AI Observatory finds 47.9% of use is non-work

The AI Observatory’s analysis of 23,158 conversations and 85,633 turns finds that applying an occupational-use filter removes 47.9% of the conversations it studied, including personal, relational, cultural, and safety-relevant interactions. The researchers released the public measurement platform and 145-feature taxonomy alongside the underlying paper, while warning that the opt-in corpus is not a representative survey. Teams designing evaluations or safety dashboards should record what their dataset excludes before treating its usage mix as a product truth; Business Arena’s reliability gap makes the same measurement point from the model side.