Loading module…
Loading module…
Module 1 of 6
Open-source intelligence — OSINT — is the discipline of turning publicly available information into answers you can act on. It is what a journalist does verifying a story, what a due-diligence analyst does before an acquisition, what a security team does profiling a threat, what a fraud investigator does tracing a network. None of it requires secret sources; all of it requires method. And into that method, AI has arrived as the most significant new tool in a generation — able to read faster than any human, surface patterns across volumes no analyst could hold, and draft in minutes what once took days. This course is about using that tool well: extracting its genuine power while defending against the specific ways it corrupts intelligence work, which are exactly the ways the previous AI courses warned about, now with higher stakes because intelligence drives decisions.
This first module establishes the frame that governs everything after it: the intelligence cycle, the disciplined process that separates intelligence from mere information-gathering, and where AI fits within it. The frame matters because the single most common failure in OSINT — with or without AI — is skipping the discipline: collecting without a question, analyzing without evaluating sources, concluding without weighing alternatives. AI makes this failure both easier to commit (it generates so fluently that skipping the discipline feels productive) and more dangerous (its confident fabrications enter your product looking like findings). So we begin not with tools but with the process into which the tools must fit, because a powerful tool applied to a broken process produces powerful errors faster.
A note on boundaries, stated once and meant throughout: this course teaches OSINT as a lawful, ethical discipline working from genuinely public information for legitimate purposes — research, security, journalism, due diligence, investigation. It does not teach surveillance of private individuals for harassment, stalking, or intimidation; it does not teach unauthorized access to anything non-public; and it treats the privacy and dignity of the people who appear in open sources as a real constraint, not an afterthought. The methods are powerful, and power in this domain carries responsibility. That responsibility runs through every module, and it is part of the tradecraft, not a disclaimer bolted to it.
The intelligence cycle is the structured process by which raw information becomes intelligence — a finished answer to a specific question, evaluated for reliability and delivered to someone who will act on it. It exists because the alternative — gathering information ad hoc and drawing conclusions as you go — reliably produces bad intelligence, and understanding why it fails is what motivates the discipline.
The classic cycle has five stages, and while different organizations name them slightly differently, the substance is consistent. It begins with direction (also called planning or requirements): deciding what question you are actually trying to answer, for whom, and to what standard. Then collection: gathering the raw information relevant to that question from available sources. Then processing: converting the raw collected material into a usable form — translating, transcribing, organizing, deduplicating, structuring. Then analysis: the core intellectual work of evaluating the processed information, weighing sources, identifying patterns, testing hypotheses, and reaching conclusions. Finally dissemination: delivering the finished intelligence to the person who needs it, in a form they can use, with its reliability honestly characterized. And running through all of it, a feedback loop: the answer raises new questions, the consumer's reaction refines the requirements, and the cycle turns again.
Why call it a cycle rather than a checklist? Because intelligence work is iterative: collection reveals that the question was wrong and needs refining; analysis reveals a gap that sends you back to collect more; dissemination produces feedback that redirects everything. The cycle captures this circularity — you rarely march through the stages once in a straight line; you loop, refining the question as the evidence teaches you what you should have been asking. This iterative character is exactly why the process matters and why skipping it fails: an ad hoc approach has no mechanism for the question to improve, so it answers whatever question you started with, which is usually not the question that matters.
The discipline's deepest purpose is to force the separation of distinct cognitive activities that, when blurred, corrupt each other — and this is the insight that makes the cycle worth internalizing rather than memorizing. Collection and analysis are different activities: collection should gather what is there, while analysis should evaluate it, and blending them causes you to collect only what fits your emerging conclusion (confirmation bias operating at the collection stage). Evaluation and conclusion are different: you must assess how reliable a source is separately from whether you like what it says, or your conclusions become a mirror of your preferences. Direction is different from everything: deciding what question matters must precede collection, or you drown in information relevant to no decision. The cycle enforces these separations structurally, making you do each activity as its own step, which is precisely what prevents the failures that unstructured information-gathering falls into. When you later see AI tempting you to collapse these stages — to collect, analyze, and conclude in a single fluent generation — the cycle is what tells you to resist, because the separations it enforces are exactly what the fluent single-pass generation destroys.
