My baseline for technology is the web, it’s where I work and it’s where the majority of the world’s services end up. If not entirely, then at least through backend server interfaces. In that space Progressive Web Apps (PWAs) are a paradigm for introducing client side features to users only when their devices are capable of using them and related to that, progressive disclosure does for cognitive load what PWAs do for device capabilities; features are opt-in (though this isn’t true everywhere). GDPR regulations in the EU also empower users to control how their data is used by companies. Given how energy-intensive & polluting language models & generative AI are and the questionable datasets & value they operative with, I think it’d be prudent for software maintainers to make AI features opt-in. That doesn’t mean they should be hidden or inaccessible, but that account/user preferences and onboarding wizards should at least make it clear how to disable them, if they are not disabled by default. The loading of client-side models and the display of generated text & images should be done as a result of explicit user interaction. This way services won’t use bandwidth, energy, or screen space to do work users won’t make use of. It saves money for companies and improves the experience for users. Making AI optional also gives conscientious users the opportunity to make their most of your software without going elsewhere entirely.
Of course, this opinion is misaligned with marketing teams and product managers who push AI as a vehicle for collecting training data, to drive up perceived market value, as a sincere effort to help users, or as a combination of the three. In these cases users may opt-out of AI features, so consideration should be made for how a user could be automatically off-boarded from an AI enhancement they aren’t using (ala auto-unsubscribe in email newsletters), with the option to come back later provided through replacement UI. There is no technical reason to impose AI features on all users and a ton of software has worked perfectly fine for decades without it. We already meet users where they are for device capabilities, workflow preferences, and data processing so we should progressively enhance, disclose, and enable with AI features too.