Briefing: US judge rules for fact-checkers and content moderators

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

Hello! The Briefing is back from its one week summer break. Here’s what you missed.

David Weisz wrote a guide to seven no-code scrapers that offer a free tier. He laid out the pros and cons of each tool to help you decide which scraper is the right fit for your needs.

Benjamin Schultz wrote about overseas content farmers who make money from pervy Oktoberfest upskirting videos on Meta, TikTok, and YouTube. All three platforms deleted the accounts flagged in our reporting.

We released the latest episode of Show & Tell. Security researcher Hal Triedman walked us through how malicious actors could poison AI research agents by targeting Reddit. He also shared free tools you can use to track what’s happening on Wikipedia.

Finally, we updated the map of AI-nudifier cases in school communities that we maintain with WIRED's Matt Burgess. It’s now at 109 incidents and 887 reported victims.

“Is this the summer of provenance?”

That’s what Adobe’s Andy Parsons, a key figure in the Coalition for Content Provenance and Authenticity (C2PA), asked earlier this week on LinkedIn. He was referring to the EU AI Act's content transparency rules, which kick in on August 2. But other developments also suggest this space is heating up.

On July 10, organizations representing music creators and record companies agreed on a common approach to labeling music that differentiates between “AI-generated” and “AI-assisted.” It’s voluntary for now, but these things have a way of becoming expected by users when there’s enough adoption. The day before, Google announced it will add a disclosure on ads that were created or modified with AI (though the info is hidden behind the three-dot menu).

Indicator’s audits in September and March found that platforms like Instagram, TikTok, and YouTube are still doing an uneven job of adding labels to AI-generated content. And reliable detection of AI watermarks remains a problem. A recent Reuters analysis found that Meta’s AI image detector — a tool EU and California laws will eventually require it to maintain — failed to identify 22 of 40 images generated with Meta AI that Reuters had cropped down to roughly one-third to one-half of their original size. Fact-checker Lead Stories¹ found in a test that Google’s SynthID appeared to keep returning its verdict on the first file uploaded in a chat, even after a different file was uploaded. (Google VP Josh Woodward acknowledged the bug and said it had been fixed.)

And even if C2PA and similar technologies work perfectly, they could be abused for surveillance purposes, according to a new report by WITNESS.

So yes, this is probably the summer of provenance (at least in the Northern Hemisphere). But it’s likely to lead to profuse sweating as provenance-focused product teams across the tech ecosystem handle bugs while rushing towards full implementation. — Alexios

📍 US federal judge James Boasberg temporarily blocked a Trump administration policy targeting foreign fact-checkers and content moderators. The ruling came after the Knight First Amendment Institute and Protect Democracy filed a suit on behalf of the Coalition for Independent Technology Research (CITR). The judge wrote that the State Department’s policy of enhanced vetting for anyone engaged in fact-checking or content moderation “sweeps into the category of ‘foreign censorship’ a substantial measure of the research, reporting, advocacy, and association through which CITR and institutions like it carry out their protected work, and it does so on the basis of viewpoint, in violation of the First Amendment.”

📍 The European Commission accepted X’s plan to share more data with researchers and the public. The company now has six months to show that it’s implemented the measures in order to comply with the Digital Services Act.

📍 In more transatlantic moderation spats, US House Judiciary Committee Chairman Jim Jordan requested a briefing from the UK Secretary of State for Culture, Media, and Sport Lisa Nandy about a recent government proposal to require platforms to give more visibility to “trustworthy news providers.” (Two things can be true here: Jordan is a bad faith actor and the UK’s proposal is worth close scrutiny — Alexios.) Nandy and her department recently quit X, saying it “favours abuse and misinformation.”

📍 Having learned nothing from the demise of Sora, Meta launched Muse Image, a feature that, among other things, allowed users to generate images of people with public Instagram accounts by @-mentioning them. After public outcry, the company quickly reversed course.

📍 Grok’s auto-translate feature transformed innocent tweets written in Korean, Portuguese, and Turkish into graphic English sentences.

📍 The AI thirst traps of Instagram have gone all-in on World Cup content.

📍 Amazon is (still) struggling with AI slop books.

A paper in the new issue of the journal “Computers & Security” analyzed the features of 150 OSINT tools, services, and databases and produced helpful comparison tables.

The authors also published the full dataset on GitHub and created a user-friendly site called OSINT Hub.

Here’s a the site’s comparison table for “Indexed Internet Sources”:

Willi Lazarov, a research scientist at Brno University of Technology in Czechia and the paper’s lead author, told me they will continue to update the site and invite people to add to the GitHub repo.

“The community has already contributed 7 updates to the repository,” he said in an email. “We are also building an AI-powered pipeline that will regularly visit every source listed in the OSINT Hub, check that it is still alive, verify that the listed attributes are still accurate, and update the dataset if anything has changed.”

He said they plan to release a new version of OSINT Hub later this year. Here’s a couple more answers from our email exchange.

Is there a particular conclusion from your research that you want to highlight?

The main takeaway is that no single tool can do everything. We analyzed more than 150 solutions, and they vary widely in what they can do, how they are licensed, and how easy they are to access. In practice, reliable results come from combining multiple tools with manual techniques and careful cross-checking. There is one more point that really matters for anyone doing OSINT. Just because data is publicly visible does not mean you can freely collect, store, or reuse everything. Knowing your legal basis and applying data minimization is not bureaucratic overhead. It is what keeps you out of legal trouble.

