The one metric that changed how we build Aisa's AI assessment
We've assessed 1,306+ people's AI skills at aisa.to and one data point keeps coming back to haunt us: Safety scores average 45 out of 100. Lowest of all 5 dimensions we measure.
At first we thought it was a data quality issue. It wasn't. People are genuinely bad at the safety side of AI usage — sharing sensitive data with models, no verification process for outputs going to clients, zero awareness of when they're creating liability.
This changed our product roadmap. We used to weight all dimensions equally. Now Safety has outsized influence on persona classification because someone who scores 85 on everything but 30 on Safety is a fundamentally different risk profile than someone at 65 across the board.
Tactical takeaway for other AI tool builders: if you're not measuring how safely people use your product, you're probably missing the full picture of adoption quality.
Full data: aisa.to/state-of-ai-fluency
The 45 safety score is striking but not surprising. We run into the same pattern with clients adopting multi-agent systems. Most teams focus on output quality and skip the governance layer entirely. Our experience is that safety is less about knowledge and more about consequences. People know they should verify outputs, but theres no structural enforcement until something breaks. Did you see any correlation between safety scores and team size or whether they had defined review processes?