Key Insight: The intelligence cycle — direction, collection, processing, analysis, dissemination, with a feedback loop — exists to force the separation of activities that corrupt each other when blurred: collecting versus analyzing, evaluating a source versus liking its content, deciding the question versus gathering information. It is a cycle, not a checklist, because the question improves as the evidence teaches you what to ask. Skipping the discipline is the most common OSINT failure, and AI makes it easier to skip and more dangerous when you do.
Every piece of intelligence work begins — or should begin — with direction: the disciplined definition of what you are actually trying to find out. This stage is the most skipped and the most consequential, because everything downstream inherits its quality: a vague question produces unfocused collection, unfocused analysis, and a product that answers nothing anyone needed. Getting direction right is the highest-leverage act in the entire cycle, and it is almost entirely a human responsibility that AI cannot perform for you.
A proper intelligence requirement has several components that vague requests lack. It has a specific question — not "tell me about this company" but "is this company financially able to complete the contract it is bidding on, and what are the signs it might not be?" The specific question focuses collection (you now know what to look for) and defines success (you know when you have answered it). It has a consumer — who needs this and what decision will it inform? — because the consumer determines the depth, the format, and what actually matters: intelligence for a go/no-go decision needs different content than intelligence for background understanding. It has a standard — how confident must the answer be, given the stakes? — because a low-stakes curiosity warrants light work while a high-stakes decision demands rigorous collection and verification. And it has scope — the boundaries of what is in and out, in time, geography, and subject — because unbounded questions produce unbounded, unfinishable collection.
The discipline of direction is to convert a fuzzy initial request into a sharp requirement before collecting anything, and this conversion is genuine analytical work. Someone asks you to "look into" a person, a company, an event — and the fuzzy request could mean a dozen different things. The direction stage is where you pin it down: what specifically do they need to know, what will they do with it, how sure do they need to be, and where are the boundaries? This pinning-down often reveals that the real question is different from the stated one, or that the request bundles several distinct questions that need separating, or that the stakes are higher or lower than the phrasing suggested. Doing this work upfront is what makes the rest of the cycle efficient and focused; skipping it — diving straight into collection on a fuzzy request — is what produces the sprawling, unfocused OSINT effort that gathers mountains of information and answers nothing.
Here AI's role is real but bounded, and the boundary is instructive. AI can help you sharpen a question — you can describe a fuzzy request and have it suggest the specific sub-questions a thorough investigation would pursue, the dimensions you might be neglecting, the ways the question could be scoped. This is genuinely useful, exactly the "help you think" use from the AI courses, and it can improve your direction. But AI cannot decide your question, because the question flows from a decision someone will make and stakes someone will bear, which are human matters the AI does not hold. The requirement — what actually matters, to whom, at what standard — is yours to set, and it is the part of intelligence work where human judgment is most irreplaceable, because it is judgment about human purposes. The person who lets AI define the question has not saved themselves the hardest work; they have outsourced the one decision that determines whether the entire subsequent effort is aimed at anything worth answering. Use AI to help sharpen the question; never let it set the question, because a brilliantly-executed investigation of the wrong question is the most expensive failure in intelligence work, and direction is where that failure is prevented or committed.
Key Insight: Direction — defining the specific question, the consumer, the confidence standard, and the scope — is the highest-leverage stage, because everything downstream inherits its quality, and it is where the real question often turns out to differ from the stated one. AI can help sharpen a fuzzy question into specific sub-questions, but it cannot set the question, which flows from human decisions and stakes. A brilliantly-executed investigation of the wrong question is intelligence work's most expensive failure, and direction is where it is prevented.
Collection and processing — gathering the raw information and converting it into usable form — are where AI delivers its most dramatic and genuine gains in OSINT, because these stages involve exactly the high-volume, mechanical work that AI accelerates and that the AI courses identified as its reliable strengths. Understanding what AI genuinely transforms here, and what it does not, lets you exploit the gains without inheriting the risks.
