Twitter Study Tracker – After the 2022 Takeover


A little bird out on a limb, looking nervously back at the tree.(Cartoon by Hilda Bastian.)

This is the home for my collection of studies and other data sources on the Musk era of Twitter. There are now 21 studies of Twitter in the Musk era published from November 2022 to the middle of August 2023 in this post, and 5 of the migration from Twitter to Mastodon, for a total of 25 studies (one study fit into both).

I've grouped the 21 Twitter studies into 4 subject categories. This post has a brief overall summary, then a list of the 4 categories, linked to a slightly fuller summary and details of the studies. Right below that list (and before the detailed sections begin), there's a section on background, process, and update history.

This tracker provided the study basis for some of my posts at Absolutely Maybe:


5 studies show an explosion of hate speech in the immediate aftermath of the takeover. For example, the use of some hate speech went from around 80 an hour to well over 4,500 an hour. In another study, hate speech quickly quadrupled, before reducing to a level still much higher than previously. GLAAD's scorecard of LGBTQ social media safety found improvement in other platforms between 2022 and 2023, but a plunge in safety at Twitter. 

There has also been a large increase in engagement with Twitter accounts sharing misinformation and links from low-credibility websites, and a small drop in engagement with Twitter accounts of traditional media (2 studies). There has been a major drop in pro-climate-science/concern tweeting, with a third of these accounts no longer active, likely as a result of increased toxicity – climate denialist tweets are now as common as tweets sharing science/concern about climate change (1 study).

There has been a 20% to 30% drop in the number of tweets (4 studies), predominantly from people from the center-left politically. From 10% to 25% of the most active users have stopped or dramatically reduced tweeting (2 studies). Three studies suggest that the reduction in activity of some groups of users may be much higher: nearly half of scientists (1 study) and people concerned about climate change (1 study), and the rate of tweets from the infosec (security) community dropped to only a quarter the previous rate. Right-wing users, including those associated with far right networks, may have had a large increase in engagement (2 studies). Note: Although it's not clear how Twitter calculates "global advertising reach", its reports to advertisers have a 33% drop in reach in the first quarter of 2023, compared to the previous quarter (Kemp 2023).

There have been 3 studies of people who could be identified as having both Twitter and Mastodon accounts in the wake of the takeover. The number of followers of the Twitter accounts was higher than the number of followers of the Mastodon accounts (3 studies). Most of the identifiable Twitter users with Mastodon accounts were posting from both of them, and the content was usually different – they posted more toxic tweets than toxic Mastodon posts (1 study). One survey found that most scientist respondents with Twitter accounts said that they were tweeting less, some not at all, and Mastodon was the most common social media alternative where new accounts were opened.
The list of categories:
  1. Hate speech and online safety [Includes 10 studies]
  2. Misinformation [Includes 3 studies]
  3. Other patterns in Twitter use [Includes 8 studies]
  4. The Twitter to Mastodon migration [Includes 5 studies]

Background and process:

This page was last updated on August 18 (update history below). I began this tracker at the end of January 2023, as a data appendix to this post – "Social Media Regeneration and Divisiveness" – at Living With Evidence. I collect the studies via Google Scholar as they appear there, and via Mastodon and other reading.

I'll be adding studies as the stream of research on what's happening, broadly defined, grows. Studies added in the last update are marked *. I'd be grateful to hear of studies I've missed. You can use comments here to contact me – I moderate them all, so let me know if you don't want me to publish a message. Or you can talk to me on Mastodon, at

Other related posts of mine:

Hilda Bastian.

This is version 5 of this post, dated August 18, 2023.

First version posted on January 30, 2023.

Version 2: Pfeffer 2023 added on February 13, 2023.

Version 3: Bin Zia, Miller, Carniel, and Uni of Michigan (all 2023) added on March 22, 2023.

Version 4: Auten, Chavalarias, Hammer, Hickey, Jeong, La Cava, and Pew Research (all 2023) added on June 4, 2023, as well as an additional study mentioned, but not included – by the Center for Countering Digital Hate (2023c). On June 6, the Iffy Quotient data was updated, and 3 studies by the Center for Countering Digital Hate were added (2022, 2023a, 2023b).

