There's a lot of very simple questions you can answer very reliably, too, like "what proportion of the users are still using a certain compatibility flag".
Interpreting data is very hard.
Funny, I’d assume the exact opposite. After all much of the understanding of privacy and statistics at scale was developed after 1980.
More than that, the interesting stats may be not even around user-visible features, but around internal mechanisms, like some cache hit rate, or how often is some branch in the compiler invoked.
As long as stats are clearly inspectable, reasonably anonymized, and are opt-out, I'd be fine with sending them.
See for example Mozilla's bizarre decision to remove the ability to change the override the character encoding of a webpage with some half-baked detector.
Meanwhile, I'll decide what I must or must not adapt to, thank you.
But, that costs money: folks don't generally like to do surveys for free, don't like to come to your site (or have you come to theirs) for free, don't like to participate in tests for free... and the caliber of person who can design and orchestrate all that is _notably_ expensive.
So, companies have discovered that it's far, far cheaper to shove remote eyes into their products. No need to schedule user observation sessions, pay for transportation and participation, have to bother with hiring people who are capable of coming up with a good set of questions and scenarios to walk through.
This actually happens sometimes: when developing the demo ".kkrieger", a first-person 3D shooter in 96 KiB, demogroup theprodukkt tried to shrink it down to get it under the 96 KiB wire. One of the tricks they used was using a profiler to identify code sections that were never reached and could be removed. One of the sections they removed was the handler for the up arrow key in the main menu, simply because the test player never pressed up in the menu.
If you think that Google or another large software organization won't misuse telemetry by cutting or neglecting important but infrequently used functionality to hit some KPI... have you ever worked in a large software organization?
All stats can be deanonymized. The more data you make available, the more you identify yourself. I do not need software I use stealthily tying up bandwidth by "phoning home" with data about me. It is simultaneously betrayal and resource theft. If I wanted to contribute to the improvement of the software, I'd file a bug report.
But since you clearly don't like telemetry, you should have a way to reliably switch it off. Here we are on the same page: there must be a well-documented and easy way to switch any telemetry off.
Telemetry should be opt-in, if it's provided at all.
Amen
Sometimes some pretty niche feature, that is used by one individual that is loud, gets more attention than a feature, used by thousands or millions that are just silent users.
It does sort of hinge on the highly suspect assumption that usefulness is correlated with use. An obvious counter-example to this is something like a fire-extinguisher, which will in the ideal case just sit on a wall until it's use-by date passes and then it's discarded having never been used; or on the flip side, an incredibly byzantine workflow that could be reduced to something much simpler will appear important and useful.
Even without these edge cases, interpreting statistics is really hard. Like people with PhDs who have studied these things for years still get them wrong all the time.
What ends up happening more often than not is it's used as a tool to quiet the critics when pushing through unpopular changes.