The Twitter filter that never was

Posted in Computing on 2023-07-06

1. The worsening of Twitter
2. The beginning of misinformation
3. Spread to English language
4. Development
5. New Twitter
6. Why did anyone think it worked?
7. Conclusion

A few years ago, a list of keywords began circulating on Twitter, with names like suggest_activity_tweet, suggest_pyle_tweet and suggest_who_to_follow. The idea was that these were internal Twitter identifiers for certain types of tweet, and that by adding these keywords to the mute list, users could block anything from advertising to followers’ likes.

There’s just one problem: they didn’t work. The whole thing was superstition.

This article explores how the idea came to be, how it spread, why people believed it, and how we know it never worked in the first place.

The worsening of Twitter

From April 2014 to July 2019, web Twitter used an older design, which you can still see with the browser extension Old Twitter Layout (Firefox, Chrome). A trait of the old Twitter design was that HTML elements had meaningful class names and attributes, allowing advanced users to customize their Twitter experience with custom CSS or adblocking tools. For example, you could highlight the tweets of people who follow you, or hide retweets entirely.

From February to March 2016, Twitter introduced the non-chronological algorithmic timeline, which was unpopular with advanced users. By March 2017, the algorithmic timeline was made worse by the inclusion of other users’ Likes as evidenced here. Likes were generally lower-quality than intentional retweets, especially considering that many people used Likes very freely as read notifications or to build engagement. The feature had previously been tested in 2014, when Likes were still called Favorites.

The beginning of misinformation

On Japanese Twitter, divided from English-language Twitter by the language barrier, people searched for a way to hide the tweets.

On 5 Mar 2017, Japanese user @strada_n suggested adding a keyword mute for the phrase “〜がいいねしました”, “~ga iine shimashita”, Japanese for “liked”. “It would be interesting to see if this works”, says the user. By 23 March, Japanese artist Nazuki makes this tweet, expressing skepticism on the method since keyword mutes should only work on the Tweet text itself.

On 6 May 2017, a correct way to hide tweets appears, seemingly independently by multiple users, which uses custom CSS or adblocker rules to hide elements with the HTML attribute data-component-context="suggest_activity_tweet". On the same day, similar methods using this attribute are tweeted in Japanese, French and English, with another English tweet using an adblock rule. A Japanese user would later extend the rule to also block promoted tweets.

It’s not until 8 June 2017 that the earlier mute method blows up on Japanese Twitter. This tweet by Spica, which received over 20,000 retweets, notes that they haven’t seen a Like in their timeline since they muted “いいねしました”, although it took a little while for the effect to kick in. After this date, Japanese Twitter is awash with tweets discussing the keyword mute. The method is widely shared, despite tweets finding that it doesn’t work.

On 12 June 2017, this article by user q7z advises users how to apply the custom CSS in Google Chrome. Originally, (Web Archive), the article makes no mention of the mute method, but it is soon added in an addendum. The author believed that the mute method worked, but took several hours to a day to take effect.

On 18 July 2017, Japanese user @mi1tu linked to q7z’s article in this tweet, saying that while the mute method was ineffective, muting the term suggest_activity_tweet did work. q7z updated the article to add it, but with the disclaimer that they hadn’t personally tested it. mi1tu’s tweet received over 4,000 retweets.

A few days later, Japanese Twitter was awash with the new method of muting suggest_activity_tweet.

Spread to English language

By the end of July 2017, muting suggest_activity_tweet began to spread to English language Twitter via bilingual users, such as this tweet. It’s repeated by this tweet by a Japanese translator and this one by an illustrator, both of whom soon post that the method turns out to be ineffective.

It’s not until 13 September 2018 that the idea of muting suggest_activity_tweet begins to spread through English-language Twitter in earnest. One of the first is this tweet, noting that it doesn’t work, followed by this one, which claims that it does. The idea soon spreads throughout English-language Twitter, while still circulating in other languages including Japanese, Spanish, Arabic and Korean.

Development

From there, users began circulating growing lists of keywords to mute. On 17 September 2018, this tweet suggested muting suggest_recycled_tweet_inline and suggest_pyle_tweet. The additing of a fourth muted word, suggest_ranked_timeline_tweet, according to this tweet, is credited with making a set which will even force the timeline to display in chronological order. This tweet, on 18 September, lists eight and credits it with keeping the timeline chronological.

There is little consistency between different users’ mute lists. The keywords most likely were drawn from HTML attributes attached to various categories of unwanted tweets. Most users appear to have accumulated their own set of keywords by copying from other tweets they saw. Some entries begin with capital letters, probably erroneously, while others omit underscores (e.g. suggestpyletweet) or write “suggested” rather than “suggest”. One such list numbers 24 entries, of which 8 are probably erroneous variants. One list even erronously mutes filter: follows -filter:replies, which is rather a Twitter search intended to show chronological timeline, and not intended to be muted.

By June 2019, mentions of the legitimate CSS or adblock method have given way to these lists of collected mute keywords.

New Twitter

In July 2019, the new web Twitter design launched, which as of July 2023 is the current design.