In collection, AI's transformation is one of scale and speed. A human analyst can read only so fast and hold only so much; AI can process volumes of text — documents, articles, posts, records — far beyond human capacity, and surface what is relevant to your question from a haystack no analyst could search manually. This matters enormously for OSINT, where the challenge is often not the absence of information but its overwhelming abundance: the relevant facts are public, buried in volumes of public material, and the bottleneck is finding them. AI-assisted collection — using tools that search, retrieve, and filter across large volumes, then summarize and extract what bears on your question — collapses this bottleneck, letting you cover ground that manual collection never could. This is a real and large gain, and it is grounded (per the AI courses) when the AI works from actual retrieved sources you can verify rather than generating from its trained memory, which is the crucial distinction to maintain: collection means gathering real sources, so use AI to find and retrieve real material, never to "recall" facts from its training, which is not collection but fabrication wearing collection's clothes.
In processing — converting raw collected material into usable form — AI's gains are equally real and lower-risk, because processing is largely about form. AI excels at the mechanical transformations processing requires: summarizing long documents so you can triage what deserves full reading, translating foreign-language sources, transcribing audio and video, extracting structured data from unstructured text, organizing scattered material into coherent form, and deduplicating overlapping sources. Each of these is a form-transformation of material you provide, grounded and verifiable, and each collapses hours of tedious manual work. The processing stage, historically a major time-sink in OSINT — the slog of reading, translating, transcribing, and organizing before real analysis can begin — is where AI's mechanical strength shines with genuine reliability, because the work is about converting real material's form, not generating claims, so the hallucination risk is contained.
But the transformation, however real, must not blur the stages the cycle separates — and this is the discipline that keeps AI-accelerated collection and processing sound. The temptation, precisely because AI is so fluent, is to let collection slide into analysis: to have the AI not just gather and summarize but also interpret, conclude, and tell you what it means, collapsing collection, processing, and analysis into one generation. Resist this. Collection should gather what is there; processing should make it usable; and the evaluation and interpretation — the analysis — must remain a separate, deliberate stage (the next module and the cycle's later stages), because collapsing them reintroduces exactly the confirmation-bias and unevaluated-conclusion failures the cycle exists to prevent, now at AI speed and scale. Use AI to collect (find real sources) and process (transform their form) at a scale and speed no human could match — this is the genuine transformation — while keeping the collected, processed material as input to analysis you still perform, not as pre-digested conclusions to accept. The gain is that you reach the analysis stage with vastly more relevant, well-organized material than manual work could provide; the discipline is that analysis remains yours, applied to that material, rather than something the AI already did for you in the same fluent pass that gathered it.
Key Insight: AI transforms collection (scale and speed — searching and retrieving relevant material from volumes no analyst could cover manually, when grounded in real retrieved sources) and processing (the mechanical form-work — summarizing, translating, transcribing, extracting, organizing — reliably, because it's about converting real material's form, not generating claims). The discipline: keep these stages separate from analysis. Use AI to reach the analysis stage with more and better-organized material, never to collapse collection, processing, and interpretation into one fluent pass that hands you conclusions to accept.
Having praised AI's genuine gains, we must map precisely where it is dangerous in intelligence work, because the dangers are severe and specific, and intelligence is a domain where AI's failures do maximum harm: a fabricated fact in an intelligence product becomes a false belief that drives a real decision. The dangers cluster at the analysis and dissemination stages — exactly where AI is weakest and where its errors are most consequential.
The gravest danger is fabrication entering the product as fact, the hallucination problem from the AI courses, now in a context where the fabrication corrupts a decision. When AI generates from its trained knowledge rather than from verified sources, it produces plausible-sounding specifics — names, dates, connections, figures, quotes — that may be entirely invented, delivered with the same fluent confidence as genuine findings. In intelligence work, where the product asserts facts that someone will act on, an invented fact is not a minor error but a poisoning of the decision at its source. The citation-fabrication disaster from the AI courses is especially acute here: an AI asked to support a finding may generate plausible-looking sources that do not exist, and an analyst who does not verify them has built an intelligence product on invented evidence. The defense is the verification discipline the AI courses drilled, now non-negotiable: every specific fact, every source, every connection asserted in an intelligence product must trace to real, verified, open-source material, and AI's trained-knowledge output is a lead to verify, never a finding to assert.
