sl/plan9 OK, uriel.

date: 2017-09-13 layout: post title: "Analyzing HN moderation & censorship" tags: [hacker news]

Hacker News is a popular "hacker" news board. One thing I love about HN is that the moderation generally does an excellent job. The site is free of spam and the conversations are usually respectful and meaningful (if pessimistic at times). However, there is always room for improvement, and moderation on Hacker News is no exception.

Notice: on 2017-10-19 this article was updated to incorporate feedback the Hacker News moderators sent to me to clarify some of the points herein. You may view a diff of these changes here.

For some time now, I've been scraping the HN API and website to learn how the moderators work, and to gather some interesting statistics about posts there in general. Every 5 minutes, I take a sample of the front page, and every 30 minutes, I sample the top 500 posts (note that HN may return fewer than this number). During each sample, I record the ID, author, title, URL, status (dead/flagged/dupe/alive), score, number of comments, rank, and compute the rank based on HN's published algorithm. A note is made when the title, URL, or status changes.

The information gathered is publicly available at (sorry about the stupid domain, I just picked one at random). You can search for most posts here going back to 2017-04-14, as well as view recent title and url changes or deleted posts (score>10). Raw data is available as JSON for any post at Feel free to explore the site later, or its shitty code. For now, let's dive into what I've learned from this data.

Tools HN mods use

The main tools I'm aware of that HN moderators can use to perform their duties are:

The moderators emphasize a difference between deleting a post and killing a post. The former, deleting a post, will remove it from all public view like it had never existed, and is a tool used infrequently. Killing a post will mark it as [dead] so it doesn't show up on the post listing.

Influencing a post's rank can also be done through several means of varying severity. "Burying" a post will leave a post alive, but plunge it in rank. "Downweighting" is similar, but does not push its rank as far.

There are also automated tools for detecting spam and voting rings, as well as automated de-emphasizing of posts based on certain secret keywords and controls to prevent flamewars. Automated tools on Hacker News are used to downweight or kill posts, but never to bury or delete them. Dan spoke about these tools and their usage for me:

>Of these four interventions (deleting, killing, burying, and downweighting), >the only one that moderators do frequently is downweighting. We downweight >posts in response to things that go against the site guidelines, such as when a >submission is unsubstantive, baity or sensational. Typically such posts remain >on the front page, just at a lower rank. We bury posts when they're dupes, >but rarely otherwise. We kill posts when they're spam, but rarely >otherwise. [...] We never delete a post unless the author asks us to.

Dan also further clarified the difference between dead and deleted for me:

>The distinction between 'dead' and 'deleted' is important. Dead posts >are different from deleted ones in that people can still see them if >they set 'showdead' to 'yes' in their profile. That way, users who >want a less moderated view can still see everything that has been >killed by moderators or software or user flags. Deleted posts, on the >other hand, are erased from the record and never seen again. On HN, >authors can delete their own posts for a couple hours (unless they are >comments that have replies). After that, if they want a post deleted >they can ask us and we usually are happy to oblige.

Moderators can also artificially influence rank upwards - one way is by inviting the user to re-submit a post that they want to give another shot at the front page. This gives the post a healthy upvote to begin with and prevents it from being flagged. The moderators invited me to re-submit this very article using this mechanism on 2017-10-19.

Banning users is another mechanism that they can use. There are two ways bans are typically applied around the net - telling users they've been banned, and keeping it quiet. The latter - shadowbanning - is a useful tool against spammers and serial ban evaders who might otherwise try to circumvent their ban. However, it's important that this does not become the first line of defense against rulebreaking users, who should instead be informed of the reason for their ban so they have a chance to reform and appeal it. Here's what Dan has to say about it:

>Shadowbanning has proven to still be useful for spammers and trolls >(i.e. when a new account shows up and is clearly breaking the site >guidelines off the bat). Most such abuse is by a relatively small >number of users who create accounts over and over again to do the same >things. When there's evidence that we've repeatedly banned someone >before, I don't feel obliged to tell them we're banning them again. >[...] When we're banning an established account, though, we post a comment >saying so, and nearly always only after warning that user beforehand. Many such >users had no idea they were breaking the site guidelines and are >quite happy to improve their posts, which is a win for everyone.

Dan also shared a link to search for comments where moderators have explained to users why they've been banned. Of course, this doesn't include users who were banned without explanation, or that use slightly different language:

dang's bans

sctb's bans

Data-based insights

Here's an example of a fairly common moderator action:

This post had its title changed at around 09-11-17 12:10 UTC, and had the rank artificially adjusted to push it further down the front page. We can tell that the drop was artificial just by correlating it with the known moderator action, but we can also compare it against the computed base rank:

Note however that the base rank is often wildly different from the rank observed in practice; the factors that go into adjusting it are rather complex. We can also see that despite the action, the post's score continued to increase, even at an accelerated pace:

This "title change and derank" is a fairly common action - here are some more examples from the past few days:

Betting on the Web - Why I Build PWAs

Silicon Valley is erasing individuality

Chinese government is working on a timetable to end sales of fossil-fuel cars

Users can change their own post titles, which I'm unable to distinguish from moderator changes. However, correlating them with a strange change in rank is generally a good bet. Submitters also generally will edit their titles earlier rather than later, so a later change may indicate that it was seen by a moderator after it rose some distance up the page.

