Adding a similar thumbs up/down affordance that we're using in Simple Summaries right now to give us a qualitative signal at scale for the recommendations and a potential control mechanism. We could introduce a requirement to remove read next and flag it to contributors after a certain number of thumbs downs.

Mimicking the media viewer settings functionality which would deep link a user to the "Reading Preferences" section of the Preferences pages from the Read next module:

Right now that page looks like this:

The primary rationale for doing the deep link is because I want to introduce relatively complex filtering options in "read next" for reader safety and control. More than I would want to include in an overlay on an article page. The tradeoff is bringing the reader away from the article page to do the config.

The Reading Preferences section would look like this:

zero state:

hide read next:

Conditional filters for hiding certain types of recommendations:

NsB: Categories filters would be a multi-select lookup codex component, keyword filters would be an open ended chip input from codex.

Alternative locations:
Right now, having the recommendations in the top right corner of the page implies that a reader will immediately want to jump to another article before reading the article that they're currently on. There are two educated guesses on how the user flow could work that would more closely mimic behaviours we know readers are already doing:

1. After intro

2. After infobox (if one exists)

Most of our linking happens from the infobox today for rabbit hole reading sessions. This assumes that the reader didn't find the link they needed from the intro or infobox and offers alternatives.

In both cases, the recommendations would "stick" in the sidebar below the sticky header after a reader scrolls past these points. These patterns would also "cannibalize" fewer of the blue link clicks that already happen in our articles today. The read next feature would be more of an additional knowledge network exploration affordance rather than a competing one.

Obviously the filtering features (at least) would need storymapping and critique from the design team, but this iteration of the design is done-ish for now, so I'll move to signoff @ovasileva

@KColeman-WMF from a trust and safety perspective, are there other filters/categories here that you think I should consider? Are there other ways we can protect readers from harmful algorithmic recommendations?

In T381337#10407797, @JScherer-WMF wrote:

@KColeman-WMF from a trust and safety perspective, are there other filters/categories here that you think I should consider? Are there other ways we can protect readers from harmful algorithmic recommendations?

Hi Justin, thanks for the ping! I like that you are including an explicit way for users to control their recommendations. The following might be out-of-scope but springs to mind for future exploration:

• In the preferences it could be made even clearer to users what these recommendation preferences will do and how the data will be used / what algorithm drives it / link to the data policy. E.g. "We will use these preferences to recommend articles to you. Your data will remain private, see data policy."
• Similarly, you may want to add a small line of copy in the recommendations box explaining what is driving these suggestions, especially if it is ML generated.
• I did initially wonder if there should be a way to hide recommendations from the outset (within the box itself), similar to a "don't show again" checkbox. You could have a popover/dialog that contains basic settings with a link to set more complex filtering on Special:Preferences if needed.
• The thumb up/down is a good way to assessing overall usefulness of the recommendations box, but I wonder if users could also pinpoint the articles that they are not interested in? Similar to the way you can snooze or hide suggested posts from Instagram.
• On social media platforms it's common practice to have sensitive content controls for feeds. I'm not sure if that's needed here as we aren't dealing with user-generated content in the same way, but there might be learnings we can take from the way those platforms provide simple pre-defined settings (e.g. more, less sensitive content).
• This is all just food for thought, from a T&S perspective I like what you have already and look forward to seeing the results!