VerifiedIntelligence
Seven Checks That Looked Complete and Weren't
My book, The Hallucination Economy, published 3 October 2026 and is now available in Kindle, paperback and hardcover. It asks how organisations act on AI claims nobody outside the producer can check, and what boards should ask instead. How I used AI tools in writing it, and how that work was checked, is set out in its own chapter.
This week's seven pieces keep finding the same gap, qualified: a governance check that exists is not the same as a check that still measures what it claims to. A vendor cleared every sanctions and certification test while its own product line went unquestioned. A satellite operator's spectrum licence survived a deadline that counts orbit, not effort. A kernel's AI-disclosure tag sat unenforced for six months before anyone compared it to reality. A benchmark's confidence interval became a headline rate, and a pricing label followed a model into a comparison it was never built for. In each case, on this week's evidence, the number or label was accurate on its own terms and still answered the wrong question once trusted at face value. The practical habit, for a boardroom, a procurement form or a vendor pitch: ask what population, condition or date a check actually covers before relying on it.
This week's Saturday analysis, on four AI claims that pass a date check and still answer the wrong question, is worth ten minutes before trusting any benchmark, price or certification at face value.Read the full analysis
Cybersecurity Governance: The Threat Actor Had an Invoice Address
Ten government agencies named a Shanghai-listed vendor, not a state actor, as the tooling behind a hacking operation. It passed every standard check: financials, certification, sanctions. None asked what else its own product line could do.
Orbital Compliance: A Thousand Built, Four Hundred Flying
Amazon built a thousand satellites; only 396 fly. The FCC's spectrum deadline counts orbit, not effort, so building faster never restores lost priority.
AI Governance: The Forty Per Cent Is Not Your Forty Per Cent
McKinsey's 40 per cent scaling AI agents describes only firms above a billion dollars revenue; smaller organisations stayed flat at 22 per cent.
The Same Tag, Pointing Both Ways
The Linux kernel measured its own AI-disclosure tag at 10 per cent use against a believed 90 per cent, and killed it. GCC made the same tag mandatory, but only below a size a human can actually check.
Before the Tool Call
The book's closing chapter arrives: a Logic Firewall checkpoint between an AI agent's decision and its tool call, ruling on identity, scope, capability and reversibility before acting.
The Word With No Meaning
Open source has no agreed meaning for AI until checked against a specific standard. Under one such test, three models pass and four major ones, including GPT-4 and Claude, do not. A vendor's open claim is marketing until verified.
AI Verification: Right Number, Right Date, Wrong Population
Four AI claims this week passed the date check and were still wrong: the population or convention behind the number did not match its use.
Pick one check your organisation already relies on, a vendor clearance, a benchmark, a certification, and ask: what population or condition was it actually measured against, and has anyone checked since?