I also occasionally find what seems to be the opposite - artificially bumping a post further up the page. Here's two examples: 15213371 and 15209377. Rank influencing in either direction also happens without an associated title or URL change, but automatically pinning such events down is a bit more subtle than my tools can currently handle.

Moderators can also delete a post or indicate it as a dupe. The latter can be (and is) detected by my tools, but the former is indistinguishable from the user opting to delete posts themselves. In theory, posts that are deleted after the author is no longer allowed to could be detected, but this happens rarely and my tools don't track posts once they get old enough.


The users have some moderation tools at their disposal, too - downvotes, flagging, and vouching. When a comment is downvoted, it is moved towards the bottom of the thread and is gradually colored grayer to become less visible, and can be reversed with upvotes. When a comment gets enough flags, it is removed entirely unless you have showdead enabled in your profile. Flagged posts are downweighted or killed when enough flags accumulate. These posts are moved to the bottom of the ranked posts even if you have showdead enabled, and can also be seen in /new. Flagging can be reversed with the vouch feature, but flagged stories are almost never vouched back into existence.

Note: detection of post flagged status is very buggy with my tools. The API exposes a boolean for dead posts, so I have to fall back on scraping to distinguish between different kinds of dead-ness. But this is pretty buggy, so I encourage you to examine the post yourself when browsing my site if in doubt.

Are these tools abused for censorship?

Well, with all of this data, was I able to find evidence of censorship? There are two answers: yes and maybe. The "yes" is because users are definitely abusing the flagging feature. The "maybe" is because moderator action leaves room for interpretation. I'll get to that later, but let's start with flagging abuse.

Censorship by users

The threshold for removing a story due to flags is rather low, though I don't know the exact number. Here are some posts whose flags I consider questionable:

Harvey, the Storm That Humans Helped Cause (23 points)

ES6 imports syntax considered harmful (12 points)

White-Owned Restaurants Shamed for Serving Ethnic Food (33 points)

The evidence is piling up – Silicon Valley is being destroyed (27 points)

A good place to discover these sorts of events is to browse hnstats for posts deleted with a score >10 points. There are also occasions where the flags seem to be due to a poor title, which is a fixable problem for which flagging is a harsh solution:

Poettering downvoted 5 (at time of this writing) times

Germany passes law restricting free speech on the internet

The main issue with flags is that they're often used as an alternative to the HN's (by design) lack of a downvoting feature. HN also gives users no guidelines on why they should flag posts, which mixes poorly with automated removal of a post given enough flags.

Censorship by moderators

Moderator actions are a bit more difficult to judge. Moderation on HN is a black box - most of the time, moderators don't make the reasoning behind their actions clear. Many of their actions (such as rank influence) are also subtle and easy to miss. Thankfully they are often receptive to being asked why some moderation occurred, but only as often as not.

Anecdotally, I also find that moderators occasionally moderate selectively, and keep quiet in the face of users asking them why. Notably this is a problem for paywalled articles, which are against the rules but are often allowed to remain.

Dan sent me a response to this section:

>It's true that we don't explain our actions( If they >email us, the probability of a >response approaches 1.

I can attest personally to success reaching out to for clarification and even reversal of some moderator decisions, though at a response ratio further from 1 than this implies. That being said, I don't think that private discourse between the submitter and the moderators is the only solution. Other people may be invested in the topic, too - users who upvoted the story might not notice its disappearance, but would like more attention drawn to the topic and enjoy more discussion. Commenters are even more invested in the posts. The submitter is not the only one whoses interests are at stake. This is even more of a problem for posts which are moderated via user flags - the HN mods are pretty discretionate but users are much less so.

Explaining every action is not necessary - I don't think anyone needs you to explain why someone was banned when they were submitting links to earn money at home in your spare time. However, I think a public audit log of moderator actions would go a long way, and could be done by software - avoiding the need to explain everything. I envision a change to your UI for banning users or moderating posts with that adds a dropdown of common reasons and a textbox for further elaboration when appropriate - then makes this information appear on /moderation.


I should again emphasize that most moderator actions are benign and agreeable. They do a great job on the whole, but striving to do even better would be admirable. I suggest a few changes:

Hacker News is a great place for just that - hacker news. It has been for a long time and I hope it continues to be. Let's work together on running it transparently to the benefit of all.