Chapter 4 — Reading the Pile

Between April and June 2020, French and Dutch authorities read more than 115 million conversations belonging to over 60,000 users of an encrypted phone network called EncroChat. The downstream result: 6,558 arrests worldwide, around 900 million euros in assets seized, more than a hundred tonnes of cocaine, 160 tonnes of cannabis, over 900 weapons. [ESTABLISHED] (Eurojust; Computer Weekly)

The following March, the same pattern with Sky ECC: over 160,000 accounts monitored, more than a billion messages accessed, around 10,000 arrests, 121 tonnes of drugs. [ESTABLISHED] (Eurojust annual report 2021)

Now the correction that must travel with those numbers, because it is the most commonly abused fact in this entire field: the breakthrough was not artificial intelligence. It was a technical intrusion into a network — a hacking operation. No model broke the encryption.

What machines did was everything after. A billion messages cannot be read by people. Triage, entity extraction, keyword and link analysis at that scale is automated or it does not happen. So these operations are legitimately evidence that automated processing changes what is investigable — and they are not evidence that AI catches criminals. Anyone who cites EncroChat as an AI success is either confused or selling something.

This chapter is about drawing that line precisely: what a machine actually does to a stack of case material, and what it categorically does not.

The claim you will hear, and its grade

Bedfordshire Police seized around two dozen smartphones in a fraud investigation — 1.4 terabytes of data. An AI system reportedly read and translated more than 100,000 messages in a single day, work said to have otherwise taken months or years. Six men were convicted over a scheme that took £800,000 through more than 3,000 cashpoint withdrawals. The system is named Nectar and it is supplied by Palantir. [CLAIMED] (IBTimes UK)

The convictions are real. The volume claim and the counterfactual timeline are police self-report via press, and I found no independent verification of either. Nobody timed the alternative.

The same platform, examined by journalists, integrates around 80 data streams and touches 11 categories of special-category personal data — race, sexual orientation, political opinion, health, sex life, trade union membership. Bedfordshire Police and Palantir respond that it aggregates data already lawfully held and adds no predictive layer. Senior MPs and Amnesty International describe it as indiscriminate mass surveillance with inadequate oversight. [CONTESTED] (Liberty Investigates)

Both things are true about the same product in the same year. That is normal in this field and the book will not resolve it for you. What it will do is insist you notice which grade each sentence carries.

What the machine actually does

Strip away the marketing and a language model applied to investigative material does five things well. All five are unglamorous. All five are real.

It translates. Case files, chat logs, and scam scripts in a language the investigator does not read become readable. For Thai material specifically this is not a minor convenience — it is the difference between a corpus being searchable and being a wall.

It extracts. Given ten thousand free-text complaint narratives, it pulls the structured entities buried in prose: account numbers, phone numbers, LINE identifiers, wallet addresses, company names, the URL the victim was sent. These exist in the record already. They are simply not in fields, which is why no database query can reach them. This is the single highest-value task in the entire chapter, and everything that follows depends on it.

It normalises. The same bank account written six ways, the same phone number with and without country code, a company name transliterated differently by four officers. Making these agree is tedious, mechanical, and exactly what machines are for.

It clusters. Given normalised entities plus the language of the narrative itself — the phrasing scammers reuse, the story structure, the promised return — it groups reports that resemble each other. Not "solves." Groups.

It summarises. Two hundred complaints become a page a human can hold in mind. This is where hallucination risk is highest and where the chapter's warning lives.

What it does not do

It does not decide. A cluster is a hypothesis about which files belong together. Treating such an output as a conclusion is the single failure mode behind every documented disaster in this field.

It does not testify. No model's assertion is evidence. The evidence is the underlying document the model pointed you at, obtained and preserved under a documented chain of custody. A summary is a finding aid, not an exhibit.

It does not know what it does not know. A model asked to summarise two hundred complaints will produce a summary of two hundred complaints. It will not tell you that the three hundred most relevant complaints were never filed because those victims did not report.

And it will invent. In November 2025, West Midlands Police used Microsoft Copilot to compile an intelligence assessment before a football match. The model fabricated a prior fixture between Maccabi Tel Aviv and West Ham that had never taken place, and the invented incident went into the document used to justify banning Maccabi Tel Aviv supporters from attending. Police initially denied the AI was at fault, then conceded it. The force withdrew Copilot access; its chief constable resigned in January 2026. [ESTABLISHED] (The Register, 19 Jan 2026; The Register, 28 Jan 2026)

Reporting at the time indicated that only eight UK forces had confirmed Copilot may not be used in investigative or intelligence work; at least 21 continued using it. [ESTABLISHED] (ResultSense)

Read that last pair of numbers as the real state of the art. A general-purpose assistant invented a fact that changed a policing decision, the failure was public and career-ending, and most forces carried on.

The rule that follows

There is a discipline that makes all five useful capabilities safe, and it is one sentence:

The machine may only produce pointers, never assertions.

Every output must resolve to a document you can open. A cluster is a list of case numbers. A summary is a set of quotations with file references. An extracted account number is a span of text in a specific complaint. If an output cannot be traced back to a source document, it does not leave the machine.

This is not caution for its own sake. It is what makes the work checkable by someone who does not trust you, which is the only kind of work that survives a courtroom or a hostile review. It also happens to eliminate the hallucination failure mode structurally rather than probabilistically: a fabricated fixture has no source document, so it never passes the gate.

Note that the West Midlands failure would have been caught by this rule at zero cost. The invented match had no citation because it had no source. Nobody looked.

Where the honest evaluations are

If you want to know whether models genuinely help investigators, the state of evidence in August 2026 is thin and I will not dress it up.

The UK College of Policing is running a formal evaluation of an AI assistant called CoPA, built on Azure OpenAI and restricted to College of Policing, CPS and gov.uk sources, piloted at Greater Manchester Police and British Transport Police. The project started July 2025 and was due to complete March 2026. No results were published as of this writing. [ABSENT] (College of Policing)

So the position is: a controlled evaluation exists and has not reported; meanwhile an unevaluated general-purpose assistant is in use at more than twenty forces and has already caused one documented harm. That is the actual landscape. Anyone describing it as settled in either direction is not reading it.

The Move — Demand a citation for every machine-produced sentence. Configure whatever tooling you build so that no output is emitted without a pointer to the source document and the exact span it came from, and refuse to act on any sentence that lacks one. Apply it to your own summaries too. The rule is trivial to implement and it converts the dominant failure mode of language models from a probability into an impossibility.

What would change my mind: if the CoPA evaluation, or one like it, showed that citation-gated assistance produced no measurable improvement in investigator accuracy or speed against a control, then the five capabilities above are a story I find plausible rather than a benefit anyone has demonstrated. As of August 2026 that evaluation has not reported, and I am explicitly writing this chapter on a promise rather than a result.