<i>> Basically with one touch you can tell Allo that you want to "Always chat in incognito mode going forward," and from that moment on all your messages will be end-to-end encrypted and auto-deleted. You can still interact with the AI, but only if you explicitly invoke it</i><p>YES! That's the <i>minimum</i> privacy level that all messengers should have. From my point of view, the "best of both worlds" would be with the Cloud AI <i>disabled by default</i>, maybe with a one-tap enabling inside the chat window and a notification to the other people in the conversations.<p>But if for whatever reasons companies don't want to make that compromise, then we should <i>at least</i> get an option so that each individual <i>could</i> make end-to-end encryption the <i>default</i> if they <i>want to</i>.<p>Again, this should be the <i>minimum</i>, and not the fact that the messenger simply offers end-to-end encryption somewhere burried in the settings and only works when you explicitly enable it everytime you want to use it. That's not acceptable for an app that cares about privacy from my point of view because it's too <i>inconvenient</i> even for those that actually care about their message privacy.<p>As for the future, homomorphic encryption could be one option (although who knows how long it will take to get here), but if you just want to offer your users "AI enhancements" and you don't just offer that as an excuse to <i>collect their data</i>, then companies could also enable <i>client-side AI</i> in the future. I think this could be a lot closer than many think. If the demand is there we could start seeing AI-accelerators as co-processor in our smartphone SoCs within the next 5 years, especially as Moore's Law slows down to a crawl and we start getting diminishing returns from the "faster 8-core CPUs" in our smartphones.<p>Such chips are already starting to pop-up, but because there's no current demand for them to be in smartphones yet, they are relegated to other more niche markets. In the next 5 years they could become a lot faster, too, especially that now Google is showing the way with ASICs as well.<p><a href="http://www.movidius.com/solutions/machine-vision-algorithms/machine-learning" rel="nofollow">http://www.movidius.com/solutions/machine-vision-algorithms/...</a>