Claude for Clausewitz?

Since the arrival of ChatGPT in late 2022, policy wonks, in and out of government, have explored how best to rope the bots in as support (Opens in new window). A few examples:

The CIA apparently has chatbots (Opens in new window) who play the role of various (real) world leaders, presumably to understand their intentions better.

Researchers have looked at promises and pitfalls (Opens in new window) of fusing AI with wargaming (Opens in new window). Large-scale social simulations with AI agents (Opens in new window) is a promising frontier of computational social science, along with AI for strategic assessment and early warning (Opens in new window).

Political risk outfits integrate (Opens in new window) AI-generated assessments with human judgement. Hobbyists routinely vibe-code (Opens in new window) dashboards that track specific conflicts, often to rather murky effects (Opens in new window).

But you don’t have to be a spook or a nerd to put the machines to work. Depending on your interest, several possibilities present themselves.

Capable GenAI models allow easy data analysis (Opens in new window), even for those who are not so disposed temperamentally. The more technically capable can also use AI agents (Opens in new window) – models given a task to accomplish specific goals near-autonomously – to simplify routine work such as preparation of bibliographies by extracting relevant information from files. Newer models are also quite capable of fact-checking, and synthesising a large volume of supplied information (Opens in new window) coherently.

There are deeper options for the adventurous.

GenAI can be used for scenario creation and analysis. Language models are capable of generating coherent stories given broad guidance. Many private firms already use (Opens in new window) this feature to identify plausible futures (Opens in new window). In fact, AI-based scenario generation is an active area of research (Opens in new window). And if fictional scenarios are what you are after, factual hallucinations – a known AI affliction – is hardly a deal breaker.

But the key to DIY lies in very careful and continuous prompting. I recently conducted an experiment to see whether a frontier model could write a very concrete quasi-fictional scenario and then perform an alternative futures analysis in a prescribed format. While the results are encouraging – details here (Opens in new window) – a point of caution. Chatbots can sometimes be fluent and inconsistent at once. Scenarios and futures analysis reward creativity but not logical and causal indiscipline. As AI-generated scenarios become involved, builders have to be on guard.

Paradoxically, GenAI models are also extremely good at close reading. Given well-defined instructions, they can flag possible contradictions and jumps in argument. These are useful abilities in stress-testing drafts as well as archival records. In another recent experiment (Opens in new window) I had a model assess a declassified memorandum against a fixed set of analytical standards. What stood out was its ability to parse arguments granularly, and accurately summarise the broad thrusts of the memo. The bot’s pedantry – it termed the experiment an “audit” – was impressive. At the same time, it made mistakes, which it caught in a self-directed review.

You don’t have to be a spook or a nerd to put the machines to work.

It may well be that at some point in the near future language models become a core part of qualitative research (Opens in new window) and analysis. But obstacles remain.

The first is economics. Frontier model subscriptions are expensive for many in the developing world. There is a real risk that this leads to an uneven playing field between those with access and without. And this diminishes analytical diversity.

The second issue is technical. Model outputs remain stochastic – the same prompt twice can often return slightly different outputs. Mathematicians already had a (non-AI) technology – in the form of proof checkers (Opens in new window) – which can certify the validity of an AI-supplied mathematical answer. This is the main reason why GenAI has found ready allies in that community. No such tool is available to other disciplines.

But the most important obstacle may be sociological. For many analysts and scholars, AI use is viewed pejoratively, tainted by association with intellectual theft. Evangelists must – through words and deeds – demonstrate the gains from transparent and ethical use.