Keith Coleman 🌱😀🙌

@soba_roblox Awesome to have another writer on the platform! 🙌 Reach out if your have questions along the way.

これこそメイン垢で呟くべきな気がしますが、コミュニティノートAIライターを作りました。 x.com/i/communitynot… あなたのタイムラインにも、私のAIライターが書いたノートが現れるかも。 誤情報を訂正し、少しでも社会貢献につながれば嬉しいです。 成果が出たら改めてメイン垢に投稿します。

@medicgordus @nicksortor Yeah, it appears to be an erroneous media match -- note was written on a very similar (but different) video on another post, and matched to this one. Note has flipped (no longer showing) and we're looking at the match issue. Thanks for the report.

@nicksortor @kcoleman: the community note on this post has nothing to do with the post or the video. Please consider punishing CN abusers, it wastes so much of our time.

🚨 JUST IN: President Trump DOUBLES DOWN on his $10 BILLION lawsuit against the BBC for DELIBERATELY misrepresenting his words, acting like he incited J6 in a speech Good! Time to BANKRUPT those fake news losers 🔥 "What they did is they had me speaking something I never said!" "They put words in my mouth, but I didn't say them. It was AI generated! I said I NEVER said that." "They ADMIT they're guilty!

THANK YOU FOR YOUR CONTRIBUTIONS!!

The best part is that the deamplification happens organically — people see the note, share it way less, virality plummets. All by the people.

Today we are revising our Creator Revenue Sharing policies to maintain authenticity of content on Timeline and prevent manipulation of the program. During times of war, it is critical that people have access to authentic information on the ground. With today’s AI technologies, it is trivial to create content that can mislead people. Starting now, users who post AI-generated videos of an armed conflict—without adding a disclosure that it was made with AI—will be suspended from Creator Revenue Sharing for 90 days. Subsequent violations will result in a permanent suspension from the program. This will be flagged to us by any post with a Community Note or if the content contains meta data (or other signals) from generative AI tools. We will continue to refine our policies and product to ensure X can be trusted during these critical moments.

Thousands of AI-written notes are showing on X, having been rated helpful by people from different viewpoints. But so far it's been only English-language notes. Today, we're bringing AI note writing to all languages 🌎🌍🌏

@martin_casado @yenkel @GeoffreyHuntley The bottleneck.

@yenkel @GeoffreyHuntley yeah, I often have mine nearby open while I eat. What have we become?!

Haha, the number of people with open laptops nearby has increased massively in the last few weeks. Noone wants to interrupt their agent's flow :)

Epic stats from AI note writers. *Notably*: 1. Their notes are are highly additive to human-written notes, covering substantially disjoint sets of posts in many cases 2. They are far from perfect... which is why we just launched Collaborative Notes (community + AI co-writing) ... but they're certainly off to a solid start!

Is the hypothesis basically: if in-network posts viewed down ⇒ less sense of community? Q for you (and to be very clear this is entirely hypothetical & out of curiosity): Let's say one were to make algo changes and obvious metrics like positive engagement, time spent, etc go up. But it's unknowingly affecting that sense of community. What's a metric the system itself could/should observe that would tell it that, so it can adapt?

@kcoleman I am trying to determine a key parameter of the X algorithm that is no longer documented. And might explain some user experience changes folks have been talking about (less sense of community)

I'm trying to estimate the proportion of tweets (or should I say "SpaceXeets" now?) which are in-network, that is, from people you follow. I used Claude Cowork to scrape and estimate my own prevalence at 16%. It used to be 50%, per old Twitter algo docs. Would anyone here be willing to do an estimate on their own feed? It's one prompt... I want to estimate the fraction of posts on the Twitter "For You" feed that come from people I follow. 1. Create a local file with a list of the twitter handles I follow, by using browser automation to scrape them from the "N following" link on my twitter profile. All following, not just verified. Ensure the number of entries in the file matches the N displayed on my profile page. 2. Now estimate the fraction of tweets in my For You feed that are from accounts I follow. Use browser automation to scroll through first ~300 posts in this feed and count how many are from my follows. Save state along the way so you can recover from javascript errors etc. hopefully without restarting from scratch.

@ChristianAdam33 @CommunityNotes Unexpected. Will look into it.

@kcoleman @CommunityNotes No sir I do not, it is set to English.

Note suggestions are working! Contributors are helping improve collaborative notes with better sources, clearer phrasing, and more. We’re expanding access: now any contributor who can write notes and has a non-negative Writing Impact can suggest improvements.

@digiphile That's where we're headed!

@kcoleman Does the dial go all the way to 11?

Human + AI co-writing is working — people are suggesting ways to make notes better, and Grok is updating with the new insights. Just turned up the dial so more people can make suggestions! 🎚️📈

@ChristianAdam33 @CommunityNotes Do you have Spanish on in your X language settings?

@CommunityNotes Awesome, keep up the good work guys. It would be a bit better if the community notes I get recommended we’re not in Spanish 💔

Been great to get feedback from people diving in & contributing to collaborative notes. Q: Now that you've experienced your suggestions being incorporated (or not 😂) should we add a "Partially incorporated" status to reflect how much of the suggestion got worked into the updated note?

And again! Fiber down. @Starlink up. Only reason I’m able to watch the #SuperBowl !

Haha 😂 This is awesome!

@FanCNsFTW Cool.

@kcoleman It gave me a solid place to start. I spend hours researching for notes; Grok has cut that time significantly. The collaborative suggestion cuts that time even more.

First Collaborative Note to go live on X -- and it went live via our new scoring algorithm, no less! All the new updates coming together. h/t to the new GaussianModel: #gaussian-model" target="_blank" rel="nofollow noopener">communitynotes.x.com/guide/en/under…

@FanCNsFTW How did it feel relative to writing from scratch?

@kcoleman Yes. I upvoted it.

@heskelbalas @Altimor @jaybaxter Yeah a key would be determine accuracy (reliably, with high precision)

@Altimor @kcoleman @jaybaxter Maybe it should be a 2 step process. Is the note true or false. If true then it needs polish but could be shown with a caveat (to increase truth first). As the contributors refine the note then it can update until it hits 'rev 1.0' or whatever you want to call it

I got accepted into X's community notes reviewer program a couple of mos ago, which means I see ~10x more community notes than before And it's made me strongly convinced that the current algorithm is very miscalibrated — 80%+ of the "community note candidates" I see and review 1/ prove that the post is complete misinformation 2/ never get rolled out I can't tell you how many times friends send me tweets like "omg did you see that??" and I have to respond "yeah there's an unposted note on this, turns out it's false" Careful out there, if a claim makes you angry it's very likely bs