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.
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
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?
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.
So under this regime Meta, or any other social media site, can do pretty much any recommendation system they want, so long as they have a UI cluttered with labels and confirmation prompts disclaiming that liking someone, joining a group, adding a friend will affect your feed and recommendations.
> Imagine if every time you clicked on a video, YouTube prompted you to confirm that viewing that video would alter future suggestions.
In practice, I suspect this will make nearly every online interactions - posting a comment, viewing a video, liking a post, etc - accompanied by a confirmation prompt telling the user that this action will affect their recommendations, and pretty quickly users just hit confirm instinctively.
E.g. when viewing a youtube video, users often have to watch 3-5 seconds of an ad, then click "skip ad", before proceeding. Adding a 2nd button "I acknowledge that this will affect my recommendations" is actually a pretty low barrier compared to the interactions already required of the user.
The end result: a web that's roughly got the same recommendation systems, just with the extra enshitification of the now-mandated confirmation prompts.
I really do think that would be annoying enough to snap a good amount of people out of the mindless autoscrolling loop. There's a reason why companies love 1 click buying, for example. At scale, any extra interactions costs real money. The example of the ad is a good one where that's already a kind of high friction interaction, so one extra click is not that much more annoying, but that's not the case for the vast majority of interactions.
Granted, some people will certainly have a higher tolerance for this kind of enshittification. If companies find that the amount of money they can extract from highly targeting a given amount of users is greater than the amount of money they can make from more numerous but less targeted users, then they could choose to go down that path. That's a function of how tolerant the average user is of the confirmation prompts, and how much more money they can make from a targeted user.
We can't control that last variable, but we ultimately could control that first one. If we find that a simple confirmation prompt is not annoying enough for as many people as we'd like, we could make the confirmation prompts more annoying. Maybe make every confirmation prompt have to be shown for at least 5 seconds. Or require a cooldown between multiple confirmations. Or add captcha like challenges. And so on.
In the limit, I think you'd agree that if you had to wait 24 hours before confirming, that would probably be enough to dissuade almost everyone from going through with it, to the point most platforms would try to not have any personalization at all. (I wouldn't be happy with this end result either)
I think even a single, instant confirmation prompt would be enough to cause a sizeable difference. Maybe not enough. Maybe you're right, and it would make barely any difference at all. Then I'd be totally in favor of these more annoying requirements. But as a first step, I'd be happy with a small requirement like this, and progressively making requirements stringier if it proves not enough.
> Doesn't mean you want to see more flipped cars down the road.
It absolutely does mean that seeing as how everybody wants to see the flipped car on the side of the road. The local news reports on the car flipped on the side of the road and not the boring city council meeting for a reason.
That's a mindboggling take, to be honest, to the point I can't help but suspect that you're being contrarian just for the sake of it. I'm absolutely sure that you, yourself, have gone by some terrible scene which you couldn't help but stare at for at least a bit, which you would not classify as something you would like to see more of.
There's a huge difference between something people want to see and something people can't ignore. There is some intersection between those categories, but they are by no means one and the same. And news reports, headlines, thumbnails et al. optimize for the latter, not the former.
And? These are still user's choices. They choose how long to view videos or scroll past them.