Briefing: New tools for AI image verification

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

Hey everyone, Indicator turned 1 on Tuesday! Thanks to all 12,359 of you for being here with us. We have big plans for year two, so stay tuned…

Craig wrote a guide to using Skills in Chrome for OSINT. Skills are custom AI prompts that you can run directly in Chrome against a webpage or multiple tabs. We built and shared three Skills that you can install: a social media profile scanner, entity extractor, and image analyzer.

Next week’s members workshop will be a hands-on demo of how to create and run Skills, as well as other simple, repeatable AI prompts that can help with investigations. (Plus: scroll to the Tools & Tips section for an update to the image Skill).

Our summer fellow Ethan McCarthy wrote about new US legal requirements for platforms to remove nonconsensual deepfake nudes, and reviewed the removal flows at 16 companies. He found that they “vary significantly across platforms and aren’t always easy to find.”

Finally, we published the first edition of our ethics policies. This formalizes the rules we’ve been abiding by in terms of AI use, corrections, and more. If you have any questions or comments, drop us an email at [email protected].

📍 Big week for undisclosed AI use in books. First, several winners of the Commonwealth Foundation Short Story Prize were suspected to have used AI in their submissions. (Though relying on AI detectors to make such arguments is dicey.) Then New York Times reporter Ben Mullin found AI-fabricated quotes in Stephen Rosenbaum’s book The Future of Truth. One misquoted source was Kara Swisher, who told Mullin that the quote made her “sound like I have a stick up my butt, according to ChatGPT.”

📍 Full Fact policy lead Phoebe Arnold told British MPs that platforms should have a statutory obligation to provide effective media literacy. (The British fact-checking org previously published a full briefing on protecting elections from misinformation.)

📍 NewsGuard found four videos and six articles impersonating Western media to falsely claim that hantavirus is spreading rapidly in France. The operation resembles previous disinformation campaigns from Russian-affiliated actors.

📍 Thomas Dietterich, chair of the computer science section of arXiv, said the open-access academic repository has clarified its penalties for submissions containing incontrovertible evidence of unchecked AI use, such as "hallucinated references, meta-comments from the LLM." The penalty is a one-year ban from arXiv, after which authors must have submissions accepted at a peer-reviewed venue before resubmitting. This comes as academia increasingly struggles with synthetic submissions.

📍 British regulator Ofcom strengthened its code of conduct on illegal content by recommending that platforms use hash matching to detect explicit deepfakes.

📍 The US Federal Trade Commission sent a letter to twelve operators of AI nudifiers asking them to add a form so that people can request removal of nonconsensual content. The FTC threatened them with fines for noncompliance. (Given that most nudifiers don’t host the content they generate, it’s unclear if this will have much of an effect. — Alexios)

📍 Amnesty International accused the administration of Indonesian President Prabowo of using false narratives to cast civil society leaders and activists as foreign agents.

Google and OpenAI this week announced new measures that will help with synthetic media detection and verification. SynthID, Google’s watermarking technology, is at the core of the efforts. It embeds invisible watermarks in synthetic media, enabling easier detection.

Here’s what you need to know:

It’s good news that big players like Google and OpenAI are working to standardize content provenance. Of course, the news comes after the information environment is inundated with synthetic images and video, and when platforms' promises to label synthetic content have fallen short. We should also be aware of the limitations of watermarking and metadata. As OpenAI said in its announcement, “No single provenance technique is enough on its own.”

But this helps investigators and the public.

Note: the Image Analysis Chrome Skill that I created and released earlier this week is now more powerful. When I created the skill, you couldn’t use Gemini in Chrome to check for SynthID watermarks. Now you can. I updated the Skill prompt to check images for SynthID and to make users aware of OpenAI’s verification tool. Members can grab the updated prompt from our recently-launched AI Prompts page. (If you’re not a member, you can upgrade here.)

I’ll also update the prompt when it’s possible to check for C2PA metadata via Gemini in Chrome. —Craig

📍 Andy Lehren and Kuek Ser Kuang Keng published, “Guide to Mapping Analysis Using QGIS.”

📍 Skip Schiphorst of I-Intelligence shared a list of resources to learn about international naming conventions, transliteration, and international identity structures.

📍 Hana Lee Goldin wrote, “YouTube is a Research Library. Here's How to Search It Like One.”

📍 The OSINT Newsletter wrote, “Call Data: OSINT on Phone Numbers.”

📍 Applied Awareness wrote, “Finding Public Files That Probably Shouldn’t Be Public.”

📍 D4rk_Intel wrote, “How To Investigate A Person Of Interest In 2026.”

📍 Researchers at the Rochester Institute of Technology found that AI models remain vulnerable to repeating and expanding on falsehoods when nudged by their interlocutor. The lead author writes that “this tactic isn’t a hypothetical. When people talk, conversational pressure can emerge naturally. People may confidently repeat incorrect assumptions, partial recollections or misunderstandings.”

📍 Rest of World interviewed several Filipino virtual assistants and agencies who described the widespread use of AI to create thinkfluencer posts for clients on LinkedIn.

📍 404 Media's Samantha Cole wrote a long piece about how a Pennsylvania school mishandled a case of deepfake image-based sexual abuse. A student used AI to create explicit images of four of his ninth-grade classmates.

📍The marvelously named paper “Why Slop Matters” (h/t Reuters Institute) argued that “AI Slop will be an increasingly prolific and impactful part of our creative, information, and cultural economies; we should take it seriously as an object of study in its own right.”

📍 In a preprint, a group of researchers from George Washington University and the University of Maryland argue that “architectural” constraints in platforms have a bigger impact than algorithmic recommendations on the spread of information. Their findings stem from an agent-based simulation of four different types of architectures (e.g. a Reddit-like “tree” or TikTok-like “graph”) and two types of recommendation algorithms (reverse chronological, and popularity-based).

📍 Cornell’s Gordon Pennycook and David Rand published a voicy review of academic studies on the psychology of misinformation. Pennycook wrote on Bluesky that “I am concerned that the research has become increasingly siloed. People with substantive disagreements are disengaging with each other to avoid disputes.”

Want more studies on digital deception? Paid subscribers get access to our Academic Library with 75 categorized and summarized studies:

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