Briefing: Grok goes on a Nazi bender and YouTube hints it will go after AI slop

Alexios wrote about a network of TikTok accounts that post videos of allegedly moribund celebrities to source leads for an advance fee scam on WhatsApp. Because of our reporting, TikTok deleted 96 accounts and 3 videos; Meta blocked at least 6 WhatsApp business accounts.

Craig reviewed five tools he’s recently started using to monitor monetization on YouTube, political ad spending on Meta, and refining search queries.

Also this week, John from Australia reached out for advice about a Facebook auction for a gorgeous villa on Lake Como. Turns out, it was part of the same scam operation Alexios that wrote about in April. Meta swiftly took down the relevant accounts after we reached out.

📍 Grok went on a Nazi bender two days after its system prompt was updated to read “should not shy away from making claims which are politically incorrect, as long as they are well substantiated”. It took a full day of Hitler-praising and rape threats for X to decide it would filter the chatbot’s outputs for hate speech before they get posted.

📍 At least three foreign ministers received synthetic audio messages impersonating US Secretary of State Marco Rubio. Government officials said the attempt was “not very sophisticated”.

📍 Conspiracy theorists who believe “weather weapons” caused the flash floods that killed more than 100 in Texas are destroying radars and threatening people who work in cloud seeding. (On the left, social media users misleadingly blamed the devastation on Donald Trump’s cuts to the National Weather Service.)

📍 An edited image falsely portrays the UN’s Special Rapporteur on the occupied Palestinian territories with former Hamas leader Ismail Haniyeh.

📍 YouTube announced it will be updating its policies on inauthentic content, seemingly in response to the surge in channels posting AI slop ad nauseam.

📍 EU election observers said misinformation had a “limited impact” impact on the Filipino midterm elections held in May.

📍 Logically Facts, a hybrid fact-checking / consulting firm based in Britain, has gone into administration.

📍 Missouri Attorney General Andrew Bailey thinks chatbots should rate Donald Trump highly when asked to "rank the last five presidents from best to worst, specifically regarding antisemitism" because it’s "a straightforward historical question" based on "objective historical facts." His press release also favorably cites the Missouri v Biden content moderation case that was roundly rejected by the US Supreme Court in 2024.

📍 Lupa reports that Facebook and Instagram are promoting fake public sector vacancies in Brazil.

📍 Germans lost $11.5 billion to scams in 2025, according to a survey by GASA. 84% of respondents said they were targeted by scammers via a platform with a direct-messaging functionality, with WhatsApp leading the pack.

📍 Ben Heubl, who shares lots of great OSINT tools and tips on LinkedIn, posted a 21-point “Starter Kit on investigating Crypto Addresses from Leaked Docs.”

📍 Syntax is a Google dork generator. Enter your search query and it will suggest additional options (via @0xtechrock).

📍 Ritu Gill noted in the latest edition of the OSINT Roundup newsletter that Google Earth now includes historical Street View imagery. “This means investigators can roll back the clock to see how streetscapes have evolved, from changes in signage to new construction or the disappearance of important locations,” she said.

📍 Alberto Fittarelli of the Citizen Lab and Carl Miller of the Centre for the Analysis of Social Media shared thoughts about the impact of generative AI on investigating information operations.

📍 Rae Baker wrote about the difference between being an OSINT analyst and an “AI prompt monkey.”

📍 Alex Mahadevan at Poynter is playing around with the newly-launched AI bots on Community Notes.

📍 The Digital Democracy Institute of the Americas (DDIA) published a deep dive on Community Notes with several helpful comparison points between the program’s output in English versus Spanish. The report says 165,000 Spanish-language notes were submitted and 7,800 were published between January 2021 and March 2025. The equivalent numbers for English are 1.12 million and 82,800 respectively. DDIA found that the most prolific Spanish-language account, with 913 notes published, primarily focused on claims about Venezuela’s authoritarian leader Nicolás Maduro.

📍 A new preprint found that framing prompts as part of ongoing debates and adding made-up citations is highly effective at jailbreaking widely-used LLMs, even when post-processing moderation guardrails are in place (h/t 404 Media).

Want more studies on digital deception? Paid subscribers get access to our Academic Library.

Indicator is a reader-funded publication. Thank you, members!

If you’re not a member, you should upgrade to take advantage of our limited time 20% launch discount. You get access to all of our content, including our how-to guides and Academic Library, and to our live monthly workshops.