Version 5: Journal versions for Hickey and Jikeli preprints added, and the note for the Bitterman pre-registered study. New studies added: Chang, GLAAD, Jacobs, Siebert, Valero.

Studies of hate speech and online safety

Summary: 5 studies show an explosion of hate speech in the immediate aftermath of the takeover – racism, anti-semitism, and/or homophobia were measured in these studies. For example, the use of some hate speech went from around 80 an hour to well over 4,500 an hour. In another study, hate speech quickly quadrupled, before reducing to a level still much higher than previously.

A further study concluded that Musk's personal tweets appear to be appealing to people who retweet hate groups' tweets. Finally, GLAAD found a major drop in provisions for LGBTQ online safety at Twitter in 2023 compared to 2022.

(Note: Musk took over Twitter on October 27.)

Studies not counted: 

If there's no reasonable research report publicly available or at least meaningful information about how data was gathered, I don't include studies. The following 2 were more borderline, so I've detailed them here in case people are interested in them.

(1/2) Center for Countering Digital Hate (2023c).

This study involved the CCDH reporting 100 tweets promoting hate from Twitter Blue users. I couldn't find enough detail about the methods for this, and what's reported shows some biases. For example, there's no control data, and Twitter's actions only within 4 days of the reports are reported. (Only 1 of the tweets were taken down in that time, and the tweets were still apparently getting boosted by Twitter's algorithm favoring Twitter Blue.)

(2/2) Washington Post article by Joseph Menn discussing unpublished research (2023).

This article discussed upcoming research suggesting that hate speech on Twitter coincide with increases in anti-semitic and homophobic violence in the US, but there wasn't enough information to weigh that conclusion.

Included studies, from the most recently published:

* GLAAD (2023).

GLAAD, a LGBTQ advocacy organization in the US, analyzed 5 social media platforms in 2022 and 2023, using a Social Media Safety Index (SMSI): Facebook, Instagram, TikTok, Twitter, andYouTube. Each of them is rated on a score with12 indicators for policy and practice in a scorecard, and the companies are contacted to review their data. Twitter came last of the 5 by a long way – scoring 33% compared to Instagram's 63%. That was a drop of 12 points from the previous year – the other 4 each had an increase (of between 9 and 15 points).

Trevor Auten (2023).

This is a masters thesis. Using data from Twitter's API, Auten created networks of tweets from 1 week in 2021 and 1 in 2022, 6 weeks after Musk's takeover, and looked at the relationship between Musk's own Twitter account and those of hate groups (3 types: White nationalists/alt-right, anti-semites, and anti-LGBTQ). He concluded that people who retweet hate groups' tweets were also more likely to retweet Musk's tweets, suggesting that Musk's rhetoric is encouraging to hate groups.

Daniel Hickey and colleagues (2023a preprint; *2023b paper).

The authors created a lexicon of 49 keywords for hate speech, supplemented with a filter for detecting toxic content to identify "a rude, disrespectful, or unreasonable comment that is likely to make people leave a discussion," and filter out porn (the Perspective API). After identifying a group of tweets between October and November 2022 from the Twitter research API that met their criteria, the researchers collected all the tweets from those accounts in the month before and after the Musk takeover. They also identified a control group the same way (using non-hateful tweets), and used Botometer to look for bot activity.

The hateful tweeters' hateful tweets quadrupled after the takeover, then settled back to a level higher than previous to the takeover.

Scores for bot activity changed. While the overall amount of bot activity remained similar, spammer and fake follower accounts increased for both hateful and control accounts, while astroturf bot activity decreased for the control accounts. (Astroturf bots here were ones that artificially inflate support for political candidates and smear opponents.)

Center for Countering Digital Hate (2023b).

This is based on a search of tweets that used both a keyword for LGBTQ+ together with any of 3 specific slurs (while excluding wedding/marriage terms). The researchers used SNScrape (Python code for scraping Twitter). They did not use any other method (such as a filter) to determine how often the context was hate speech. That's critical, because people can use a term to call out someone else's hate speech.

The researchers identified over 1.7 million such tweets between January 1, 2022 and February 28, 2023. Before the takeover, the average was over 3,011 per day, rising to 6,596 after the takeover.

