Briefing: X handed fact-checking to the crowd. Then it became the crowd.

Our weekly Briefing is free, but you should upgrade to access all of our reporting, resources, and a monthly workshop.

Craig shared more than two dozen tips and tools he discovered at USC's Open Source Journalism Conference. One highlight is an awesome tool for investigating GitHub accounts and organizations.

We also published the latest episode of Show & Tell. Our guest Shayan Sardarizadeh told us all about his approach to debunking viral fakes for the BBC Verify service. Check it out on YouTube, Spotify or Apple Podcasts.

Craig and Tom Vaillant, the developer of the new Indicator Lab tools, are hosting a free webinar on Oct. 28. They’ll share hands-on tips and case studies on using AI agents to enhance open-source investigations. Info and registration here.

This opening item is co-authored with special guest Renée DiResta.

Two weeks ago, a group of X engineers published a preprint about the Community Writer, an open-source AI notewriter that contributes to Community Notes, the platform’s crowdsourced fact-checking program. They describe the tool as its “largest AI API client.” That’s an understatement. 

In September, the six accounts powered by the Community Writer accounted for more than 90% of all notes contributed by AI writers. The six accounts also wrote 55% of all notes rated “Helpful,” which are those shown to X users below the original posts.

X’s decentralized fact-checking feature has relied on AI bots for most of its published content for a while now. But the paper, related code, and this thread by Community Notes engineering lead Jay Baxter reveal something that was not previously clear: The bots behind most of the published notes were actually operated by X. 

A core goal of Community Notes was to move away from a system where X (then Twitter) decided what was misinformation, and hand moderation power to its users. As its own Community Writer-powered accounts now dominate the program, X has ended up reclaiming much of that power.

Notes must still be approved by a diverse set of human raters, so the program retains a community-based filter. It’s also possible that X’s Community Writer accounts are more successful at earning “Helpful” notes because their content is better. But the data on their activity suggests there may be some structural advantages that give X’s accounts a leg up over other bots.

First, X assigns AI writers a quota of notes they can contribute, which starts at 10 notes a day for new bots. They can earn their way into publishing as many as 500/day by achieving a good overall helpfulness rate. However, this requires that a bot’s notes actually get enough ratings — a core challenge for humans and bots alike. Eight-eight percent of all notes are currently stuck in a “Needs More Ratings” purgatory, and that share has been pretty stable over the past 3 years.

X’s AI note writers aren’t necessarily better than the others. Taking into account deleted notes, about 9% of their notes get rated “Helpful.” That’s in line with other top-performing AI bots and worse than some of the most prolific human writers.

What X has on its side is access and scale. It runs its own models on its own systems and can afford to draft tens of thousands of notes. X.com is also a top cited source in Community Notes, because corrective context is often available in a different post. An outsider would have to pay to use Grok or access the X API in order to source the same data; X’s engineers can do it for free.

Scale compounds the benefits of access. According to table 2 in X’s preprint, its six bots generated about 240,000 draft notes in June, of which about 3,600 were rated helpful. It would be expensive for an unpaid outsider to match such output. Data shows X’s AI writers also arrive first with a proposed note on 90% of posts that are also “noted” by another AI bot. And because its writers are top-rated, they have access to real-time data on how their notes are getting rated. (Smaller contributors like Renée have to pull data from the public files which have a ~2 day lag). It all makes for a significant advantage that reinforces the status of those on top because they can delete notes before they are recorded as “Not Helpful,” if that’s the direction user ratings are taking. It’s important that the program not allow for spammy bad writers that waste raters’ time. But the combination of limited human raters and the inability to delete notes likely influences what small note writers contribute.

All this is important because it reshapes the proposition of Community Notes. The contributions of X’s six AI bots dwarf those of the other dozen or so continuously active AI writers. X has essentially re-centralized the content, if not the application, of misinformation labels on the platform. This also impacts what gets fact-checked. 