The new Twitter made a major change, which is to eliminate all of the useful attributes from the HTML tags. The new HTML is horribly inefficient, with some elements possessing ten or so class names, none of which are descriptive. A single sample tweet within a timeline, consisting of 186 characters, uses 13,491 bytes of HTML text, more than 72 times as much. A page for a single tweet consists of a full quarter of a megabyte.

Notably, this change means the keywords circulating in mute lists no longer appeared anywhere in web Twitter’s front end. This did not prevent people from continuing to circulate the existing keyword lists, nor did it discourage belief in their efficacy. However, it did prevent new entries from being added to the list, since they no longer had a source of clearly defined HTML attributes to draw from.

For example, on 1 August 2019, this article presented the mute list strategy, along with 15 keywords, all of which appear to be correctly spelled and drawn from Twitter’s HTML attributes. The article presents them as a solution to new Twitter design, following a complaint on the matter by security expert @thegrugq.

The mute lists continued to be shared, even by highly technical users who ought to have been more skeptical. On 3 January 2020, @IanColdwater posted twittermute.txt to GitHub, a list of 21 muted keywords including three misspelled with missing underscores. It received 1,892 Stars at current count, and was lauded with praise in the comments, including Javascript snippets to automatically add the entries to the mutelist. Even the Hacker News thread on the list was limited in its skepticism.

As of July 2023, mute lists are still being distributed on Twitter. In many cases, the list is truncated to a list of 12 entries for a total of 277 characters, constrained by the tweet limit of 280 characters.

Why did anyone think it worked?

There’s a famous gaming anecdote about the beta of Dungeons & Dragons Online. Due to an oversight, it was possible to use the Diplomacy skill on treasure chests. Players began to believe that doing so would increase the treasure, and refused to take part in raids without a Diplomacy master in the party. They continued to believe in it even after the developers explained that it didn’t work. It was complete superstition.

A superstition consists of a set of actions which are which is supposed to attain a desired benefit. Proof of efficacy is not necessary, as long as there is no immediate disproof. A superstition is validated by the number of people who appear to believe it. Once it becomes established, contrary evidence is rationalized away for fear of losing the benefit.

Some users certainly shared the mute list without checking if it actually worked. Other people are saying it, after all, and those people had no particular motive to lie.

Of the users who did try it, surely some did not report when it failed, meaning users had poor feedback on the system’s lack of efficacy. Failure to report negative outcomes is even a problem in mainstream science.

Some users even reported that it seemed to work, but because they did not apply the scientific method, they were unable to prove that the mute list, rather than other associated or unrelated factors, were removing Twitter junk. Possible conflating factors might include:

• Suggested tweets were shown to on first visit after a time, to allow users to catch up on what they missed. Fewer promoted tweets were shown later. Refreshing the page after editing the mute list would therefore trigger the reduction. In fact, the effect would be triggered by the refresh, even if you didn’t edit the mute list.
• Many users would also switch to chronological Latest Tweets timeline as part of the process. This timeline may have been shown less junk content.
• A/B testing may cause users to be shown more ads at first, and fewer ads later on.
• Over the long term, the suggested tweets feature may have been reduced for all users following user feedback or analytics.
• The placebo effect may cause users to assume the mute list was doing something, when it had no effect. An excess of optimism could have biased their perception.
• Users may apply additional methods to block junk at the same time, and conflate the efficacy of those other approaches. One user altered their uBlock Origin settings at the same time as the mute list, and was later unable to replicate the effect.
• Adblock lists may have automatically updated to block Twitter junk. Users of adblock may have seen long-term reductions in suggested tweets and falsely attributed it to mute lists.

Some users noticed that it didn’t work, but invented rationalizations to explain why. Actual theories among mute list users included:

• You need to wait a few hours to a day for it to take effect. (By that point, you have already shared the mutelist and forgotten about it, so the idea spreads.)
• It used to work, but Twitter changed the keywords. As soon as someone discovers the new keywords, you need to need to add them to the mute list.
• It works, but unreliably. It only works sometimes, or only for certain people.
• It does work, but it only reduces the junk tweets, rather than eliminates them.
• It works on Japanese language users, but the feature may not have been rolled out yet for English language users.
• It works, but you have to perform additional steps, such as refreshing the page, force-refreshing the page, or switching to Latest Tweets timeline.
• Twitter Support says it doesn’t work, but they’re probably lying to keep the feature a secret because they don’t want people using it.
• It used to work, but has recently stopped working.

Conclusion

The two biggest factors, almost certainly, were that refreshing the page and switching to the Latest Tweets timeline were effective in clearing junk in the short term. Users who edited their mute list would often perform both actions, allowing the illusion that the mute list was the active ingredient.

At no point did anyone prove that keyword muting worked, or demonstrate a provable mechanic by which it would work. The keywords did work as adblock targets, but only until the July 2019 redesign. Using them as keyword mutes never had any effect because mutes only ever applied to tweet text. Yet, as of writing, people are not only still sharing Twitter keyword mute lists, but also insisting that they’ve observed it working.