Intelligence agencies spent seventy years learning to think clearly about a deceptive, uncertain
world. The methods are public. AI can now run many of them at scale — and quietly break others. Here are the
highest-leverage analytic techniques, the ways of seeing behind them, and what AI does to each.
Drawn from the public literature — the CIA's
Tradecraft
Primer (2009), Richards Heuer's Psychology of Intelligence Analysis, Sherman Kent's
Words of Estimative
Probability, and Philip Tetlock's Superforecasting — and connected to our own
research platform, which is built to run these techniques.
The core insight
Heuer's central finding, and the reason this whole discipline exists: the analyst's problem is rarely too
little information. It's how the mind processes what it has. We see what we expect, close on the first good
answer, and assume our adversary thinks like us. Structured techniques exist to externalise
reasoning — to get it out of one head and onto a page, in a form others can challenge and build on.
That is also exactly what makes them a fit for AI: a technique that forces the steps into the open is a technique
a machine can help run, and a human can check.
The warning that comes with it: AI amplifies the biases too. A model that agrees with you is confirmation
bias on tap; a model that fabricates a citation is manufactured evidence; a model trained on your own culture
mirror-images by default. So the rigour matters more with AI in the loop, not less.
Six ways of viewing the world
Before the techniques, the mindsets — the lenses that separate deep analysis from opinion.
Estimative, not predictive. The job is not to be certain; it's to state calibrated odds
and be honest about them. When Sherman Kent wrote "serious possibility," colleagues read it as anywhere from
20% to 80%. The fix: attach numbers to your words, and score yourself over time (Tetlock's
superforecasters won on calibration and base-rate training, not genius).
Outside view first. Start from the base rate — how often things like this happen in
general — then adjust for the specifics. Most overconfidence is a failure to anchor on the base rate.
Competing hypotheses, judged by disconfirmation. Don't ask what supports your favourite
theory; ask what would refute each one. The surviving hypothesis is the one with the least evidence
against it.
The mosaic. The picture assembles from many public pieces, none decisive alone. Patents,
filings, hiring, permits and talks together reveal what no single source states — the legal heart of
open-source intelligence.
Indicators before events. Decide in advance what you would observe if each
scenario were unfolding. Then watch for those, instead of rationalising whatever happens after the fact.
Assume the possibility of deception. A source that is too good, a signal that is too
clean, an absence where there should be noise — all are data. The question "what would I see if I were being
deceived?" is itself a technique.
The high-leverage techniques
The Tradecraft Primer sorts these into diagnostic (what's going on), contrarian (challenge it), and
imaginative (what else could be true). For each: what it is, and what AI does to it.
Diagnostic
Key Assumptions Check. List every assumption your analysis rests on, then ask which
would break it if wrong. The single cheapest, highest-return technique — most analytic failures trace to one
unexamined assumption.
AI amplifier: a model is good at surfacing the assumptions you didn't notice you were making, and
at generating the "what if this is false" chain for each. The trap: it will also confidently invent assumptions
you don't hold — keep a human editing the list.
Analysis of Competing Hypotheses (ACH). Heuer's method: lay out all hypotheses, rate
each piece of evidence as consistent or inconsistent with each, and rank by inconsistency. It defeats
confirmation bias by construction.
AI amplifier: generating the full hypothesis set — including "deception" and "nothing
happened" — fights the most common ACH failure, a missing hypothesis. The trap: let a different model, or a
human, fill the matrix than the one that proposed the hypotheses.
Quality-of-information check. Rate each source for reliability and each claim for
credibility (the Admiralty code: A–F for the source, 1–6 for the information), and trace a claim to its earliest
appearance. Repetition is not corroboration.
AI amplifier: deduplicating syndicated copies and tracing first appearance at scale is exactly
what the collection platform does. The trap: a model rates fluent text as more credible — reliability is about
the source, not the prose.
Indicators / signposts of change. Define observable markers for each scenario ahead of
time, then monitor them.
AI amplifier: continuous monitoring is cheap now — an agent can watch filings, feeds and news
for your indicators daily and flag movement. This is where AI most clearly beats a human analyst's attention
span.
Contrarian
Devil's advocacy. Assign someone to build the strongest case against the
consensus.
