Just in time, Proton is offering their version of privacy respecting AI service. This seems to be middle ground somewhere between the worst offenders and running your own AI in you own computer. See this blog announcement for more details. Although it seems muuuuch better than the alternatives when it comes to privacy and security, there’s very little info on (other) ethical aspects, which is a shame.
[Edit to add: If you are interested, Proton has a feedback form. My questions about this are below. Anyone can send similar (there or any other AI service). NOTE: I got a reply to this - more or less an AI generated response with “Thanks” and links to read all the info that I was giving feed back - not a good sing 
]
Questions
Hi. I read about you AI service announcement. Could you please elaborate, preferably publicly, on the ethical aspects of your AI service, AI system and AI models, as details we lacking in the blog post. Here are a few questions off the top of my head, but I encourage you to look for others as well, via ethicality frameworks (for instance: https://d-nb.info/1229620141/34) or such.
- What data was used to train them models, how, and where was it gathered - or is that fully unknown?
- What kind of quality, filtering, selection and/or censoring methodologies and techniques were used?
- Were humans used to check and moderate the data - if so, where are they based and what was their compensation and support (regarding filtering worst of internet)?
- What is the nature and mix of the energy used to run the services and what is/was it when the models are/were trained?
- Are the models filtered or censored - do they prevent or limit the processing of sensitive information, like in medical context or minority groups?
- Does the system have any safety features/fences regarding inputs - and can user set those, for example to protect from accidental input of personal data (own or others)?
- You advocate to use the service for even anything sensitive, even work related, but that may seriously be against organizational policies and even local law - is there a checklist or legal document that can be used to demonstrate compliance?
- How is the feedback, corrections and notifications of areas where models have troubles handled and notified?
- How is it determined that the given output is “correct enough”, how is that level of confidence communicated to user and can user modify this setting?
- What are the processes or possible standards used to oversee the quality of models without logs and how does this impact usability?
[Edit to add: Meanwhile (with a tip of the cap to mr. Colbert’s segment), Windows 11 is implementing more AI to read what you have on your screen (starting in the US). It’s not called Recall anymore…]