Here are the facts about where Twitter’s Trust & Safety and moderation capacity stands today:
tl;dr: While we said goodbye to incredibly talented friends and colleagues yesterday, our core moderation capabilities remain in place.
Yesterday’s reduction in force affected approximately 15% of our Trust & Safety organization (as opposed to approximately 50% cuts company-wide), with our front-line moderation staff experiencing the least impact.
Last week, for security reasons, we restricted access to our internal tools for some users, including some members of my team. Most of the 2,000+ content moderators working on front-line review were not impacted, and access will be fully restored in the coming days.
More than 80% of our incoming content moderation volume was completely unaffected by this access change. The daily volume of moderation actions we take stayed steady through this period.
With early voting underway in the US, our efforts on election integrity — including harmful misinformation that can suppress the vote and combatting state-backed information operations — remain a top priority.
What AREN’T we doing? In the short term, we’ve had to deprioritize a few workflows — like account access (lost password requests), and some suspension appeals. We’re working to get these back online in the days to come.
I’ll continue to share updates about our Trust & Safety work as things evolve in the days to come. Our team's mission of enforcing our policies and protecting the conversations happening on Twitter remains unchanged.
Recruiters: My DMs are open — I have recommendations for incredibly smart and compassionate talent that is now available.
• • •
Missing some Tweet in this thread? You can try to
force a refresh
Since Saturday, we’ve been focused on addressing the surge in hateful conduct on Twitter. We’ve made measurable progress, removing more than 1500 accounts and reducing impressions on this content to nearly zero. Here’s the latest on our work, and what’s next.
Our primary success measure for content moderation is impressions: how many times harmful content is seen by our users. The changes we’ve made have almost entirely eliminated impressions on this content in search and elsewhere across Twitter.
Impressions on this content typically are extremely low, platform-wide. We’re primarily dealing with a focused, short-term trolling campaign. The 1500 accounts we removed don’t correspond with 1500 people; many are repeat bad actors.
Let’s talk for a minute about slurs, hateful conduct, and trolling campaigns.
Bottom line up front: Twitter’s policies haven’t changed. Hateful conduct has no place here. And we’re taking steps to put a stop to an organized effort to make people think we have.
Our Rules prohibit Hateful Conduct. This includes targeting people with dehumanizing content and slurs.
This DOESN’T mean we have a list of words that are always banned. Context matters. For example, our policies are written to protect reclaimed speech. help.twitter.com/en/rules-and-p…
Over the last 48 hours, we’ve seen a small number of accounts post a ton of Tweets that include slurs and other derogatory terms. To give you a sense of scale: More than 50,000 Tweets repeatedly using a particular slur came from just 300 accounts.
Since the start of the Russian invasion of Ukraine, our aim has been to remediate abuse at scale and be transparent about our work to protect the conversation happening on Twitter. Today, we’re sharing two key updates about government affiliated accounts. blog.twitter.com/en_us/topics/c…
Beginning today, we will require the removal of Tweets posted by government or state-affiliated media accounts which share media that depict prisoners of war in the context of the war in Ukraine. help.twitter.com/en/rules-and-p….
We’re doing so in line with international humanitarian law, and in consultation with international human rights groups. To protect essential reporting on the war, some exceptions apply under this guidance where there is a compelling public interest or newsworthy POW content.
We’re adding labels to accounts and Tweets sharing links of state-affiliated media outlets in Belarus after detailed reporting about their role in the war in Ukraine. This builds on our years-long work to add context to state media outlets and limit their reach on Twitter. 🧵
Last week, we launched labels on Tweets sharing links to Russian state-affiliated news media.
Early data suggests that our interventions here are working: We've seen a 30% drop in impressions on Tweets labeled under this expanded policy.
Why label Belarusian outlets now? Expert voices have highlighted Belarus’ involvement in the war in Ukraine. Coupled with what we’re seeing on the service around engagement with state-affiliated media, we see transparency about these outlets as key.
Have been sitting with the Apple announcement for a couple of days to try to avoid tweeting a regrettable hot take. This thread from @alexstamos just about captures it for me — as well as @gruber’s writeup: daringfireball.net/2021/08/apple_…
More than anything, announcing these 3 fundamentally distinct features together feels like a (rare) colossal PR misstep by Apple. Grouping together anti-CSE tech (broadly good) with parental controls (somewhere between meh and dangerous) muddies the waters unproductively.
Apple in 2021 has a trust deficit with the security and privacy community — with good reason (nytimes.com/2021/05/17/tec…). The choices they’ve made about how to do business in China enable the slippery-slope arguments we’ve seen here.
We've seen no evidence to support the claim that "nearly half of the accounts Tweeting about #COVID19 are likely bots.” 🧵 with a few thoughts on the subject... npr.org/sections/coron…
First, we should get our terms straight: "Bot" means a lot of different things to a lot of different people — and doesn't necessarily refer to coordinated, manipulative, or inauthentic behavior.
There are lots of different ways of using Twitter. Some people Tweet a lot; some never Tweet at all, or only Tweet sporadically. There's no right or wrong way to use Twitter — and many "bot" studies wind up dismissing a lot of real activity as inauthentic.