A second danger is AI mistaking correlation and coincidence for meaningful connection, the causal-reasoning weakness from the AI courses, which in OSINT manifests as false links. Intelligence analysis often involves inferring connections — is this person linked to that organization, did this event cause that outcome, are these accounts operated by the same actor? — and this is exactly the causal-and-inferential reasoning where AI is weakest, prone to asserting connections that are coincidental, spurious, or based on superficial pattern-matching rather than genuine evidence. An AI that "finds" a connection between two entities may be pattern-matching on incidental similarities, and an analyst who accepts the connection without scrutinizing the actual evidence for it can build a false picture — linking people or entities that are not actually linked, which in intelligence work can do serious harm to the innocent and mislead the decision. The defense is to treat every connection AI surfaces as a hypothesis to be evidenced, scrutinizing the actual basis for the link rather than accepting the AI's assertion of it.
A third danger is the false authority of AI-generated intelligence products. A well-formatted, fluent, confident intelligence product carries authority, and AI produces exactly such products effortlessly — which means AI can generate intelligence that looks authoritative and rigorous while resting on unverified or fabricated content, and the polish makes the lack of rigor harder to detect. This is the intelligence-work form of the AI courses' warning that sophistication increases the danger of misplaced trust: a crude, obviously-rough product invites scrutiny, while a polished AI-generated product invites acceptance, so the very quality of AI's output can smuggle unverified content past the scrutiny it needs. A fourth danger is over-reliance eroding the analyst's own judgment: an analyst who lets AI do the analysis ends up with conclusions they did not reason through and cannot defend, a borrowed understanding that fails under questioning and that leaves the analyst unable to catch the AI's errors because they never engaged with the evidence themselves. Across all four dangers, the pattern is the AI courses' pattern: AI is dangerous exactly where intelligence work requires genuine reasoning, evaluation, and judgment (the analysis stage), and safe where it requires mechanical scale (collection and processing) — so the map of danger is the map of the cycle, with the risk concentrated precisely at the stage where the intellectual work of intelligence actually happens.
Key Insight: AI's dangers concentrate at analysis and dissemination — where intelligence work needs genuine reasoning and where errors corrupt decisions: fabrication entering the product as fact (invented names, dates, connections, and non-existent sources, delivered with fluent confidence), false connections from AI's causal-reasoning weakness (coincidence mistaken for meaningful links, which harms the innocent), the false authority of polished AI products (sophistication smuggling unverified content past scrutiny), and over-reliance eroding the analyst's judgment. The map of danger is the map of the cycle: risky exactly where the real intellectual work lives.
Given AI's transformation of collection and processing and its danger at analysis, the correct division of labor becomes clear, and it defines the role of the intelligence professional in the AI age: AI handles the mechanical scale, the analyst does the analysis, and the analyst's judgment governs the whole. Understanding this division — and defending it against the pressure to let AI do more — is what keeps AI-assisted OSINT a discipline rather than an automated generator of confident error.
The analyst's irreducible contributions map onto exactly the stages where AI is weak or absent. Setting the question (direction, Section 2) is the analyst's, because it flows from human decisions and stakes. Evaluating sources is the analyst's, because judging the reliability, bias, and provenance of a source requires understanding the source's nature and motives — an AI can note what a source says but cannot reliably judge whether to trust it, which is a judgment about the source's credibility that draws on knowledge and reasoning the AI lacks. Weighing evidence and reaching conclusions is the analyst's, because this is the genuine reasoning — considering alternatives, weighing conflicting evidence, distinguishing what the evidence supports from what it merely suggests — that AI's plausibility-generation cannot reliably perform. Verifying is the analyst's, because AI provides no honest signal of when it is fabricating, so confirming that findings rest on real evidence is a human act. And owning the product is the analyst's, because someone must stand behind the intelligence, vouching for its reliability to the consumer who will act on it, and that accountability cannot be delegated to a tool.
