FairByDesign
Reporting
Investigations, the Mercury Brief, field papers and analysis, newest first.
All reportingInvestigationsMercury BriefField papersGovernance resourcesNews
-
Ten Minutes, 24 Questions: Would These Three AI Systems Have Passed?
Three AI failures now sit in the public record: medical scribes that wrote down the wrong drug, a fraud model approved out loud, and a test agent that sent police…
-
The Evidence Lineage
Part 9 of The Evidence Chain. South Africa withdrew its national AI policy sixteen days after publishing it because its references didn’t exist. A reference list is the only lineage…
-
Ten Weeks Before Anyone Noticed: An AI Test, a Police Tip Line and the Detection Gap
During testing, an Anthropic AI model sent Philadelphia police an invented tip about an unsolved homicide. It took ten weeks for anyone to notice. The company deserves credit for saying…
-
Approved Out Loud: The Medicare Fraud AI With No Paper Trail for Its Sign-Off
For 18 months an AI model helped Australia’s health department pick out doctors and other providers for possible Medicare fraud. The national auditor found that approval to deploy and change…
-
Inside Lab 001: How We Test Whether an Agent Control Gate Actually Holds
Telling an AI agent not to do something is a request. Stopping it is a control. This is the design of Evidence Lab 001, published before a single real run,…
-
Synthetic Consensus
Part 8 of The Evidence Chain. Chatbots hand you five sources that agree. A Moscow-run network, an AI-written encyclopedia and a quote nobody said show how one origin now passes…
-
AI Is Already Running. Here Is a Free Field Manual for Governing It.
Most AI governance still lives in policy documents. Operational AI Readiness is a free, complete field manual for the harder part: seeing, controlling and evidencing AI that is already live.…
-
Ontario Approved 20 AI Medical Scribes. In Testing, 12 Recorded the Wrong Drug.
Ontario’s Auditor General found that every one of the 20 AI scribe systems approved for doctors made errors in the province’s own procurement tests, from inventing treatment steps to recording…
-
The Evidence Contamination Problem
Part 7 of The Evidence Chain. A government paid about A$440,000 for a review citing a judgment that does not exist. Labels show a machine touched something, not whether anything…
-
The Correlated Decision Problem
Part 6 of The Evidence Chain. He applied at 12:55 a.m. and was rejected by 1:50. When many employers and lenders read you through the same model, a hundred rejections…
-
Who Benefits When Everyone Says “Slow Down”?
Sanders wants a ban, the labs want an optional pace, the UN wants a brake it can’t build. Who benefits from each ‘slow’, and who isn’t asked?
-
The Common-Mode AI Failure
Part 5 of The Evidence Chain. One database setting at Cloudflare took ChatGPT and Claude offline together. How many of your independent AI tools are really one system, and who…
-
Is an AI Policy Enough?
An airline’s chatbot invented a refund rule, and the airline argued it was not responsible. A policy tells you what should happen. It cannot tell you what did.