The "95% of AI pilots fail" figure lost its hedge in its own report's executive summary. Anthropic's self-replication result lost "upper bound" on its way to 133,000 impressions. In both cases the dropped hedge was the line that told you what to watch. My rule now is to carry the unit, the sample and the strongest hedge, or leave the number out.
"Open" in AI usually means you can download the weights under a license. Open source means you can rebuild, modify, and redistribute. The useful question is what each one lets an outsider check.
AI revenue headlines often mix gross bookings, revenue shared with partners, and run-rate extrapolations. Which number you report changes what the business looks like, and the gap is where the hype lives.