They concluded that 5 accounts are driving the "grooming" slur on Twitter. From December 15, when Twitter began releasing tweet impressions, to the end of February, tweets from those 5 accounts were getting an average of over 18 million impressions daily. (Impressions are the number of times Twitter serves the tweet into a timeline or search.)

Screenshots of ads from major companies appearing on those accounts are included, along with an estimate of how many millions in advertising revenue Twitter could be making on these accounts alone over a year.

Center for Countering Digital Hate (2023a).

This is a tally of "tweet impressions" for 10 reinstated Twitter accounts. Tweet impressions are the number of times that a tweet has been served in a timeline or search result. The accounts had previously been banned for hateful activity, misinformation, or both.

The 10 accounts posted 9,615 tweets from December 15, 2022 to January 30, 2023, with over 2.5 billion impressions – averaging over 54 million a day – with 1.6 billion of them for a single account. 

Screenshots of ads from major companies appearing on those accounts are included, along with an estimate of how many millions in advertising revenue Twitter could be making on these accounts alone over a year.

Carl Miller and colleagues (2023).

This study classified tweets, using Twitter's API, identified by language models as "plausibly antisemitic", in English only. Between June 1 of 2022 and February 9 of 2023, there were 325,739 of these tweets sent from 146,516. Before the takeover the weekly averages was 6,204. Afterwards, the weekly average was 12,762.

The researchers found a significant surge in new accounts posting these tweets: 3,855 such accounts were created in the week after the takeover.

While some of these tweets are being taken down, possibly by Twitter, the rate of takedowns did not keep up, according to the researchers, "with the increases in absolute volume of antisemitic content."

Bond Benton and colleagues (2022).

A study of the use of a previously moderated homophobic slur before and after the mass shooting at an LGBTQIA nightclub in Colorado on November 19/20. The authors analyzed tweets from November 15 to November 23. They used Tweet Binder, a Twitter analytics marketing service.

The authors explain the QAnon context of the use of forms of the term, groomer. There was some increase in use of the term in the days before the shooting, and then a dramatic increase in the use of the term after the shooting, both as homophobic slurs and denouncing those tweets. The authors concluded, "Cumulatively, results would seem to indicate what results of a less stringently moderated Twitter may look like."

Gunther Jikeli and Katharina Soemer (2022, *2023).

A study of anti-semitism on Twitter between January and November 2022, with a focus on October 1 to November 6, "when the number of conversations about Jews more than doubled," and the period after Musk's takeover. Several millions of tweets are included. Of the more than 100,000 about Jews in the first peak after Musk's takeover (October 30), the authors randomly sampled and annotated 100.

Most tweets were decrying anti-semitism, but there was an increase in anti-semitic tweets. They concluded that some users were trying out whether they could use a particular slur in the new Twitter.

(This study presumably used Twitter's API – they say they "scraped" tweets from "the Twitter archive.")

Center for Countering Digital Hate (2022).

This is a fact check of a Musk claim about hate speech declining at Twitter in November 2022. They used a social media analytic tool, Brandwatch, to gather data on the frequency of tweets using a slur (from a list of 6 related to race and sexuality) and retweets of them in the week beginning October 31, compared to the 2022 average. However, they did not use any other method (such as a filter) to determine how often the context was hate speech. That's critical, because people can use a term to call out someone else's hate speech.

The Center identified a major increase in the use of these words, up to triple.

The report includes a link to a tweet from Twitter's then-head of trust and safety (Yoel Roth) acknowledging that there had been a surge in hate speech, but the team had succeeded in reducing the prominence of these tweets, so that they did not spread as much.


Bond Benton and colleagues (2022).

A study of the use of particular hate terms (eg racist, anti-semitic, and homophobic slurs) from October 22 to October 28, with sentiment analysis (estimating whether or not the use of the word was hostile).

The hate speech they tracked went from 80 tweets an hour before the takeover to well over 4,500 an hour afterwards. The researchers used the Tweet Binder tool.



Summary: There was a small drop in engagement with traditional media Twitter accounts, and a very large increase in engagement with misinformation superspreaders and accounts sharing links from low-credibility websites (2 studies).

