If the system is up front about what the inputs are then I think most of the problems evaporate. The user doesn't need to understand the math behind how 0-5 star ratings get translated into movie suggestions, if the user understands that the star rating is the input and not dumb implicit stuff like dwell time, then the user will have a more agreeable experience with the system.
I've recently created a system for assigning tags to files. The system recommends tags that might be applicable to a file given what the file is already tagged with and what other files are tagged with. There are no other inputs, I'm not trying to extract information from the filename or running image recognition on images or anything like that. It's straight forward multi-label classification similar to a naive bayes spam filter. Straight forward but effective; the relevant tags are often at the very top of the suggestions list.
I've had a few non-technical users try it out and so far the response has been very positive. They don't know how the tag suggester works. All I told them is that when they enter a tag, the suggestions for other tags are refined. If all they know about the algorithm is what inputs it takes, they're comfortable with it. If the inputs are well defined, then the whole thing becomes "just a machine". Similar to how a non-technical user doesn't need to understand how magnetrons work to be comfortable with microwaving some popcorn, a user doesn't need to know how machine learning works to be comfortable with a program employing it. Just make sure they understand the general relationship between the buttons on the front and what the outcome is going to be.
Don't make a microwave that tries to be clever by turning up the power if the user is standing impatiently in front of it. If you do that, the users will start to get paranoid/superstitious about the microwave (and rightfully so!)
This is a cool topic, but the examples - FB feed, Netflix, etc. - don't really apply. Those systems aren't trying to maximize user control, they are trying to sustain engagement while giving as much control back to the platform as possible. I'd like to see more work on ML driven UI that is specifically trying to enable user -driven intent.
That's a good distinction; those systems are fundamentally adversarial and I think a lot of users pick up on that, at least unconsciously. Many of the ways users interact with these systems could be described as defensive.
The article talks about how the problem is users don't know how to change the inputs to effect the outputs but I don't think this is really desirable. Users shouldn't be thinking "If I hover my mouse over this for too long then it will stuff up my suggestions" that ends up with users altering their behaviour to suit the AI when the AI should just give you correct suggestions without the user having to understand how it works.
That's a great ideal, but the technology just isn't there.
With any system that tries to adapt to my preferences, I find myself thinking constantly about how my behaviour will influence it. I'm careful about what I upvote, I try to avoid pausing scrolling on uninteresting things, I even avoid reading articles which I don't want to see more of, even if I find the article itself interesting (remember that users are not necessarily logical - e.g. I don't wait clickbait even though I'm just as drawn as anyone else by the titles).
I would happily trade all of that mental load for a couple of buttons after each item "more like this" or "less like this" and a guarantee that nothing else affected the algorithm. No advance in the last ten years has changed my mind about that, and my hopes aren't high for the next ten doing so either.
I think there's a case to be made that even hard AI wouldn't be sufficient to solve the problem. When my bank manager suggests a product to me, I'm doing the same thing - wondering "why are you suggesting this to me, what's motivating this?". I still want inputs and reasoning exposed, not merely for the answer to be, on reflection, correct.
The article describes a major problem with AI/ML-based systems, but the examples they picked are poor.
With recommendations, users don't really care about controlling the results. If you know which movies should be recommended, you would just search for them directly.
The real problem arises in systems where the user does want to achieve some specific results:
* Home automation for lights, heating and cooling,
* Microwaves with clever programs that never do what you want,
* User interfaces that adapt to user behaviour,
* Self-driving cars for other drivers on the road,
I argue that pie menus are much easier for users to operate, understand how they work (selection base purely on the direction between clicks, not the path), and predict how they will respond, and thus trust, than general gesture recognition (often implemented with a black box based on machine learning). Because the pie menu gesture recognition algorithm is obvious, transparent, and Fitts' Law friendly.
Pie menus completely saturate all of "gesture space" with valid easy to distinguish gestures of uniform and high area. But gesture recognition systems like Palm's Graffiti only uses a small subset of all possible gestures, each with different areas and distances from each other (i.e. how easy is it to make each gesture and how distinct each gesture is varies), and most gestures are invalid. There are no possible invalid pie menus gestures -- they all map to well defined selections, so no gesture space is wasted.
Ideally each gesture should have a large area in "gesture space" so it's easy to make, and be as distinct (well separated) from other gestures as possible, so it's harder to make mistakes (like confusing an "n" with a "u"). Pie menus optimize for those ideals. It's also extremely useful that you can correct mistakes in-flight (i.e. reselection, browsing, moving out further to get more "leverage"), so no matter how much you mess up the gesture, you can always correct your mistake and refine your selection.
Once you start writing a sloppy "n" and realize that it may think you're writing a "u", there's no way to correct your mistake, and all you can do is scribble to make sure you input an invalid gesture that's ignored (and hope it isn't mistaken for the "delete all" gesture!), then try again.
Preference is mostly social. A good recommendation system would need to feel more like a friend than an algorithm. Not only does it need to understand your preferences, it needs to convince you that it understands "where you're coming from" in general so you can relax and be honest with it. And even if that can be achieved technology-wise, how could you comfortably accept an AI system as your friend if its being "controlled" on the back end by a whole organization of people you don't know? There's more than just tech missing from this equation.
Ugh, I accidently watched a skateboarding video on facebook about a month agi. Now my feed is full of skating videos and I can’t figure out how to tell facebook to calm down with skating videos and adverts
I've recently created a system for assigning tags to files. The system recommends tags that might be applicable to a file given what the file is already tagged with and what other files are tagged with. There are no other inputs, I'm not trying to extract information from the filename or running image recognition on images or anything like that. It's straight forward multi-label classification similar to a naive bayes spam filter. Straight forward but effective; the relevant tags are often at the very top of the suggestions list.
I've had a few non-technical users try it out and so far the response has been very positive. They don't know how the tag suggester works. All I told them is that when they enter a tag, the suggestions for other tags are refined. If all they know about the algorithm is what inputs it takes, they're comfortable with it. If the inputs are well defined, then the whole thing becomes "just a machine". Similar to how a non-technical user doesn't need to understand how magnetrons work to be comfortable with microwaving some popcorn, a user doesn't need to know how machine learning works to be comfortable with a program employing it. Just make sure they understand the general relationship between the buttons on the front and what the outcome is going to be.
Don't make a microwave that tries to be clever by turning up the power if the user is standing impatiently in front of it. If you do that, the users will start to get paranoid/superstitious about the microwave (and rightfully so!)