We publish original reporting, in-depth investigations, and practical tutorials on open-source intelligence (OSINT) tools and techniques. Our expert research equips you with the knowledge and skills to navigate a chaotic digital landscape filled with scams, search engine and social media manipulation, disinformation, trolling, mobile app abuse, spyware, AI slop and more.
2026 is shaping up to be a significant year for investigators trying to identify who really owns a company.
Changes in the EU, UK, and British overseas territories mean beneficial ownership data that was previously locked away is slowly becoming accessible.
But the reality is complicated: different countries are at different stages, fees can be prohibitive, and some registries are barely functional.
My new guide for Indicator breaks down what's actually changing, which EU countries are worth starting with for legitimate interest access, what's happening with the BVI and Cayman Islands, and where the US stands.
It's for journalists, researchers, OSINT analysts, and anyone who investigates corporate ownership. Thanks to Stephen Abbott Pugh and to the experts at The Organized Crime and Corruption Reporting Project, Transparency International, Data and Research Center, and OpenSanctions who spoke to me.
Link in comments — free to read for Indicator members.
Today on Indicator: I uncovered a coordinated network of AI "models" repeatedly promoting the gambling website 1win on Instagram.
This branding was what ultimately gave them away. Despite a potentially unlimited range of prompts, the accounts often posted videos featuring the 1win logo that were virtually indistinguishable from each other.
Instagram deleted 21 of the 23 accounts I sent over. They had ~2 million followers in total.
If money exchanged hands, legal experts told me these posts would violate advertising regulations in the United States.
I couldn't prove such a payment took place. If it did, it is yet another way in which the AI thirst traps of Instagram are diversifying into new revenue streams.
A more intriguing hypothesis was suggested to me by the ever-wise Craig. What if the accounts included 1win on their 'clothes' to evoke the real adult content creators who *do* have partnerships with the website?
Placing the logo of a restricted gambling site would have therefore acted as a form of inverse AI label, a marker of humanity. It's an entirely speculative conclusion. But ridiculous as it feels to write, it cannot be entirely dismissed.
Check out the story: https://lnkd.in/ewmtuaJS
🍴 Au menu de notre dernière veille : le feedback émotionnel de DER STANDARD, le tracker de rumeurs d’Indicator, la nouvelle navigation temporelle que nous avons conçue pour Basta!, et bien plus encore !
C'est par ici ⬇️
Feeling lucky? When it comes to social media platforms labelling AI generated content, it's still closer to a slot machine than stable digital infrastructure.
Alexios Mantzarlis just published his second audit on how reliably social media platforms label AI generated content in his must follow newsletter Indicator.
This includes content featuring a variety of provenance signals, including IPTC metadata and Coalition for Content Provenance and Authenticity (C2PA) Content Credentials, as well as proprietary watermarking such as Google's SynthID.
Despite Alexios verifying the presence of these signals in the media he tested, the surfacing of them as AI generated content labels was wildly inconsistent (see full results at the link below)
Like the last investigation, factors including which AI tool was used to generate the content, the content format (image or video), and how the media was uploaded to the platform (desktop or mobile) had a significant impact, but in ways that were not always obvious or explainable.
All the platforms Alexios tested have policies for labelling AI generated content, but none performed reliably, with LinkedIn and Pinterest best overall at 67% of images labelled, while Instagram only managed 14%.
So what's the message to take from these findings?
As i said in my comments for the investigation, “platforms may be making the right kinds of noises, but the unified front we need to see for the entire media lifecycle is just not there yet.”
I'm not looking to trivialise the understandable difficulties in achieving this unified front; we're talking about an incredibly complex patchwork of organisations, pipelines, and media formats.
That said, with legislation from the likes of India, South Korea, and California mandating the labelling of AI generated content, don't expect much sympathy from these governments when it comes for enforcement, regardless of whether their demands are currently feasible or realistic.
I've not been shy about critiquing approaches to content authenticity that focus on labelling AI noise rather than securing authentic signal, but given the current beleaguered state of the information landscape, it is better than nothing.
Yet ultimately, if platforms don't reliably surface the provenance information in an accessible format, it will be just that for the vast majority of users.
Be sure to read the full investigation below and subscribe to Indicator. I cannot recommend it highly enough for those looking to understand the modern information ecosystem- both Alexios Mantzarlis and Craig Silverman are killing it!
https://lnkd.in/eJUiPjZz
Alexios Mantzarlis just issued a new study looking at the current state of AI signalling in social media platforms.
