it isn’t the tech that is bad. Self hosting a model for a task that suits it like speech recognition for a disabled person is righteous and liberating.

Even that is tricky. One must check how the model itself was trained, namely :

• was the trained data acquired rightfully (e.g. not stolen images without creator permission)
• was the training done properly (e.g. not out of a datacenter running on polluting generators)
• was the annotation of the training data done within worker rights (e.g. psychological support if violence or porn data to remove)

rather than just solely have a positive use case and being privacy preserving.