The engagement-weighted proportion of shared links from news websites classified as "iffy" (frequently sharing misinformation) has been higher at Twitter post-Musk than at any other time since 2016 other than after the 2020 US presidential election (1 study), and consistently 2 or 3 times higher than the proportion at Facebook.

The profile of climate denialism is far higher since the Musk takeover. While climate denialist tweets were outnumbered 2 to 1 previously, they are now roughly similar. About a third of people who had previously posted pro-climate-change-science stopped doing so after the Musk takeover (1 study). That's believed to be because of an increase in toxicity to them and a reduction in views due to algorithmic changes driving them off Twitter.

Studies, from the most recently published:

(See also Center for Counteracting Digital Hate (2023a) in the hate speech group above.)

University of Michigan Iffy Quotient (2023).

The Iffy Quotient is "a metric for how much content from 'iffy' websites has been amplified on Facebook and Twitter." It's run by the School of Information Center for Social Media Responsibility at the University of Michigan. Ratings of websites as "iffy" for this project are those made by NewsGuard and Media Bias/Fact Check. "Iffy" means the site is biased, or frequent reporters of false information.

Then, a media monitoring service, NewsWhip, is used to monitor links shared on Facebook and Twitter from over 400,000 sites, and provide an engagement score. Those sites haven't all been rated, so the sites NewsWhip is monitoring can be rated "OK" or "iffy", or be classified as "unknown". NewsWhip's engagement score is used by the University project to weight popular links.

The weekly quotient for Twitter and Facebook have been charted charted since January 2016. It's the proportion of the monitored shared links that are "iffy": Higher is worse. The rate of links from unrated websites has been fairly constant across time.

In the early years, Facebook had a higher "iffy" quotient than Twitter. However, Twitter has been higher almost constantly since April 2019. The highest ever engagement-weighted "iffy" quotient for either platform was 30% at Twitter in November 2020 (the time of the US presidential election). However, in the week prior to Musk's takeover, it was 11%. It climbed quickly after that: Musk-era Twitter includes the second-highest peak, at 26%. It has fluctuated a bit, but almost every week has been higher than 2022 pre-Musk, and hasn't been below 10%. At Facebook in the same period, it's been generally below 5% (with a peak of 8% and a low of 4%).

David Chavalarias and colleagues (2023).

This is a report of the Climatoscope project, where Twitter debates about climate change in English and French has analyzed over 400 million tweets from 2016 to March 2023, drawn from Twitter's API. The researchers found a sharp decline after the Musk takeover in "pro"-climate change accounts – about a third ceased tweeting – and those of their supporters, as well as in daily tweeting concern about climate change. There was only the usual seasonal change in quantity of climate change denialist accounts and tweeting.

The result of this shift was that while climate-concerned tweets were nearly twice as common as denialist ones pre-Musk (6.6 million vs 3.5 million), post-Musk they were similar (3.3 million vs 3.1 million).

The researchers suggest several factors that appear to have contributed to this: An increase in toxicity towards people concerned about climate change, algorithmic changes favoring denialism and toxicity, and both these likely contributing to people concerned about climate change deserting the platform.

The researchers concluded that post-Musk, Twitter became "an active place of conversion to climate skepticism". They note that Russian misinformation activity and aggressive tweeting had also contributed to this trend from the time of the invasion of Ukraine.

Bastien Carniel and colleagues (2023).

This study was based on a database of websites labeled as propagators of misinformation by fact-checking organizations in multiple countries and languages. A list of 514 Twitter accounts were labeled misinformation superspreaders, as they repeatedly shared links from these websites. Those superspreaders tweeted 1.5 million times between September 1 and December 31, 2022 (before and after the takeover). For comparison, the researchers collected over 640,000 tweets from 130 traditional media Twitter accounts.

The researchers found a small drop in engagement with the tweets from traditional media (– 6.3%) and a large increase in engagement with the superspreaders (42.4%): "Immediately after the acquisition, total engagement with tweets posted by superspreaders spiked, almost doubling overnight and remaining at elevated levels ever since. Since the number of tweets from these accounts has remained roughly constant, this jump is only explained by an increase in the average engagement with these accounts' posts."