AI's contributions, correspondingly, are the mechanical amplifications of the analyst's reach. It extends collection across volumes no human could cover. It accelerates processing — the summarizing, translating, transcribing, extracting, organizing. It can assist analysis as a thinking aid (surfacing considerations, generating alternative interpretations to evaluate, stress-testing the analyst's conclusions), exactly the "help you think better" use from the AI courses, provided the analyst does the actual reasoning and evaluates the AI's contributions rather than accepting them. And it can assist dissemination by drafting the product, which the analyst then verifies, corrects, and owns. In every case AI amplifies what the analyst can do — more sources covered, faster processing, more angles considered, faster drafting — while the analyst supplies the judgment, evaluation, verification, and accountability that AI cannot. This is the augmentation relationship from the AI courses, in intelligence form: the analyst is amplified, not replaced, and remains the analyst.
The pressure to violate this division is real and worth naming, because it is where AI-assisted OSINT goes wrong. AI is so capable at collection and processing, and so fluent at generating analysis-shaped and product-shaped output, that it is tempting to let it do the analysis and write the product with the analyst merely reviewing — or barely reviewing. This is the path to automated confident error: the analyst who lets AI analyze and concludes by rubber-stamping the AI's output has abdicated the judgment that intelligence work exists to provide, and produces products that look rigorous while resting on unverified AI generation. The discipline is to hold the division firmly: let AI amplify the mechanical stages fully, use it as a genuine aid to analysis, but keep the evaluation, the reasoning, the verification, and the accountability human — because the value of intelligence is precisely the trustworthy judgment it embodies, and a product where the judgment was AI's and unverified is not intelligence but the appearance of intelligence, which is worse than none because it is trusted. The analyst in the AI age is more powerful than ever, amplified across the mechanical work, freed to focus their judgment where it matters — but they remain the analyst, and the judgment remains theirs, or the product is not intelligence at all.
Key Insight: The division of labor defines the analyst's role in the AI age: AI amplifies the mechanical stages (collection across vast volumes, processing, drafting) and assists analysis as a thinking aid, while the analyst keeps the irreducible work — setting the question, evaluating sources, weighing evidence, reaching conclusions, verifying, and owning the product. The pressure to let AI analyze and rubber-stamp its output is the path to automated confident error. The analyst is amplified, not replaced, and the judgment stays human, or the product is only the appearance of intelligence — worse than none, because it is trusted.
We close the foundational module with the constraint that runs through all of OSINT and that AI makes more pressing, not less: the ethical and legal boundaries of intelligence work. These are not a compliance afterthought appended to the tradecraft; they are part of the tradecraft, because intelligence work operates on information about real people and organizations, and how you gather and use that information determines whether you are doing legitimate research or causing real harm.
The foundational boundary is the line between open and lawful collection and everything else. OSINT works from genuinely public, lawfully accessible information — what is published, posted publicly, filed in public records, available to anyone. It does not involve unauthorized access to non-public systems or data, deception to obtain access, or any method that crosses from observing what is public into intruding on what is not. This line is both legal (crossing it can be a crime) and definitional (crossing it is no longer open-source intelligence but something else). AI does not change this line, but it can blur it in practice by making large-scale collection so easy that the volume obscures the question of whether each source was lawfully and appropriately obtained — so the discipline is to keep the line clear regardless of the tool's ease: gather only what is genuinely public and lawfully accessible, and let no capability tempt you across the boundary into unauthorized access, because the ease of the tool is not a justification and the law does not exempt efficient methods.
