Safety scores 45/100 — the one AI skill nobody's practicing
412 assessments in on aisa.to and the data keeps telling the same story. People are getting decent at prompting, pretty good at picking the right tool for the job, but genuinely bad at the safety side — data handling, output verification, knowing where models fail.
Average Safety score: 45/100. Every other dimension clusters around 50-55. That gap isn't noise.
The weird part is it's not a knowledge problem. Most people can tell you AI hallucinates. They just don't have a process for catching it in their own work. They accept output that "sounds right" without checking a single claim.
People who score well here aren't paranoid — they've built habits. Spot-check two facts per output. Think about what data you're pasting in. Have a mental model of what the model is actually bad at.
Building the assessment around observable habits rather than knowledge recall was probably the best product decision we made. Quiz someone on AI risks and they'll ace it. Watch them actually use AI and the verification step just... doesn't happen.
aisa.to/state-of-ai-fluency — full breakdown across all 5 dimensions.
This matches what I see with internal tools too: the weak point is rarely the prompt, it's the acceptance step after the model gives a plausible answer. I try to make every AI-assisted change leave behind something boring to check, like a test case, a dry-run log, or one source link that a human can verify later.