Other patterns in Twitter use

Summary:  After a few months, there had been a roughly 20% to 30% overall reduction in tweets (2 studies), with about a quarter of the previously most active US Tweeters no longer as active (1 study). In another study, about 10% of US users surveyed said they had stopped using Twitter. (In the misinformation section of this post, a study of the climate debate in English and French found that about a third of the pro-climate-concern accounts were no longer active.)

Some groups may have a much higher rate of reduced use, including around half of scientists (1 study) and people who tweet concern about biodiversity and climate change (1 study), with infosec tweets dropped to about a quarter of the previous rate (1 study). People who stopped or reduced tweeting were more likely to be from the center-left politically (3 studies).  However, right-wing and far-right accounts, had a large increase in activity and engagement (2 studies), and non-right-wing Twitter Blue subscribers also experienced an increase in retweets (1 study).

There was a major divergence in opinions about their experience on post-Musk Twitter for US users who are women, black, and Democratic or Democratic-leaning: They were more likely to say they had taken a break and were unlikely to still be using Twitter in a year, and that incivility and misinformation were major problems (1 study).

Note on baseline data: Musk took over Twitter on October 27. Jürgen Pfeffer and colleagues (2023) collected a complete 24-hours' of tweeting before the Musk takeover. On September 21, 375 million tweets were sent from over 40 million accounts. About 1% of the tweets came from less than 100 accounts, and roughly 175,000 accounts (0.44% of accounts) produced 50% of the tweets.

Studies, from the most recently published:

* Myriam Vidal Valero (2023).

In this study for a journal news feature, a database of email addresses of over 170,000 corresponding authors who had tweeted about their own publications was used. (The database had been compiled by Wenceslao Arroyo-Machado and colleagues for other analyses [2023]. The authors had been identifiable in February 2023, and the papers – gathered from the Web of Science – covered over 250 disciplines.)

The authors were emailed in July, and asked if their Twitter use had changed in the previous 6 months, and if they had opened accounts on any other social media platform in the previous year. The response rate was extremely low – about 5% – and no data is provided to give an idea who the 9,153 respondents were, and how they compared to non-respondents.

Of the respondents, 7% said they had stopped using Twitter, 24% that their use had "significantly decreased", and 23% that it had "slightly decreased". Of the other half, most said their use hadn't changed (37% of the total), while for 9% it had increased significantly (3%) or slightly (6%).

Nearly half (46%) said they had opened a new social media account in the previous year. The social media platforms with proportions of these scientists in the 2-digits were:

  1. Mastodon (47% - 1,976 scientists)
  2. LinkedIn (35%)
  3. Instagram (29%)
  4. Threads (25%) (it had just rolled out at the time of the survey)
  5. Facebook (22%).

*  Jay Jacobs (2023).

This is an analysis of tweets from the Twitter API for Infosec Twitter – discussions on common information security vulnerabilities. Tweets (but not retweets) with particular keywords had been collected from 12 July 2021, ending in July 2023 when charges for the data were introduced.

Across that time, there had been over 1,200 of those tweets each week day, and under 500 on weekend days. There was no major change in the first few months after the takeover. However, from April to May 2023, Jacobs describes Infosec Twitter as dying. By the last 2 weeks, there were less than 300 tweets a week day. He wrote, "And with that, we say 'so long' to infosec Twitter."

Charlotte Chang and colleagues (2023).

In this study, the authors compare 380,000 participants in "Environmental Twitter" (frequently posting about biodiversity and mitigating climate change) with a control group of 458,000 participants in what they called "Politics Twitter" (accounts posting on the 2020 US presidential election). Commenting on that election was so widespread, the authors believed those accounts represented "a broader slice of Twitter as a whole". (The accounts tweeting on both subject areas were deleted from the "Politics Twitter" group.)

The authors analyzed the proportion of accounts that were active in each 15-day period between July 2019 and April 2023, with active defined as an existing account tweeting at least once in the period. The data was grouped into pre- and post-takeover samples. By April 2023, 53% of the "Environmental" group were still active versus 79% of the "Politics" group, with a rapid decline from the time of the takeover.