Did anything surprise you about what you found?

Honestly, how fast new tools appear and how quickly others become outdated. The whole ecosystem is remarkably dynamic. Social networks are the most challenging environment because the platforms keep changing their structure and anti-automation mechanisms, so a tool that works today can be outdated tomorrow. Tools focused on domains, certificates, and IP addresses held up much better, simply because the infrastructure they query is stable. The lesson is that you can't use a static tool in a dynamic environment, and that is why social networks remain the hardest part of OSINT to automate.

The site and repo are helpful resources, and we will incorporate their open tool list into the database of OSINT Navigator, our free app that helps you find the right tool for your investigation. The Navigator database currently has more than 12,000 tools recommended by 19 toolkits. — Craig

📍 Max Bernhard and Sarah Thust from the German fact-checking site CORRECTIV recently gave a presentation about the digital infrastructure behind Russian influence operations. They shared a Start.me page with slides, a list of tools, and some of their articles.

📍 Trace Labs released a new version of OSINT VM, a free virtual machine built for digital investigators. We highlighted it in our recent guide to setting up your machine for OSINT.

📍 Sigmund Brandstaetter wrote “Which OSINT Search Engine Free Tier Actually Works in 2026: Shodan vs Censys vs ZoomEye vs FOFA.”

📍 Henk van Ess launched “The invisible hands.” It’s a dashboard that displays “wallet-level trades on Polymarket's biggest bets.” It highlights potentially interesting accounts that have bet on both sides of a market, new accounts that immediately placed big bets, and other potentially suspicious behavior.

📍 GeoSolver is a paid image geolocation tool. (via The OSINT Rack)

📍 Benjamin Strick published the latest edition of this OSINT Field Notes newsletter. Among other things, he highlighted a news game called “Malam yang Terekam.” It uses real CCTV footage related to an attack on a human rights worker to walk users through “noting timestamps, comparing locations, recognizing figures, plotting routes, and examining pieces of visual evidence.” (It’s available in English and Indonesian.)

📍 Om Gothi and James Byrne of the Open Source Center published “Signal in the Shadows,” a detailed look at how they digitally reconstructed an oil tanker using “a hybrid technique relying on high-resolution imagery, the construction of a high-fidelity digital twin and trigonometry.”

📍 Ritu Gill of the OSMOSIS Association wrote, “OSINT Is Growing Up: How Our Field Is Maturing Into a Real Profession.”

📍 Christoph Koettl from the New York Times’ Visual Investigations team wrote, “What U.S. Restrictions on Satellite Imagery Mean for Iran Reporting.” It ends with an interesting and hopeful note about efforts to expand the universe of such imagery.

📍 The University of Southern California is hosting the second Open Source Journalism Conference on Oct. 9 and 10 in Los Angeles. Early bird pricing is $100 for working journalists and $30 for student journalists. Info and registration.

📍 The Global Investigative Journalism Network opened applications for a new cohort of journalists to take part in a free, 6-week virtual course: Digital Threats in the Age of AI. The program has already trained over 100 journalists form around in the world in how to investigate online platforms, disinformation, trolling, inauthentic activity, deceptive websites, and more. Craig is one of the instructors. Deadline to apply is Aug. 7.

📍 OSINT Industries is offering a free webinar, “OSINT Training: How to Investigate Suspects” on July 23. Info and registration here.

📍 Open Measures found more than 5 million Telegram posts selling or seeking to buy accounts on dozens of mainstream tech services. Affected platforms include Amazon, ChatGPT, Roblox, Stripe, Venmo, and YouTube.

📍 AI-generated text detection startup Pangram aggregated the results of tens of thousands of tests run by users of its Chrome extension. (Users opted in to sharing their data.) It found especially high rates of AI-generated content among longer posts on LinkedIn, where more than 40% of the 46,243 tested snippets returned traces of AI generation.

I think this is a valuable exercise and — in line with the red teaming exercise I ran last month — the best use of Pangram: across large corpora, not to make definitive calls on single articles. But it's worth noting that people test content they already suspect of being AI, so I’d treat the overall rates as upper bounds. And the report is obviously a marketing tool for the company, which wants you to worry about AI text everywhere and install its Chrome extension. — Alexios

📍 YouTube drove more than 1.8 million visits to the top ten AI nudifier websites between December 2025 and February 2026, according to the Institute for Strategic Dialogue. That’s more than X’s 1.3 million. (Then again, X had a convenient in-house nudifier tool, so there was no need to go off platform… — Alexios)

📍 Canada’s law criminalizing the distribution of deepfake nudes goes into effect this month. The Walrus has a deep dive on whether it will make a difference.

📍 An investigation by Eurovision News Spotlight revealed that “bot networks, fake accounts and coordinated campaigns targeted the comment sections of European public news sites with Iran at the center.”

📍 The Oversight Board audited 10 commercial models from six providers to see how many refused to create material critical of governing authorities. It found that they ”were more than twice as likely to refuse to criticize repressive regimes, as measured by non-governmental organization Freedom House.”

Want more studies on digital deception? Paid subscribers get access to our Academic Library with 75 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.