4 ms·

Even that, I don't think is entirely true. I'm pretty sure they use signals as implicit as how long you took to scroll past an autoplaying video, or if you even

Even that, I don't think is entirely true. I'm pretty sure they use signals as implicit as how long you took to scroll past an autoplaying video, or if you even hovered your mouse pointer over the video but ultimately didn't click on it. Same with friends, even if you have the exact same friends, if you message friend A more than friend B, and this otherwise identical account does the opposite, than the recommendation engine will give you different friend-related suggestions. Then there's geolocation data, connection type/speeds, OS and browser type, account name (which, if they are real names such as on Facebook, can be used to infer age, race, etc), and many others, which can also be taken into account for further tailoring suggestions. You can say that, oh, but some automated system that sent the exact same signals on all these fronts would end up with the same recommendations, which I guess is probably true, but it's definitely not reasonable. No two (human) users would ever be able to achieve such state for any extended period of time. That's why we are arguing that only explicit individual actions should be allowed into these systems. You can maybe argue what would count as an explicit action. You mention adding friends, I don't think that should count as an explicit action for changing your content feed, but I can see that being debated. Maybe the ultimate solution could be legislation requiring that any action that influences recommendation engines to be explicitly labeled as such (similar to how advertising needs to be labeled), and maybe require at least a confirmation prompt, instead of working with a single click. Then platforms would be incentivized to ask as little as possible, as otherwise confirming every single action would become a bit vexing.

Manuel_D 11mo ago
> Even that, I don't think is entirely true. I'm pretty sure they use signals as implicit as how long you took to scroll past an autoplaying video, And? These are still user's choices. They choose how long to view videos or scroll past them.
- rafabulsing 11mo ago
  People also slow down to look at the flipped car on the side of the road. Doesn't mean you want to see more flipped cars down the road. Either way. Do you have any points other than that you think any and every action, no matter how small, is explicit, and therefore it's ok that for it to be fed into the recommendation engine? Cause that's an ok position to have, even if one I disagree with. But if that's all, I think that's as long as this conversation needs to go. But if there's any nuance I'm failing to get, or you have comments on other points I raised such as labeling of recommendation altering actions, I'm happy to hear it.
  - Manuel_D 11mo ago
    I'm mostly interested in getting concrete answers as to what people are referring to when they talk about "algorithmically served" content. This kind of phrasing is thrown around a lot, and I'm still unsure by what people are referring to by this phrase and I've rarely found people proposing fleshed out ideas as to how to define "algorithmically served content". Some people take the stance that even using view counts as part of ranking should result in a company listing section 230 protections, e.g. https://news.ycombinator.com/item?id=46027529 https://news.ycombinator.com/item?id=46027529 You proposed an interesting framing around reproducibility of content ranking, as in two users who have the exact same watch history, liked posts, group memberships, etc. should have the same content served to them. But in subsequent responses it sounds like reproducibility isn't enough, certain actions shouldn't be used for recommendation even if it is entirely reproducible. My reading is that in your model, there are "small" actions that user take that shouldn't be used for recommendations, and presumably there are also "big" actions that are okay to use for recommendation. If that's the case, then what user actions would you permit to be used for recommendations and which ones would not be permitted to use? What are the delineation between "small" and "big" actions? Labeling is another idea but Meta does, in fact, disclaim which user actions are used for content recommendations: https://transparency.meta.com/features/ranking-and-content/ https://transparency.meta.com/features/ranking-and-content/ Basically, I still don't have a clear picture of what does and doesn't qualify as "algorithmically served" content in your model.
    - rafabulsing 11mo ago
      As I pointed out, I agree that defining what should be deemed acceptable and what shouldn't is a bit subjective, and can definitely be debated. Reasonable people can disagree here, for sure. That's why I proposed that maybe the solution is: 1. only explicit actions are considered. A click, a tap, an interaction, but not just viewing, hovering, or scrolling past. That's an objective distinction that we already have legal framework for. You always have to explicitly mark the "I accept the terms and conditions" box, for example. It can't be the default, and you can't have a system where just by entering the website it is considered that you accepted the terms. 2. explicitly labeling and confirming of what is an suggestion algorithm altering action and what isn't. And I mean, in band, visible labeling right there in the UI, not a separate page like that Meta link. Click the "Subscribe" button, you get a confirmation popup "Subscribing will make it so that this content appears in your feed. Confirm/Cancel". Any personalized input into the suggestion algorithm should be labelled as such. So companies can use any inputs they see fit, but the user must explicitly give them these inputs, and the platforms will be incentivized to keep this number as low as possible, as, in the limit, having to confirm every single interaction would be annoying and drive users away. Imagine if every time you clicked on a video, YouTube prompted you to confirm that viewing that video would alter future suggestions. I'm ok with global state being fed into the algorithm by default. Total watch time/votes/comments/whatever. My main problem is with hyper personalized, targeted, self reinforcing feeds.