A second boundary concerns privacy and proportionality even within public information. That information is technically public does not mean any use of it is ethical: aggregating scattered public details about a private individual into a comprehensive profile can constitute a harm even when each detail was public, because the aggregation creates something — a detailed dossier on a private person — that the scattered public facts did not. Intelligence work, especially AI-amplified intelligence work capable of aggregating at scale, must weigh proportionality: is this collection proportionate to a legitimate purpose, or is it intrusion dressed as research? The purpose matters enormously — investigating a public company's finances for a legitimate business decision is different from compiling a private individual's life for no legitimate need — and the discipline is to hold a legitimate purpose for the intelligence you gather and to keep the collection proportionate to that purpose, rather than gathering everything possible because AI makes it possible. The capability to profile anyone in detail is exactly why the restraint matters: the tool removes the practical limits that once constrained intrusion, so the ethical limits must be held deliberately.
A third boundary concerns use and harm: intelligence, once gathered, can be used to help or to harm, and the analyst bears responsibility for the use they enable. OSINT methods can serve legitimate ends — security, journalism, due diligence, fraud investigation, legitimate research — and can serve illegitimate ones — stalking, harassment, doxxing, intimidation. The methods are the same; the purpose and use differ, and that difference is where the ethics live. The discipline is to apply these methods only to legitimate ends and to refuse the illegitimate ones, recognizing that the power to gather and analyze information about people carries the responsibility not to use it to harm them. This connects to the deepest point of the module, which ties intelligence ethics to the AI courses' theme: powerful tools have no ethics of their own, and the entire ethical weight falls on the human wielding them. AI amplifies OSINT's power enormously — more collection, faster analysis, aggregation at scale — and in doing so amplifies its potential for both legitimate value and real harm, placing more ethical responsibility on the analyst, not less. The professional intelligence analyst holds these boundaries — lawful and open collection, proportionate and purposeful use, refusal of harm — not as external rules but as part of what it means to do the work well, because intelligence work done without them is not skilled tradecraft but the efficient production of harm, and the skill of the tradecraft includes the judgment of where not to point it. That judgment, like all the judgment in this module, is irreducibly human, and it governs everything the powerful tools of the following modules will teach.
Key Insight: Ethics and legality are part of OSINT tradecraft, not an afterthought, and AI makes them more pressing by removing the practical limits that once constrained intrusion. Three boundaries: open and lawful collection only (never unauthorized access, no matter how easy the tool makes it); privacy and proportionality even within public information (aggregating scattered public details into a dossier on a private person can be a harm; hold a legitimate purpose and keep collection proportionate); and responsibility for use (the same methods serve security or stalking — apply them only to legitimate ends). AI amplifies the power and thus the responsibility; the entire ethical weight falls on the human analyst.
Concept without execution is half a course. This section turns the intelligence cycle into concrete tools and moves you can run today. Everything here works from lawful, public sources; the ethical and legal boundaries of Section 6 govern all of it.
Organize your tools by the stage they serve, so you reach for the right one at the right moment rather than collecting a random pile of "OSINT tools."
whois, dig/host (DNS), crt.sh (certificate-transparency logs — reveals subdomains via issued TLS certificates), dnsdumpster, SecurityTrails, and Shodan / Censys for internet-exposed hosts and services. (Only ever observe what is publicly indexed — never connect to or probe systems you are not authorized to.)web.archive.org) and archive.today — indispensable for content that has been changed or deleted; a deleted page often still lives in an archive.(Tool names current as of authoring; the durable skill is knowing which category of tool each cycle stage needs — the specific product for that category you confirm is current when you use it.)
The single highest-leverage collection skill is precise querying. Plain keyword search is a blunt instrument; operators turn a search engine into a scalpel. The core set (Google syntax; most engines have equivalents):
site: — restrict to one domain. site:linkedin.com/in "supply chain director" "Acme"filetype: (or ext:) — restrict to a file type. filetype:pdf "annual report" "Acme Corp"intitle: / inurl: — term appears in the page title / URL. intitle:"index of" "backup" surfaces open directory listings."exact phrase" — force an exact match. "John Q. Smith" "Acme Corp"-term — exclude. Acme -jobs -careers strips recruiting-site noise.OR (or |) and ( ) — alternatives and grouping. ("chief financial officer" OR CFO) "Acme"before: / after: — date-bound. "Acme merger" after:2024-01-01 before:2024-06-30A real collection move combines them: to find documents a company published but may not link prominently, site:acme.com filetype:pdf then scan the results; to find an executive's public footprint, run the exact-name phrase across site: a few platforms in turn. Certificate transparency is its own high-value trick: visiting crt.sh/?q=%25.acme.com lists TLS certificates issued for acme.com subdomains, mapping an organization's public web surface in seconds — all from public logs.