The authors don't report the language of the accounts, but it appears to be English-speaking Twitter. Nor do they report the geographical spread, so it's not clear if a more international group is being compared to a more US-based group.

Luca Hammer and Martina Schories (2023). (In German – this is my own translation.)

This is a study of over 1.2 billion tweets in German between December 2020 and May 2023, with a gap of several months at the start of 2023 because of access issues with Twitter's API. The researchers estimate the sample is around 90% of all German-language tweets. There is information for over 4.2 million Twitter accounts, and follow networks for March 2023 of over 600,000 accounts that had tweeted more than 100 times in the sample. 

The researchers determined that the data for April were the highest quality in 2023, so they compared April data from 2021, 2022, and 2023. The number of tweets dropped from close to 4 million that month in 2021 and 2022, to just over 3 million in 2023. There were over 1.5 million tweets per day in April 2023. However, right-wing accounts posted 64% more tweets in April 2023 than in April 2021, the only cluster with a major increase in tweeting. Otherwise, people's bubbles on Twitter were getting smaller, and people were tweeting less.

Pew Research Center,  American Trends Panel survey, Wave 123 in March 2023 (2023b, 2023c).

This is a survey of 10,701 people drawn from a representative panel of internet users in the US, with Asian, Black, and Hispanic users oversampled. (Methods.)

Although Twitter-using respondents from this panel in 2021 hadn't differed much along partisan political lines on whether civility or misleading information were a major problem on Twitter, they were very divided in March 2023. From 36-39% believing incivility was a major problem in 2021, 50% of Democratic users said it was in 2023 vs 27% of Republican users. From 52-54% believing misinformation was a major problem in 2021, 68% of Democratic users said it was in 2023 vs 37% of Republicans.

More Democratic users than Republican ones believe harassment and abuse from other users is a major problem: from 50% of Democratic users in 2021 to 65% in 2023, vs 41% of Republican users in 2021 dropping to 29% in 2023.

They report that 60% of Twitter users said they had taken a break of weeks or months from Twitter in the past year, with women more likely to (69% vs 54% of men). Black users were also more likely to say they took a break than white users (67% vs 60%). However, there was no context to interpret this, as the researchers didn't report a comparison question from past surveys, or a difference before and after the Musk takeover.

Asked if they expected to be using Twitter a year from now, 25% said they were not very likely to, or not at all. The rate of Democratic women saying this, however, was 35%. Again, there were no comparisons from the past, or pre- and post- the Musk era. Overall, the rate for Democratic users was 29% versus 20% of Republican users.

Note: Democratic above is a combination of Pew's "Democratic and Democratic-leaning"; Republican, "Republican and Republican-leaning."

Pew Research Center, American Trends Panel survey, Wave 119 in December 2022 and respondents' tweets from January 2022 to April 2023 (2023a).

This is a survey of 11,004 people drawn from a representative panel of internet users in the US. (Methods.) The researchers also used Twitter's API to collect respondents' tweets from January 1, 2022 to April 10, 2023.

Awareness of the Musk takeover was high: About 40% had mentioned Musk in a tweet or tweets since early 2022, particularly Republican Twitter users.

As usual, a minority of people do the bulk of the tweeting: 20% of this group posted 98% of the tweets. The bulk of them were Democratic users (61%) – again, similar to previously. However, 20% of the most active tweeters before the Musk takeover were no longer among the most active after it; and a quarter of the top 10% before were no longer so after the takeover. And the most active users were also tweeting less (about 25% fewer tweets).

Note: Democratic above is a combination of Pew's "Democratic and Democratic-leaning"; Republican, "Republican and Republican-leaning."

Jonathan Schulman and colleagues (2023).

More than 24,000 Americans were surveyed in December 2022/January 2023. Surveys had also been run at 3 prior points in the 2022, each with over 20,000 participants. There was an overall drop of 3 percentage points in people reporting they used Twitter – or roughly 10% of individual users. Democrats had been at 37-38% at each of the earlier 3 time points, but were 33% in December/January. The authors concluded: 

"This decline was driven by Democrats. For Republicans and independents, there were statistically insignificant declines in Twitter usage between the two surveys, but the percentage of Democrats fell from 38% to 33%. Prior to Musk purchasing Twitter, our surveys had shown reported Twitter usage to be stable within each party, where Democrats and independents have been consistently more likely to report using Twitter than Republicans.

