This is false.

eg Gemini enterprise has a note under each chat session explicitly saying it’s not used to improve the models

ZDR policies are a thing. Maybe you can't trust all of them, but when a provider like AWS says it, you can probably believe them. They'd be setting themselves up for an endless stream of massive lawsuits from other large companies if they were lying about it.

Local models are also a thing. The Qwen 3.8 I'm running on a spare computer is definitely not training on my inputs.