Fuzzy request: "Look into Acme Corp before we sign." Direction (sharpen it): Question — "Is Acme financially and operationally able to deliver the 18-month contract, and what public signals suggest risk?" Consumer — the deal owner, for a sign/no-sign decision. Standard — high; this commits budget. Scope — Acme and its named principals; last 3 years; public sources only. Collection (run specific queries):
site:sec.gov "Acme""Acme Corp" (layoffs OR restructuring OR lawsuit OR default) after:2023-01-01crt.sh/?q=%25.acme.com for subdomains; Wayback Machine on acme.com to see how their public claims changed over time.site:linkedin.com/in, news, and court-record portals.
Processing (AI-accelerated): feed the collected filings and articles to the model with: "Summarize what these documents say about Acme's financial health and contract-delivery capacity. Quote the exact source line for each point. Flag anything you cannot ground in the provided text."
Hold at analysis: you now have organized, sourced material — but the judgment (is this risk real? do the sources check out? what is the verdict?) is the next stage, done by you, not the model.The safest, highest-value AI move in OSINT is grounded extraction — the model works only from sources you paste in. A reusable template:
"Below are [N] sources I collected about [subject], each marked with a number. Extract every factual claim relevant to [the question]. For each, give: the claim, the source number, and the exact quoted sentence it comes from. If a claim appears in multiple sources, note which. Do not add any fact that is not in the sources. List separately anything the sources imply but do not state outright. Sources: [paste]."
This forces citation to real text, exposes the model's reach so you can check it, and separates stated fact from inference — turning the model into a fast, verifiable extractor rather than an unreliable oracle. Then you verify each quoted line against the source, because (per Module coming and the AI courses) a cited line is only verification once you have actually confirmed it.
Key Insight: Execution turns the cycle into moves you can run: a toolkit organized by stage (search, infrastructure, people, historical, imagery, AI processing, aggregation), search operators (
site:,filetype:,intitle:, exact-phrase, exclusion, date bounds, and certificate-transparency via crt.sh) that make querying a scalpel, and a grounded-extraction prompt that forces the model to cite real source lines. Run the full cycle on one target — sharpen the question, fire specific queries, AI-process the haul, then hold at analysis — and OSINT stops being "search around" and becomes disciplined, repeatable tradecraft.
You now have the frame that governs AI-powered OSINT: the intelligence cycle, and where AI fits within it. You understand the cycle as a discipline that separates the activities — collecting, evaluating, concluding — that corrupt each other when blurred, and why AI both tempts you to blur them and makes blurring them more dangerous. You understand direction as the highest-leverage stage, the human-set question that everything inherits. You know where AI genuinely transforms the cycle (collection and processing — the mechanical scale) and where it is dangerous (analysis and dissemination — where reasoning lives and errors corrupt decisions). You understand the division of labor that keeps the analyst the analyst: AI amplifies the mechanical work, the human keeps the judgment, evaluation, verification, and accountability. And you hold the ethical and legal boundaries as part of the tradecraft, made more pressing by the power AI adds.
The next module, Automated Research & Information Gathering, goes deep on the collection and processing stages where AI's gains are largest — the concrete methods of using AI to search, retrieve, filter, and organize open-source information at scale, and the discipline that keeps AI-accelerated collection grounded in real, verifiable sources rather than sliding into fabrication. Everything builds on this module's frame: the methods serve the cycle, collection stays separate from analysis, findings trace to real sources, and the ethical boundaries hold. You will learn to gather at a scale no manual effort could match, while keeping the gathering disciplined, lawful, and grounded — the foundation on which the analytical modules that follow will build the judgment that turns gathered information into genuine intelligence.
osintframework.com) is a living directory when you need a tool for a specific data type.