A Reed Group white paper. Companion to Signals in the Noise.

Algorithms now decide what most leaders see before any editor does. That makes the modern feed a collection environment with built-in bias, and it should be handled the way intelligence analysts handle any source of uncertain reliability.

This paper makes three claims:

The framework is tested against AI risk coverage, a domain where volume is high, stakes are real, and many voices carry undisclosed commercial interest.

The audience is fragmenting, but not in the clean way either camp claims. Cable news is shrinking across the board, while creator news reach has plateaued rather than surged.

The last row matters most. Most people do not choose their creator sources. The feed chooses for them, ranked by predicted engagement. That is the core analytic problem: the selection mechanism is invisible and optimized for attention, not accuracy.

Counterpoint worth holding: editorial selection was never neutral either. Legacy outlets also chase ratings. The shift is from a visible, accountable gatekeeper to an invisible, personalized one.

Grade the source and the claim separately. Analysts have done this for decades with the Admiralty system: a letter for source reliability (A to F) and a number for information credibility (1 to 6). A reliable source can carry a weak claim, and an unknown source can carry a strong one.

Step 1. Grade the source (A to F)

Step 2. Grade the claim (1 to 6)

Step 3. Apply three incentive checks

An item graded F6 is not false. It is unusable for decisions until it moves. Most of the feed lives in C3 to F6, and that is fine as long as leaders treat it as a lead, not a finding.

AI risk is the right test domain because all three source types compete on it at once, and most carry a stake in the outcome.

Worked example: the "rogue AI drone" (June 2023)

A USAF colonel told a London conference that an AI drone in a simulation "killed" its operator to complete its mission. Within a day it ran as fact across Vice, Insider, Fox News and creator accounts. The Air Force then denied running any such test, and the colonel said he misspoke and was describing a hypothetical thought experiment. Vice · Channel 4 FactCheck

Run it through the framework:

The lesson for leaders: a high source grade does not lift a low claim grade. B3 is still a lead.

Counterexample: when the independent got there first (MH17)

Independents are not the problem by default. In May 2016, the open-source collective Bellingcat matched seven unique features in public photos and traced the Buk launcher that downed MH17 to Russia's 53rd Anti-Aircraft Missile Brigade. RFE/RL The official Dutch-led Joint Investigation Team reached the same conclusion in May 2018, two years later, and much of the evidence it cited had already appeared in Bellingcat reports. RFE/RL

Why it graded well despite a new, non-institutional source:

The lesson: grade the method, not the masthead. A transparent F-grade source can earn its way to B. An opaque A-grade source can coast on reputation.

AI-specific counterexample: prompt injection

Independent developer Simon Willison popularized the term "prompt injection" on his blog in September 2022 and spent the following months documenting the flaw publicly. Wikipedia Mainstream recognition came later: Ars Technica covered the Bing "Sydney" leak in February 2023, and the UK National Cyber Security Centre warned in August 2023 that the problem may be inherent to LLMs. OWASP now ranks it LLM01, the top LLM risk. A solo blogger with no institution behind him graded C at the start. A transparent method, reproducible demos and a two-year track record moved him to B.

The pattern across all three: the most shared framing sits furthest from the evidence. The measured finding is narrower and duller every time.

Neither case proves a hidden payment. That is the point. A reader cannot tell, and under the framework, an unknown financial relationship fails the money check until it is disclosed. The FTC Endorsement Guides require clear disclosure of any material connection, including free travel and products.

Partly, yes. Security and threat-intelligence vendors publish research as marketing, and their content teams tune it with the same engagement metrics creators use. Alarming reports earn coverage, coverage drives inbound demand, and demand funds the next alarming report.

The steelman: vendors see telemetry no journalist or regulator can access. Some of the most important early warnings of the last decade came from vendor research. Dismissing it wholesale is its own bias.

The test that separates the two:

Worked example: the $10.5 trillion cybercrime figure

The most quoted number in security marketing claims cybercrime would cost the world $10.5 trillion a year by 2025. Cybersecurity Ventures reported it in a report sponsored by security vendor INTRUSION, and it compounds from a 2015 base. Critics quoted by SecurityWeek found no independent evidence supporting figures at that scale.

Grade: C4. The number may still point to a real, growing problem. It should never anchor a board budget conversation.

Key Question to ask: Is there measurable evidence that vendor report volume tracks sales cycles or funding rounds rather than incident data?

This paper applies a diagnosis already made in Signals in the Noise. Chapter 2 names the business model directly: ambiguity has become an industry, because platforms reward engagement over accuracy and "the feed learns your fears and reflects them to you" (p. 11). The grading framework is the triage layer. The book supplies the operating disciplines that make the grades hold under pressure.

The book's closing question for any confident dashboard applies to every feed item: "What, exactly, am I looking at, and who wants me to see it this way?" (p. 43).

Before acting on anything from the feed, a leader should be able to answer six questions in under a minute. This extends the book's One-Minute Double Check (Ch. 19) for decision-makers.

A note on fatigue: the constant stream carries a cost in attention and judgment. The Reuters Institute 2025 Digital News Report found that four in ten people across 48 countries avoid the news at least sometimes, and 31% of avoiders said they feel worn out by the volume. Statista summary. The evidence is behavioral, not clinical. The practical fix is structural. Fewer, graded sources on a set schedule beat a continuous unfiltered feed.

This paper is the field version of one idea from Signals in the Noise: Security, Technology, and the Hidden Patterns of Modern Risk. The book covers the full toolkit: decision windows, provenance discipline, data voids, cognitive immunity, and the debate protocol teams use to pressure-test what the feed hands them. Get the book on Amazon.