I fact-checked my own simulators. Three claims were wrong.
Building with AI makes it easy to produce plausible statistics. Before publishing ten simulators, I traced every cited figure to its source.
Over the summer I built ten interactive simulators about supply chain, AI economics and leadership. Each one opens with a real statistic to ground the scenario. Before putting them on this site, I did something I should have done at the start: I traced every cited figure back to its source.
Seven of the ten needed corrections. Three claims contradicted the very sources they cited.
The three that were wrong
The CEO study. One simulator said that in a study of 1,087 board members who fired a CEO, the number one reason was slow decisions. The study is real, from Leadership IQ. But the top reason was poor change management, at 31%. “Too much talk, not enough action” was fifth, at 22%. The simulator’s whole premise leaned on a ranking the study does not support.
The $234 billion. Another cited a Gartner release about organisations demoting autonomous AI agents after governance failures, and added “$234 billion judged at risk”. The governance finding is real, from May 2026. The $234 billion is from a different Gartner release, about application software spending exposed to agentic AI. Two true facts, joined into a false one.
The 73%. A readiness diagnostic said 73% of organisations cite data quality as their top AI challenge, and that unclear ownership appears in 89% of scaling failures. In the source, data issues were 41% and integration ranked first. Unclear ownership was 49%. The 89% referred to all five root causes combined.
How this happens
None of these were invented from nothing. Each started from a real source. The errors came from compression: a ranking becomes “number one”, two releases blur into one, a combined figure gets attached to a single cause. When you build quickly with AI assistance, this kind of drift is easy, and the result reads perfectly plausibly.
That is exactly what makes it dangerous. Nobody questions a number that sounds right.
What I changed
- Every claim that failed the check was rewritten or removed.
- Every lab page now lists the sources behind its figures, linking to pages I actually opened, not search snippets.
- Where a claim rests on a single secondary source, the app says so on screen.
- Every app page states whether its figures are checked, single-source or produced by the model itself.
The lesson
The speed of building with AI is real. So is the speed of producing confident errors. The fix is not to build slower. It is to put a check between building and publishing, the same way nothing ships to production without a test.
I would rather show ten simulators with honest labels than ten with better-sounding numbers. One wrong statistic, found by the wrong reader, discredits all the others.