As Renée will cover in an upcoming post, polarizing topics inherently struggle to attract cross-partisan positive ratings — regardless of accuracy. Because high “helpfulness” unlocks perks that boost the future likelihood of being rated helpful, the current structure disincentivizes writers from focusing on topics most in need of clarification and context. Why risk getting a “Not Helpful” rating on Gaza, vaccines, or elections if you are optimizing for getting published?

Indeed, by coding Community Writer contributions to controversial topics using a keyword method used in this paper to every helpful note since April, we find that X’s AI writers mention elections, Israel/Gaza, vaccines or immigration in about two-thirds as many notes as the other AI writers do — though more often than human writers. This may be because X’s Community Writer is optimized for identifying posts on which it has the highest chance of getting a note past raters.

In other ways, X’s AI writers don’t appear all too different from other synthetic note writers and their human counterparts. All three groups cite X itself and Wikipedia the most in their notes. X’s AI writers appear to rely more than human writers on sources that Elon Musk has expressed antipathy towards like Wikipedia, fact-checking websites, and traditional media outlets.

At a more granular level, we also manually reviewed 150 randomly sampled “Helpful” notes from X’s AI writers and the other AI writers between April 1 to Oct 9. We find that X’s writers seem to cover political topics noticeably less than the other automated fact-checkers, but otherwise the key stats are pretty similar.

There’s a lot more work to be done to understand how maximizing cross-partisan “helpfulness” scores might affect the topics an AI writer chooses. We think there is value in Community Notes-like programs to combat misinformation. There’s plenty of valuable insight around automated fact-checking in the paper. And X has opened up a lab where this can be tested in a real-world setting. It also continues to share data and the code powering the Community Writer, presumably in an effort to reduce the over-reliance on its own note writers. Meta and TikTok introduced similar crowdsourced fact-checking features to their platforms without the same level of transparency.

But Community Notes on X also has clear structural challenges. Its credibility comes from a theory of diverse perspectives: people who usually disagree agreeing that a note helped. The raters are still diverse. The writers increasingly aren’t. If this program is going to work, and not be just Grok fact-checking people through multiple channels, other entities will have to step up and fund comparable scaled systems. But they’ll only show up if X makes it worth their while. Publishing the code was a real step. Addressing the economic incentives and structural challenges is the next one. – Alexios and Renée

📍 Anthropic said in a blog post that 5.5% of sampled Claude US user conversations about politics were about fact-checking (see chart above). This sort of transparency is helpful, but as I wrote after OpenAI shared similar data about scams, it’s high time for AI labs to create public interfaces à la Google Trends so the rest of us can also evaluate what people are doing on chatbots. – Alexios

📍 Ukrainian president Volodymyr Zelenskyy said that strikes on Yandex-operated data centers made it “a little easier to breathe on many social platforms.” Politicians, news media, and online commentators took it a step further, variously claiming that a collapse in Yandex bot activity led to a sharp drop in activity tied to right-wing parties and/or conspiracy-minded personalities and communities. Over the week, a whole lot of experts weighed in to explain why there was no evidence to support such claims.

📍 MrBeast is apparently hiring for a “Principal Product Manager, CX & Platform Integrity.” Make of that what you will.

📍 Brazil’s electoral court called on the country’s two presidential contenders to stop weaponizing the process of filing legal requests for content removal. AFP reports that “the campaigns of the bitter rivals have filed a record number of censorship requests over what they call online misinformation from the opposing camp.”

📍 Also in Brazil, Meta promised to apply content bans fairly across presidential candidates after users found that its AI image generator refused prompts containing the word “Lula” but not the word “Flávio,” the first name of the front-runner Flávio Bolsonaro. (Two T&S hot takes: it looks like query-based blocklists are still a thing, even in the age of AI; and while the rules should apply equitably, I’m not sure whether it’s worse to be the candidate whose likeness could or could not be deepfaked… – Alexios)

📍 Press Gazette reports that Quartz’s website is “mainly produced by one AI-assisted reporter” who publishes more than 50 stories per day. 

📍 Philadelphia PD called on Anthropic to “strengthen its safeguards” after the company disclosed with a two-month lag that one of its agents submitted made-up information about an unsolved murder through the police department’s tip line. (The tip was never investigated.)

