Briefing: The AI image gotcha game

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

Craig shared the video, transcript, and AI-generated notes from our October workshop. He showed how to set up free alerts that catch new mentions of people, companies, and phrases, and how to track changes to webpages.

Alexios reviewed 1,669 Community Notes about US politics posted on Facebook, Instagram, and Threads in the month before the midterm elections. He found a lot of noise — including political insults and self-promoting bios — but also some signs that crowdsourced fact-checking can provide value. The full dataset is available to Indicator members at the bottom of the article.

Tom shipped Spotlight 2.0, a new version of the Indicator Lab’s agentic investigation system. It uses a swarm of AI agents to assist with OSINT research and can run locally on your machine.

The US midterms offered a case study in the lack of AI literacy among professionals who should know better.

On election night, Free Press reporter Olivia Reingold shared a photo she said was from a Democratic gathering in Michigan:

A style writer at Vanity Fair took a closer look at the image and noticed that several details looked off. They wrote:

A fellow Vanity Fair writer tweeted, “AI image detectors confirmed that yes, Olivia Reingold used an AI-generated image to make her point, without a drop of disclosure. Wild!!!!”

She added: “Thank you @pangram and @GoogleDeepMind SynthID for your attention to this matter.”

Pangram’s new image detector (which we also highlight below) and the SynthID detector offered these results for the image:

That was enough for critics to accuse Reingold of presenting an AI-generated image as real, a significant ethical breach.

But the truth was more nuanced — and a good illustration of the lack of AI literacy.

Reingold explained that she “used ChatGPT to brighten and sharpen an image of a woman in Valentino shoes and a Chanel bag. After processing, it still shows these same things: a woman wearing Valentino shoes and a Chanel bag.”

Not exactly. Look at Reingold’s original and the edited version:

ChatGPT did more than just “brighten and sharpen” the image. It distorted it. Simply put, a journalist should not use ChatGPT as an image editing tool. As Aric Toler of the New York Times pointed out:

It’s a problem that Reingold didn’t know this and a problem that she didn’t disclose her use of ChatGPT in the original tweet.

But along with Reingold’s ethical lapse, others failed to account for the nuances (and limits) of AI detection tools.

The SynthID detector result said that “This media was made or edited with OpenAI.” Pangram flagged the image as “AI-generated or edited with AI tools.” Many people assume such verdicts mean that an image was fully created with AI. That’s not necessarily the case.

The image Reingold shared wasn’t a camera original. But is it accurate to label it “AI-generated,” as some did? Does calling it “edited with AI” undersell how much it changed? (Then there’s the fact that the latest phone cameras also use AI, which is a whole other issue...)

Expect more confusion and more examples like this. People will reach for AI tools when they shouldn’t, and others will engage in the synthetic media gotcha game without accounting for the nuance and limits of detection tools. — Craig

📍 X said it will inform users through its “Under the Hood” tool when one or more of their posts have been labeled in a way that reduces their visibility (e.g., for spam, violence, or sexual content).

📍 Cyberattacks this week targeted supermarket refrigerators in at least nine EU countries, as well as French blood banks.

📍 Pangram, known for its AI text detector, launched an image detector. The company says it was 99.96% accurate on an evaluation set of 1,200 synthetic images. Aggressive cropping still fools it, though.

📍 The Washington Post reported that an obscure AI-generated website called the Ballot Brief was frequently cited by ChatGPT and Microsoft Copilot as a source for US election information, despite containing several mistakes.

📍 Microsoft filed a patent application in 2025 “for an AI product that would be able to transform live video in real time to alter speakers’ words and faces,” according to Futurism. The filing was made public last week. A USPTO examiner rejected the claim, but Microsoft can respond.

📍 Joseph Cox of 404 Media interviewed Yuriy Nazarenko, the creator of fake ID-generating website OnlyFake. After more than a year in US custody, Nazarenko was deported to Ukraine before fleeing to Thailand, where he now lives.

Lots of OSINT tools disappear, get shut down, or stop working. It’s rare to see one come back from the dead.

I recently noticed SpyOnWeb was back online after a hiatus. It used to be one of my go-to tools to search for websites that share the same Google Analytics or AdSense ID. (If you’re not familiar with the technique, check out my guide to connecting websites together using OSINT tools and techniques. Or watch the workshop video.)