AI amplifier: a model has no ego and no career risk, so it will argue the opposite as hard as you
ask. A standing "red-team this conclusion" pass is almost free.
Red-team / think like the adversary. Reason from the other side's goals, constraints and
culture, not yours.
AI amplifier: role-played adversary analysis at scale. The trap: mirror-imaging gets worse, not
better, if the model shares your assumptions — feed it the adversary's actual doctrine, not your guess at it.
Premortem. Assume the conclusion turned out catastrophically wrong, then work backwards
to explain how. It surfaces failure modes that a "what could go wrong?" never does, because it starts from
failure as a fact.
AI amplifier: models are excellent at generating many distinct failure stories fast. One of the
highest-value, lowest-cost uses of AI in analysis.
High-impact / low-probability analysis. Deliberately examine the unlikely-but-devastating
tail, so it isn't dismissed just because it's improbable.
AI amplifier: generating plausible tail scenarios you'd otherwise never staff. The trap: don't let
ease of generation inflate your probability of the tail.
Imaginative
Alternative futures / scenarios. Build several internally consistent futures (often a
2×2 of the two biggest uncertainties) instead of one point forecast.
AI amplifier: drafting four rich, distinct scenarios from two axes in minutes, each with its own
indicators.
What-if analysis. Assume a surprising event has happened and explain it. This
loosens the grip of "it hasn't happened, so it won't."
AI amplifier: a model will take the premise seriously without the human reflex to dismiss it.
Outside-in thinking & structured brainstorming. Start from the broad forces
(technological, economic, regulatory) and work inward; generate widely before converging. Starbursting — asking
who/what/when/where/why/how systematically — is the simplest form.
AI amplifier: divergence is a model's strength; it will enumerate factors and questions a tired
analyst misses. Convergence still needs human judgment.
What AI actually changes
Put together, AI shifts the analytic bottleneck. It was always generation — one analyst can only
imagine so many hypotheses, failure modes and scenarios, and fatigue narrows that further. AI removes that
limit: the full hypothesis set, twenty premortems, continuous indicator monitoring, an adversary's case argued
in full, all cheap. That genuinely fights the mind's worst habit, premature closure.
But it adds three new failure modes, each a classic bias with a new on-switch:
Sycophancy is confirmation bias on tap. A model nudged toward your view will support it
eloquently. Always make it argue the other side too, and never let the generator be the only judge.
Hallucination is fabricated evidence. A confident, well-formatted citation to a source
that doesn't exist is the most dangerous output in analysis. Ground every claim; our platform's citation
check flags any source it can't trace.
Mirror-imaging scales. A model reflects the culture it was trained on, so "think like the
adversary" fails silently unless you feed it the adversary's own material.
The rule that falls out: AI generates, the human judges. Use it to open the aperture — more
hypotheses, more challenges, more scenarios — and keep a person on the disconfirmation, the source reliability,
and the final estimate. That division is the whole game.
The shortlist: better than the rest
If you adopt only a few, these compound into most of the value, and all run well with AI generating and a
human judging:
Key Assumptions Check — cheapest insurance against the single wrong belief.
Analysis of Competing Hypotheses — the core engine; disconfirmation over confirmation.
Indicators defined in advance — turns monitoring from hindsight into foresight, and it's
where AI's tireless attention wins outright.
Calibrated probability, scored over time — the only way to know if any of it is working
(Kent + Tetlock). Write the odds down; check them later.
Premortem — the highest insight-per-minute contrarian move.
Mosaic collection with source-reliability and first-appearance tracing — the input
discipline that makes the rest trustworthy.
Run them together and you have a working analytic method: gather the mosaic with graded sources, lay out
competing hypotheses, check your assumptions, define what you'd see if each were true, premortem your leading
answer, and commit to a calibrated estimate you'll score later. AI does the heavy generation at every step; the
judgment stays human. That's the platform we're building — the collection half
and now the analytic half (source grading, a key-assumptions / premortem / red-team pass, a competing-hypotheses
engine, and calibrated forecast scoring) both exist as tools, with the human kept on the judgment at every step.
Provided as is, for information only, with no warranty and no liability. Not legal or professional advice. Security and radio material is for authorized, lawful use only. Full disclaimer.
Visits are logged on our own server: IP address, browser and device details, and pages viewed. We use them to understand traffic and spot attacks, and never share or sell them.