Despite the decrease in usage and trust of Twitter by Democrats since Musk’s takeover, there were no significant changes among Democrats in reported usage of Twitter specifically as a source for political news. There was an increase in the share of Republicans who reported using Twitter as a source of news, from 10% to 14%."

Christopher Barrie (2022).

From November 9-11 there was a release of paid "blue check" subscriptions, and this study used a database of 138,000 Twitter users that signed up. The author looked for 1,000 of those accounts with close connections to far right networks (from another database) and found 961 of them, with around 4.8 million tweets between May 17 and November 23, 2022. He also looked at tweets from 933 users from the "blue check" database that were not in the far-right-associated database.

The 961 contentious users had a 70% increase in retweets after the takeover; the other 933 had a 28% increase. (This study couldn't assess overall engagement on Twitter in that time period.)

Twitter to Mastodon migration


There are 3 studies of the Twitter to Mastodon migration analyzing a minority of Twitter users who set up Mastodon accounts post-Musk. Another study analyzed Twitter and Mastodon presence for a group of the most highly-cited scientists. Finally, a survey of scientific authors already included above, asking if they had set up new social media accounts. That survey found that the most common platform for new accounts was Mastodon.

In the early period after the takeover, people with accounts on both Twitter and Mastodon had more followers on Twitter (3 studies). In the first few weeks after the takeover, most of the Twitter users with Mastodon accounts studied were posting from both of them, and the content was usually different. They posted more toxic tweets than toxic Mastodon posts (1 study). The sample of most highly-cited scientists found Mastodon accounts were rare, while Twitter accounts were common, though generally not highly active with very large following.

* Note: A study on open science and the Mastodon migration has been preregistered by André Bitterman and colleagues (2023).

Studies (most recent first):

* Myriam Vidal Valero (2023).

This report is detailed in the "Other patterns" section above. Among the over 4,200 respondents who reported that they had opened a new social media account in the year ending in July 2023, Mastodon was the common (47% of those respondents). Altogether, only 5 platforms accounted for a double-digit proportion:

  1. Mastodon (47% - 1,976 scientists)
  2. LinkedIn (35%)
  3. Instagram (29%)
  4. Threads (25%) (it had just rolled out at the time of the survey)
  5. Facebook (22%).

Lucio La Cava and colleagues (2023).

This is a study of around 75,000 people with accounts on both Twitter and Mastodon. They were found by searching for Mastodon account names in 2 million tweets using the hashtag TwitterMigration from October 26, 2022 to January 19, 2023, as well as the half a million accounts posting those tweets. (Only about 2,500 of those Mastodon accounts included the corresponding Twitter account in the bio.)

The Twitter accounts had about 4 million followers; the corresponding Mastodon accounts, 2.5 million. About 30% of the Twitter-Twitter relationships could be found in the Mastodon-Mastodon networks. The authors suggested this implied people using #TwitterMigration were trying to replicate their Twitter networks within Mastodon. 

The researchers' analysis concluded that networks on Twitter were very loosely connected – which doesn't facilitate rapid spread of information. After analyzing the language in tweets using the hashtag, the researchers suggested that information exchange and a sense of belonging to a group or community may have driven the migration phenomenon. However, this could be a feature of using the hashtag in the first place: We don't know what proportion of people from Twitter who set up accounts at Mastodon used that hashtag – 75,000 is a very small proportion of people who set up Mastodon accounts during that time.

Ujun Jeong and colleagues (2023).

This is a study of 10,169 people with accounts on both Twitter and Mastodon (on intances accessible via the Mastodon API), from October 2022 to January 2023. This sample is skewed by not including the largest Mastodon instance ( and others that weren't open for registration throughout. The researchers developed their sample by searching for @ mentions of 10 Mastodon instances in Twitter bios, and mentions of Twitter handles in Mastodon bios.

