Information Is Not Intelligence

Ask what an intelligence firm actually does and two answers arrive before the question has finished. One is theatrical: tradecraft, a source in a hotel bar, a file nobody else holds. The other is technological: a system that ingests everything and returns the answer. The two look opposed. They are the same mistake, which is to describe a method of acquisition and call the result intelligence.

This firm's own page on the subject is deliberately flat about it. "Intelligence gathering has nothing to do with artificial intelligence nor with science fiction. It is a systematic way of collecting data and information, analyzing it, and using it to your benefit." The sentence deflates both registers in one line, which is its purpose. It is also doing more work than it appears to, because the load-bearing part is not the collecting. It is the clause that follows.

Information is abundant and nearly worthless

Begin with the least flattering fact in this business, which the same page states without hedging: a significant portion of the data is disclosed to the public. Read plainly, that concedes that a determined person with enough time could assemble part of what an intelligence firm assembles. It is true, and it is worth saying before anything else, because everything that follows depends on accepting it.

Access stopped being the scarce commodity some time ago. In March 2024 the Office of the Director of National Intelligence and the Central Intelligence Agency issued a joint open-source strategy under the title The INT of First Resort, formalising publicly and commercially available information as a discipline in its own right rather than a supplement to the secret ones. The organisations with the least difficulty obtaining secrets are the ones putting public material first. The reason is not that public material improved. It is that the marginal value of one more secret has fallen against the marginal value of reading what was already sitting unread.

The demonstration is more than sixty years old. Roberta Wohlstetter's Pearl Harbor: Warning and Decision, published by Stanford University Press in 1962, examined why an attack preceded by a great deal of relevant reporting was nonetheless a surprise. Her finding was that the failure was one of interpretation rather than collection: the relevant signals existed, and were lost among a far larger body of irrelevant ones, with attention running preferentially towards material that confirmed what was already expected. The collectors had done their work. Relevance was never established.

That is the whole argument of this piece, and it can be stated once. Information is a commodity. Intelligence is a judgement about relevance. No volume of the first produces the second.

The three steps are a ranking, not a workflow

The method set out on the service page has three steps. Retrieve. Read between the lines, deleting irrelevant or unreliable data and looking elsewhere for confirmatory data. Then deduce priority and relevance. Written as a sequence it reads like a process diagram. It is better read as a ranking of scarcity, and the ranking runs opposite to the way this work is normally sold.

Retrieval is the commodity step. Knowing where to look and how to obtain it is a real skill, unevenly distributed, and it is the part that has depreciated fastest. It is also the part clients ask about first, because it is the only part they can picture.

Subtraction is the second step, and note the verb the page uses: delete. Most of what is collected on any subject is noise, duplication, or a single origin repeated until it resembles corroboration. Ten sources are one source with nine echoes until somebody traces provenance. The measure of good work at this stage is that the file gets smaller.

Relevance is the third step, and it cannot be performed on the material at all. Priority and relevance are not properties of information. They are properties of the relation between information and a decision somebody has to take. The same fact about a counterparty's banking arrangements is decisive for one principal and immaterial to another with different exposure, different obligations and a different tolerance for being wrong. This is why the engagement begins with requirements and scoping rather than collection. Work commissioned without a decision attached produces a document, and cannot produce intelligence, because the term of comparison is absent.

State practice standardises the same thing. Intelligence Community Directive 203, the analytic standards directive governing finished analysis across the United States intelligence community, requires products to distinguish underlying information from the analyst's assumptions and judgements, to express and explain uncertainty, and, most demandingly, to identify the indicators that, if observed, would alter the judgement. Not one of its nine tradecraft standards governs collection. They govern the part somebody signs.

What artificial intelligence changes, and what it does not

Something real has changed, and denying it is a posture rather than a position.

Machine systems have made the first step cheaper and wider: volume, languages, continuous monitoring, translation, deduplication, pattern surfacing across bodies of text no team could read in the time available. They have taken on part of the second step as well, the mechanical portion of subtraction. That is a genuine gain in the collection layer, and a firm claiming its analysts still out-read a machine is describing a preference, not a capability.

Three things do not move.

Relevance still requires a client. A model can rank material by general salience. It cannot rank by consequence for a particular principal, because the consequences are not in the corpus. They sit in exposure, obligations, intentions and tolerances that the client has often not articulated even to themselves. Eliciting that frame is the analytic act, and it happens in conversation before collection begins.

Confidence and evidence come apart. The National Institute of Standards and Technology's Generative Artificial Intelligence Profile, published in July 2024, lists confabulation among its named risks and defines it as the production of confidently stated but erroneous or false content by which users may be misled or deceived. The difficulty this creates for the discipline is specific rather than general. Intelligence is bought for the correspondence between how certain a statement sounds and what actually supports it. Human sources fail too, but they fail traceably: a source has a distance from the event, a motive, a history that can be weighed. Fluent synthetic output arrives with those traces stripped away and the register uniform throughout. The technology that most reduces the cost of producing an assessment raises the cost of auditing one.

The noise floor rises faster than the signal. Every capability described above is available to the other side, and to anyone with an interest in shaping what a subject's public record looks like. Corroboration by repetition was always the weakest form of confirmation. It is now close to worthless. Provenance and judgement become more expensive, not less.

The position, stated plainly: artificial intelligence is a collection and triage technology of real value. It has changed the first step of this work substantially, the second partially, and the third not at all. A firm selling it as an analyst is selling the least scarce part of the process at the price of the most scarce.

What the position costs

If the product is judgement rather than access, the firm cannot promise the thing clients most want promised.

We do not promise that the answer exists. Some questions cannot be resolved from any material that can lawfully be obtained, and the honest close of such a mandate is a short document saying so in the third week rather than the sixth month, before a second tranche of collection is bought to return the first tranche's answer at a higher price. We will not write certainty we do not hold in order to make a report feel worth its fee, and we will not let a judgement pass as a finding. Where a conclusion rests on an assumption, the assumption is named, and so is the thing that would break it.

That obligation is also the answer to the obvious objection, which is that a firm whose product is judgement has made its product conveniently unfalsifiable. It has, unless the judgement is written in a form that can fail. Assumptions stated. Indicators stated: the observations that would change the assessment, recorded before the event rather than explained after it. A report that cannot turn out to have been wrong is not an assessment. It is a description delivered in a confident register, and confident register is now the cheapest thing on the market.

The documents cited above are public statements of state practice and of technical risk. They describe standards rather than any engagement of this firm, and nothing here is a representation about any particular person, matter or jurisdiction.

The method itself sits under Intelligence Gathering, its open-source layer under OSINT, UCOL & Tailored Analysis Products, and the finished product under Intelligence Analysis & Privy Consul Office.