Who Sells Data is a database tracking both the companies selling AI training data and those buying it from organizations. It’s a useful resource if you want to get a sense of players in this growing industry, and the kind of data being bought and sold.

As of this writing, it lists 38 data buyers, and 159 companies that sell data used for AI training, simulation environments for agents, data annotation, and more. You can filter by the type of data they provide, how much they pay for data, and other categories. It’s maintained by a company that sells a de-identification product. — Craig

📍 Skopenow released the beta version of X-Ray, a tool that “combines AI detection, metadata analysis, available provenance, and open‑web matches to help investigators assess authenticity and identify what needs further review.” It works with images, PDFs, audio, and video. You can run three free searches a day during the beta release.

📍 Thomas Jean Lekhal shared that it’s now free to run company shareholder searches in Thailand. “Until recently, you'd pay the DBD ฿50 ($1.5) for shareholder records. Now, anyone can access current shareholder names from government records online at no cost, “ he said.

📍 The operational atlas of mapping tools GEOINT and OSINT is a list of more than 70 tools and resources for geolocation and other tasks. Created by Joel Jerez and shared by Skip Schiphorst.

📍 Andrew Deck wrote, “‘Uber agents,’ undercover AI personas, and other ways newsrooms are using AI in investigations.”

📍 Molfar wrote, “OSINT Email Search: Tools, Methods and Verification Workflow.”

📍 Henk van Ess wrote, “International Research Guide: How to Find Experts and Hidden Documents on Any Topic.”

📍 William Goodhind wrote, “What I Learned from Geolocating 1,000 Satellite Images.”

📍 Benjamin Strick looked at examples of altered satellite and aerial imagery in the latest edition of OSINT Field Notes. He shared a checklist of what to look for, including:

📍 The Data Frontier in Conflict Research, a webinar series on satellite data and armed conflict, is holding a session on Oct. 19 at 9 am ET. The session will be based on the recent Nature paper, “Advancing conflict research and response through satellite-derived data.” Sign up here. (More info here.)

📍 SANS opened the call for presentations for the 2027 US SANS Institute OSINT Summit. It will be held Feb. 1 and 2 in Alexandria, Va.

📍 The Global Disinformation Index studied how Claude and Mistral answer questions on ‘data voids’ filled by Kremlin-aligned sites and concluded that both LLMs “were found to repeat contested or false claims from Kremlin-directed or aligned sites.” Mistral performed particularly poorly, failing to contradict the majority of the false or contested claims.

📍 The Citizen Lab unveiled what it calls “the first-ever public confirmation of zero-click mercenary spyware delivered via targeted ads.” The technique, dubbed AdFection, infected Al Jazeera devices.

📍 Bloomberg published a deep dive into the global scam economy, describing a shift in operations “as they spread beyond Asia’s fortified compounds, become cheaper and easier to run – and harder to kill. Stolen data, fake accounts, fraudulent websites, AI software and laundering services are all bought online, allowing smaller crews to rebuild after takedowns and move when authorities close in.”

📍 Maldita and Science Feedback reported 169 videos (87 on TikTok, 82 on YouTube) in Spanish and French that they assessed as violating the platforms’ climate disinformation policies. They found low enforcement rates after reporting the videos to the platforms; there was a similar lack of action by the Spanish out-of-court dispute settlement body and the French Digital Services Coordinator they appealed to.

📍 Full Fact launched its Public Interest AI Trust Benchmark with a first report looking at how five chatbots replied to 22 questions on a daily basis over a 70-day period.

📍 The Center for Democracy and Technology published a major report on how foundation model companies try to stop their models generating non-consensual intimate imagery. It concludes that “at present, the interventions employed by foundation model companies focus heavily on the generation of sexual content, rather than on AIG-NCII. NCII is specifically the replication of someone’s likeness in intimate contexts without consent. Yet, likeness was rarely the focus of intervention.”

Want more studies on digital deception? Paid subscribers get access to our Academic Library with 55 categorized and summarized studies:

Please upgrade to access all of our content, including our how-to guides and Academic Library, and to our live monthly workshops.