1,306 AI assessments later — the safety scores are still terrible
We crossed 1,300 assessments on aisa.to last week. Some patterns are getting clearer with more data, and one keeps standing out.
Safety and responsibility scores are stuck at 45/100 on average. Lowest dimension by far. For context, prompting scores around 55, workflow around 50. But safety? People just aren't thinking about it.
Most users can write decent prompts and chain multi-step workflows. But ask about data privacy, output verification, or when NOT to use AI — blank stares. They adopted the tools without developing the judgement.
This holds across experience levels. Junior devs and senior PMs both score low on safety. It's not a knowledge gap, it's a habits gap.
Every L&D team we talk to says "we need to upskill on AI" but when you show them the data — 62% below Proficient, safety the weakest link — the conversation shifts from "nice to have" to "we have a risk problem."
Full data: aisa.to/state-of-ai-fluency
That “habits gap” is the key bit. In real workflows, safety improves when verification and data-handling checks are built into the process, not left as something people remember after the output. I’d be curious whether scores improve when teams are assessed on an actual task vs a questionnaire.