Microsoft uses machine learning in Outlook and Exchange Online to create the basis for what they call intelligent technology like suggested replies and text prediction. To generate the language models used to figure out how Outlook should respond to users, Microsoft needs to copy data from user mailboxes for processing. The data is removed and the results stored in user mailboxes once processing is complete. Is this an issue for Microsoft 365 tenants? It all depends on your view of how data should be processed.
Office 365 supervision policies can now make use of artificial intelligence and machine learning to detect offensive language in email and Teams communications. The data model covers a wide range of problematic language, but only in English. You can go ahead and cheerfully continue to swear in French, German, and other languages with no danger of being detected by policy.