Hacker News Reader

Comment by slim | Hacker News Reader

removal of content is really like 10% of the cases. the most worrisome type of suppression of speech is what they call "shadow banning" which is really "reach limiting". to the extent people now refrain from using the word "Palestine" to avoid metas algirithms. I experienced this myself. We have a draft law in parliament concerning boyocott of israel. I posted a status about the law (notice it's unrelated to war or anything in the news rights now) and it got almost zero reach. it's not subtil. It's like two orders of magnitude less reach. I know it because I'm a political activist and I always comment on draft laws.

View on HN

> people now refrain from using the word "Palestine" to avoid metas algirithms

It wasn't algorithm, or glitch, or bug. It was a decision made by humans, including you know who. It's a form of crowd control. That's why it's important to have several channels for 'producers'. For others it's a good idea to have several sources. You'll see much more even if each source represents a single side (which is usually the case, 'democracies' are no different).

It's difficult for anyone who doesn't work there to know, but I would guess based on what I know of how ranking decisions are made that you are mistaken. For example I worked there prior to the 2020 election and on Instagram there was no specific ranking rule of any kind related to keywords like that, and no intentional bias to push up one side or another in any way, outside of ads and government officials.

Your system, including subsystems, must have a filter for prohibited words, like n-word, f-word, etc. It's very easy to add another one. That could be another subsystem which very few people have access to. I suspect there are different subsystems. One strait with explanations, another shadow without. The second is something nobody talks about in public as it's difficult to explain, may be even illegal. The same way unwanted talks are being suppressed on other portals. Note, it's independent of recommendations system, which can be played with by assigning scores to words and phrases. Engineers who build those systems don't need to know how exactly they are used and what are the key words.

I can give you a first hand example. One big sh*ty bank didn't let me see my accounts if I refused to activate a debit card. After login they showed a message box, when I pressed 'cancel' they logged me out hoping I will 'click through' to get access. This way they didn't have to explain and could claim it was a bug. It lasted for months till I closed all accounts and they confirmed, but actually didn't do it. What I mean by this is don't expect big company to play nice.

Yes it would be easy, but when I worked there it didn't exist.

From my first hand experience designing these algorithms at Meta and running experiments, and listening to feedback from accounts, I would guess that you are not getting reach because your followers do not interact with what you post. In almost every case people call "shadow ban", the cause was occasionally an actual bug (code doing something undesired and unexpected) or far more often people posting bad content.

The ML is good enough to know whether your followers want to engage with political content. If they don't, Facebook will show them less of it.

> The ML is good enough to know whether your followers want to engage with political content. If they don't, Facebook will show them less of it.

This is a very simplistic explanation. If the GP is a political activist and "always comments on draft laws", I would expect their followers to be willing to engage with political content.

It would explain a lot about how Meta operates that the engineers don't factor things like this in.

We did, they do. When I was there people did "calibration studies" for certain sensitive categories to understand and fix any model biases that would rank content higher for undesirable or unexpected reasons.

I feel a deep discomfort with algorithms optimizing for “engaging” content like this. It feels to me like a form of algorithmic brainwashing, capturing people in filter bubbles. I would prefer it if there were laws against platforms making visibility decisions for the user. If people choose to be in a bubble, that’s a form of personal freedom, if an unknowable algorithm chooses for them, not so much.

Most people on Meta's product follow too many accounts to see everything everyone posts. In an alternate reality where ranking is illegal, the game would be different and worse. Accounts would just repost the same content over and over every few minutes to stay at the top of the feed, or some other strategy to get distribution.

Right but users could be given prominent ways to deal with this - eg tools and buttons to (in bulk at first) unfollow people.

Well, it does not make sense because : 1- I know the baseline of interactions I get usually 2- it's not something that happened to my account specifically, it's everybody, including community managers, for whom posting content is a monetizable skill, who resort to obfuscation when they use the word "Palestine".

I posted 2 things recently. One was link to a game, another to an anime (neither related in any way shape or form to wars nor any news or politics what-so-ever. I got zero responses, not even likes/hearts/laughs. So I assumed FB decided not to show them to any of my friends. "Zero reach".

How do you know the reason for your reason for zero reach is any different than mine? I assumed Meta's algo decided none of my friends would be interested in what I posted, not that I was being censored.

They did respond to my next 2 posts so no idea the difference.

I have heard from others that link posts are somehow downranked on many social media platforms (not just Meta, but especially FB/IG). This is why many people recommend instead text posts without a link, with a callout to a second post containing a "link in comments." This seems common on YouTube community posts also.

I wonder if the word Palestine (and potentially also Israel) trips some ML systems up. I put "Palestinian values" into a random sentiment analysis tool and it's ranked lower than most other countries. They'd both be (understandably) linked to negative content.

Yes that's definitely possible. It's the sort of thing that Meta has teams to try to avoid, but in fast moving information landscapes it's sometimes difficult to keep up.

Reply on news.ycombinator.com