SpyOnWeb went offline around the fall of 2024. But it’s back on a new domain — spyonweb.net instead of .com — and offers many of the original features, plus new ones. You can search for domains that share the same AdSense/Analytics ID, research nameservers and historic IPs, track three domains or IPs for free, and more.

David Carrero Fernández-Baillo, the owner, told me in an email that the new site “isn't a continuation of the original spyonweb.com, and it's not the same owner. I was a regular user of the original service, and when the .com stopped working I missed having the tool, so I decided to build a new version.”

The service scans domains and extracts data including Google IDs (UA-, G-, GTM, AW-, pub-), IP addresses, nameservers, the SSL certificate, and more.

He said the data goes back to September 2025. (Services like BuiltWith and DNSlytics will likely have better historical data.)

Carrero said that if you search a domain that’s not in the database, the system queues it for a future run. “So check again a little later,” he said.

Along with three free site monitors, you can run 10 queries per hour without being logged in. — Craig

📍 CLOCKEDF12 is a new browser extension from Nicole Hurey that “automatically examines the technical information available through Chrome DevTools. It identifies and organizes network activity, API responses, JavaScript, page data, browser records, and other technical evidence.” Available for Chrome, Edge, and as a direct download.

📍 FaceSeek is a facial search tool. (via Saad Sarraj)

📍 Archivarix Retrace is a reverse image search tool. It was included in this week’s edition of the OSINT Tools Radar, a weekly Indicator email that highlights new and updated tools in the OSINT Navigator database. (Upgrade to the Indicator Lab to receive it, along with access to a suite of tools and support to help you use AI to supercharge your OSINT and digital investigative work.)

📍 Nico “Dutch OSINT guy” Dekens wrote, “Mapping an Individual X Account Footprint with Grok and Advanced Search.”

📍 Boudjenane Soufiane wrote, “Searching smarter: a guide to name-based OSINT.”

📍 Lucía Gómez wrote, “Building OSINT evidence databases: Lessons from the Amnesty Digital Investigations Summit.”

📍 Rae Baker wrote, “What Is LRIT and Why Haven’t You Heard of It?”

📍 Eliot Higgins of Bellingcat is giving a paid webinar on Nov. 9, “Verification, Disordered Discourse, and the Repair of Democratic Reality Webinar.” Info and registration.

📍 The Global Investigative Journalism Network and International Consortium of Investigative Journalists are hosting a free webinar on Nov. 24, “How to Investigate Cryptocurrency.” Info and registration.

📍 A new preprint by Thomas Renault, Mohsen Mosleh, and David Rand found that AI-generated Community Notes on X “are submitted more quickly, are longer, and cite more and higher-quality sources than human-written notes.” They were also rated helpful at a similar rate and were linked to similar drops in reposts once published. AI-written notes were twice as likely as human ones to cite fact-checking websites.

📍 Wired and AI Forensics found that a network of more than 400 innocent-looking AI apps and websites was secretly offering tools to create sexually abusive deepfakes. Wired wrote that the apps “appear to be run by Chinese developers and a Hong Kong-registered entity called KeLing Technology Limited.”

📍 Bill Adair’s Grokipedia page used to confuse him with a mid-century baseball player. Still, Adair argues the AI encyclopedia has some strengths that its rival Wikipedia could learn from.

📍 Speaking of Wikipedia, and specifically its notability policy: Wired looked at the “tragedy” of having your page deleted for being insufficiently significant. And a marketing consultant considered whether the policy makes the platform biased toward incumbents in US elections.

📍 Jeff Kosseff, of “The Twenty-Six Words That Created the Internet” fame, makes the case that the court fight between Elon Musk and Minnesota over the state’s law imposing fines on providers of AI nudifiers may define how liable AI companies can be held for harmful content more broadly.

📍 The Wesleyan Media Project concluded last week that spending on AI ads during this US election cycle “surpassed $187 million,” with just over 90% of spend coming from Republicans. WMP notes that “our calculation of spending is almost surely lower than the actual level of spending given that many ads do not contain disclaimers and some instances of AI use are difficult to determine with the human eye alone.”

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