Manuel_D 11mo ago

> Even that, I don't think is entirely true. I'm pretty sure they use signals as implicit as how long you took to scroll past an autoplaying video, And? These are still user's choices. They choose how long to view videos or scroll past them.

rafabulsing 11mo ago
People also slow down to look at the flipped car on the side of the road. Doesn't mean you want to see more flipped cars down the road. Either way. Do you have any points other than that you think any and every action, no matter how small, is explicit, and therefore it's ok that for it to be fed into the recommendation engine? Cause that's an ok position to have, even if one I disagree with. But if that's all, I think that's as long as this conversation needs to go. But if there's any nuance I'm failing to get, or you have comments on other points I raised such as labeling of recommendation altering actions, I'm happy to hear it.
- Manuel_D 11mo ago
  I'm mostly interested in getting concrete answers as to what people are referring to when they talk about "algorithmically served" content. This kind of phrasing is thrown around a lot, and I'm still unsure by what people are referring to by this phrase and I've rarely found people proposing fleshed out ideas as to how to define "algorithmically served content". Some people take the stance that even using view counts as part of ranking should result in a company listing section 230 protections, e.g. https://news.ycombinator.com/item?id=46027529 https://news.ycombinator.com/item?id=46027529 You proposed an interesting framing around reproducibility of content ranking, as in two users who have the exact same watch history, liked posts, group memberships, etc. should have the same content served to them. But in subsequent responses it sounds like reproducibility isn't enough, certain actions shouldn't be used for recommendation even if it is entirely reproducible. My reading is that in your model, there are "small" actions that user take that shouldn't be used for recommendations, and presumably there are also "big" actions that are okay to use for recommendation. If that's the case, then what user actions would you permit to be used for recommendations and which ones would not be permitted to use? What are the delineation between "small" and "big" actions? Labeling is another idea but Meta does, in fact, disclaim which user actions are used for content recommendations: https://transparency.meta.com/features/ranking-and-content/ https://transparency.meta.com/features/ranking-and-content/ Basically, I still don't have a clear picture of what does and doesn't qualify as "algorithmically served" content in your model.
  - rafabulsing 11mo ago
    As I pointed out, I agree that defining what should be deemed acceptable and what shouldn't is a bit subjective, and can definitely be debated. Reasonable people can disagree here, for sure. That's why I proposed that maybe the solution is: 1. only explicit actions are considered. A click, a tap, an interaction, but not just viewing, hovering, or scrolling past. That's an objective distinction that we already have legal framework for. You always have to explicitly mark the "I accept the terms and conditions" box, for example. It can't be the default, and you can't have a system where just by entering the website it is considered that you accepted the terms. 2. explicitly labeling and confirming of what is an suggestion algorithm altering action and what isn't. And I mean, in band, visible labeling right there in the UI, not a separate page like that Meta link. Click the "Subscribe" button, you get a confirmation popup "Subscribing will make it so that this content appears in your feed. Confirm/Cancel". Any personalized input into the suggestion algorithm should be labelled as such. So companies can use any inputs they see fit, but the user must explicitly give them these inputs, and the platforms will be incentivized to keep this number as low as possible, as, in the limit, having to confirm every single interaction would be annoying and drive users away. Imagine if every time you clicked on a video, YouTube prompted you to confirm that viewing that video would alter future suggestions. I'm ok with global state being fed into the algorithm by default. Total watch time/votes/comments/whatever. My main problem is with hyper personalized, targeted, self reinforcing feeds.