Here is my read:

Developers are refusing to complete the survey or selecting themselves out because they (apparently) don’t want to complete the non-AI task.

The also saw selection effects from a large reduction in the pay for the study (which is an unfortunate confounder here), 150/hr -> 50/hr.

They guess this makes their estimates lower bounds, but the selection effect is complicated (which they acknowledge).

Overall this is a hard problem for them in the current state. It will be challenging to produce convincing year over year analysis under these conditions.