Jacobo Castellanos and I had the joy of speaking with him, alongside other amazing Specialists, about what some of the results meant for some of the efforts trying to implement downstream transparency in AI generated or manipulated content.
https://lnkd.in/diNGJCHd
IPTC Managing Director Brendan Quinn was quoted in an Indicator study looking at the current state of AI signalling in social media platforms: https://lnkd.in/dF3Qgsj6
On the creation side, OpenAI and Google Gemini declare AI-generated content using IPTC's Digital Source Type vocabulary embedded in a C2PA manifest. (Unfortunately they each use different versions of the C2PA spec, so results are not consistent across all social media platforms, even those that read C2PA metadata.)
Meta AI uses the same vocabulary, embedded in the Digital Source type property in "regular" IPTC embedded photo metadata.
We would recommend that all AI engines uses both techniques, to give their content the greatest chance of being surfaced by all platforms.
Here's our take on the situation: https://lnkd.in/d4cPyRq2
"AI Label Slots" shows us that AI labeling is a haphazard mess, due to the failure of tech companies to work together. The clock is ticking on both California and EU laws that will come into force in August 2026 and January 2027 that will require these companies to do much better. Those laws are the California AI Transparency Acts of 2024 and '25 (SB 942 & AB 852) and the EU AI Act (Article 50) and Digital Services Act (2024 elections guidance).
Thank you so much to Alexios Mantzarlis from Indicator for this rigorous and beautiful study! The slot machine is free, but I highly recommend a paid sub to see the full report! Link in comments.
From the report:
Headline: AI labeling is still very much a work in progress
Indicator’s latest audit of three generators and five social media platforms revealed multiple gaps
AI content is everywhere, and it’s ever harder to spot.
Tech companies have promised to help limit the damage to online truth-seeking by tagging AI content with machine-readable signals that can be used to label the material if it appears on social media feeds. Billions of such labels have already been applied.
This tagging-and-labeling infrastructure has multiple failure points, however, according to a new Indicator audit.
Over the past few weeks, I created more than 200 AI-generated images and videos using Google, Meta, and OpenAI tools. I then posted them on Instagram, LinkedIn, Pinterest, TikTok, and YouTube; each of these platforms has promised to label synthetic content.
The results were inconsistent and often underwhelming. Even the best in class – LinkedIn and Pinterest – only labeled 67% of the AI content I posted. YouTube managed roughly 50% and TikTok about one third. Instagram did worst of all, labeling just 15 of the 105 synthetic images.
Whether a synthetic image or video got labeled came down to a combination of how the content was created, what device was used to upload it, and which platform it was posted on. While these failure modes were not random, the process ended up feeling like a slot machine....
David Evan Harris, who helped write California’s AI Transparency Act [of 2025] and sits on the EU Working Group on Transparency of AI-Generated Content, told Indicator that “the results of this study are confirmation that voluntary commitments by the tech companies in the study cannot be taken seriously.” Harris, a Chancellor’s Public Scholar at UC Berkeley, says consistent labeling “inherently requires some degree of coordination, which [platforms] seem either unwilling or unable to undertake.”
Regulation on AI labels has been approved or is about to go into effect in several places around the world, including California, the European Union, India, South Korea, and Vietnam.
---
#AI#DeepfakesUniversity of California, Berkeley | University of California, Berkeley, Haas School of Business | CITRIS and the Banatao Institute | California Initiative for Technology and Democracy | Brennan Center for Justice
Today on Indicator: The AI content tagging-and-labeling infrastructure is STILL not fully functional, and Meta is a big part of the reason why.
Over the past few weeks, I created more than 200 AI-generated images and videos using Google, Meta, and OpenAI tools. I then posted them on Instagram, LinkedIn, Pinterest, TikTok, and YouTube; each of these platforms has promised to label synthetic content.
The results were inconsistent and often underwhelming. Even the best in class – LinkedIn and Pinterest – only labeled 67% of the AI content I posted. YouTube managed roughly 50% and TikTok about one third. Instagram did worst of all, labeling just 15 of the 105 synthetic images.
Whether a synthetic image or video got labeled came down to a combination of how the content was created, what device was used to upload it, and which platform it was posted on. While these failure modes were not random, the process ended up feeling like a slot machine.
(Overextending the metaphor, I created an interactive that visualizes the audit results as a slot machine. Try it out!)
Tech companies are likely to make further progress as legal obligations kick in. But even then, we’ll still only be at the beginning of the path towards refitting our online spaces for trustworthy communications in the AI age.
Read, share, and subscribe: https://lnkd.in/dtDyST6q