There were close to 1.2 million tweets from these accounts and over 0.5 million retweets, and over 725,000 Mastodon posts and over 0.5 million re-posts. The researchers also crawled the followers of these accounts on both Twitter (over 53.5 million) and Mastodon (over 2.9 million). 

These people were focusing more on building connections with other users on Mastodon, where they were newcomers, than they were on Twitter, where they had established networks. However, they were getting more responses from other users on Twitter than on Mastodon, reflecting the smaller size of the networks of their new accounts: 14% received no responses on Mastodon vs 3% on Twitter.

Analysis of hashtags found that Twitter accounts with higher followers predominantly used social issue hashtags (such as Covid and Ukraine) whereas those with fewer followers were topic-oriented. On Mastodon, however, hashtags were topic-oriented, across the board. They concluded that this reflected Twitter's algorithmic push of users with higher follower counts, whereas Mastodon relies more on shared interests, an "approach which provides an opportunity for users from diverse occupations to gain popularity".

The researchers reported some analyses of the groups of people who directed their time back to Twitter and those who remained active on Mastodon after the early surges. Unfortunately, they don't report the numbers here. They suggested that people who are mostly focused on attracting a large user response "tend to discontinue using the platform [Mastodon] over time." Those who stayed were more likely to be having frequent conversations with people from a variety of parts of Mastodon.

A major focus of this paper is their data on inferred occupations of the users. However, with inferences available on only 57% of the people, I'm not including that here.

* Maximilan Siebert and colleagues (2023).

This study analyzes presence on Twitter and Mastodon between December 2022 and February 2023 for a sample of 900 most highly cited scientists derived from Scopus data. To be included, the scientist had to have published a paper in 2022, and then 300 of the top 2% were randomly selected in each of 3 "age" bands (from years they were publishing). Then authors searched for Twitter and Mastodon accounts for the 900 manually. The Twitter API was used to extract data on Twitter accounts; Mastodon data was gathered manually.

Twitter accounts were common (29%), but only 9 Mastodon accounts were found (1%). Of the 262 Twitter accounts, 6 had more than 10,000 followers and none had more than 100,000 followers. Between December and February, 4 of the 262 left Twitter (2%). Of the 9 Mastodon accounts, all had been opened between October and December, and all of the scientists had Twitter accounts as well (3% of the total group with Twitter accounts). No data was included on those accounts.

Haris Bin Zia and colleagues (2023).

This is another study of Twitter users that had Mastodon accounts in October and November 2022. The researchers looked for all tweets with links to Mastodon instances, as well as a group of keywords related to Twitter migration. In the weeks after Musk's takeover (Oct 26 to Nov 21), they found over 2 million tweets from over 1 million Twitter accounts. They looked for Mastodon user accounts in the tweets, and in the bios of the users posting those tweets, identifying 136,009 Twitter users with Mastodon accounts. (72% of the accounts had the same name on Twitter and Mastodon, and 4% were blue-check-verified.)

The researchers then crawled the Mastodon and Twitter timelines of those 136,009 users (except for just over 5% of Twitter accounts which were deactivated, suspended, or in protected status, and some Mastodon instances which were down at the time of the crawl). (Note: that's in line with traditional social media research practice, but not with the expectations of Mastodon users, some of whom explicitly object to having their data downloaded without their consent.)

The researchers estimated that just over 2% had deactivated their Twitter accounts. 9% of the users hadn't posted from their Mastodon accounts and 4% didn't follow anyone on Mastodon. 21% of the Mastodon accounts had been created before the takeover.

There were over 16 million tweets from these users, and over 5.7 million Mastodon posts. The median number of followers for the Twitter accounts was over 700, and by late November, the median number of followers for the Mastodon accounts was less than 50.

The authors concluded that there was "a continuous growth in user activity on Mastodon after the acquisition of Twitter. However, the activity of migrated users on Twitter do not decrease in parallel, i.e. our migrated users are using both their Twitter and Mastodon accounts simultaneously." Most (85%) were posting differently on the two platforms, though some were mirroring posts in the two places.

There was a low rate of posts measured as toxic for these users, but the rate of toxicity was higher on their Twitter accounts: 6% versus 3% on Mastodon.

No comments:

Post a Comment