Participants need to know whether they can truly remain anonymous, how their contributions will be used and what influence they can realistically expect to have. Without a visible feedback loop, consultation can become extractive: people share their experiences but never learn whether they were heard or what followed. A conversational system can also steer or persuade, so its questions and responses require careful human oversight.
In the tools
Among the tools presented here, AI Consultation and the AI Interviewer use this approach to engage participants at scale through structured, responsive conversations.
Composition turns analytical work into clear, purposeful writing or visualisation: a briefing note, workshop report, infographic, summary or draft language for a process document.
Writing up has always been the slow end of the work. A workshop generates days of discussion; the report used to take weeks of rapporteur work and arrive after the moment it could have informed. With a model drafting from transcripts and analysis, a first version exists in minutes, and the same material can be rendered differently for different readers - a two-page note for a principal, a full report for the record - without redoing the analysis.
Where it breaks down
This is also where a model's fluency becomes particularly risky. A polished text can sound convincing whether the analysis behind it is strong or weak, and a well-written draft often invites approval rather than scrutiny. A model can produce a plausible report from thin or incomplete material just as readily as it can from sound evidence. Drafting may be delegated, but responsibility for anything published under a person's or organisation's name cannot be.
In the tools
Composition is in nearly everything presented here. Every tool ends its run by writing something a person will read - and act on.
Translation carries meaning across languages - positions, idiom and register, not just words. In Arabic, Farsi, Russian or French contexts, it decides whether the other five crafts are available to participants in their own language or only to analysts working in English.
Language has always been a filter on this work. Analysis ran on what existed in English or on what scarce interpretation could cover; participation was limited to those comfortable in a process's working language; translation was rationed to whatever seemed important in advance. Multilingual models loosen the filter: material can be searched and read in the language it was written in, and a participant can answer in their own dialect and be understood in full.
In practice this means consultations where Yemeni or Sudanese participants speak as they actually speak, not in a survey's formal register, and monitoring that reads a country's press in the country's own language rather than through wire-service English.
Where it breaks down
Translation is not neutral, and that is where it breaks down. Models behave differently across languages: research has shown the same model giving materially different casualty figures for the same events depending on the language of the question. Dialects poorly represented in training data are served worst - often the languages of the people the work most concerns. And the reader who needed the translation is the person least able to check it.
In the tools
Translation runs through everything multilingual on this site: it is what lets Media Monitoring read the Iranian press in Farsi, and what lets AI Consultation hear participants in their own words.