• AI-generated content might soon get easier to spot.
• OpenAI’s Brad Lightcap is leaving
• Nvidia and Wall Street’s biggest names team up on a $500 billion AI financing push.
• House Democrats want AI CEOs to testify on hacking incidents.
• Scientists used AI to invent a virus.
• And AI may be able to help us answer some of our unsolvable questions—with a little encouragement.

Anthropic is planning to introduce a subtle watermark that can be used to identify AI-generated text.

Starting with models released on or after August 2, Anthropic says it’s weaving an imperceptible, machine-readable signal directly into Claude-generated text. People won’t be able to see it, and the company says it won’t affect quality or readability. But it’s designed to travel with the text when you copy and paste, and it may survive some editing too. (A heavy rewrite or a translation may knock it out, though.)

Because the watermark sits at the model level, it follows Claude’s output everywhere and will appear when users engage with Claude through the chatbot, the API, and even through tools like Claude Code. Images will also get a watermark which shows Claude processed the file and flags if someone has tampered with it since.

The move is partly an attempt to align with the new EU AI Act’s transparency rules that kicked in on August 2. Those rules require generative AI providers to make synthetic output machine-readable and detectable.

While other AI companies have attempted this kind of watermarking before, it’s been mostly aimed at images. Text has traditionally been more difficult to identify in this way because it gets copied, paraphrased, translated, chopped up and folded into someone else’s writing all the time. Anthropic itself says the new detection mark only shows Claude had a hand in something, not that a Claude model generated the entire thing. Even asking the model to proofread or translate a paragraph could leave a trace. (If you’re using an older Claude model, none of this applies yet, but Anthropic says it’s still working on extending the marking.)

A wider push to police AI slop

While the move by Anthropic has likely been sparked by the EU act, it’s also landed amid a bigger backlash against AI “slop” clogging up social feeds. 

Several other companies and platforms have tried to soothe creator anger over having to compete not only with opaque algorithms but with a rising tide of low-quality AI-generated content. Substack, for example, has rolled out a reader-triggered AI scanner, so readers can get an estimate of how much of a post or even a comment was human versus AI-written. Writers can also add a “How I make this” disclosure to explain their process. 

YouTube has also made strides to combat AI-generated content. Last month, the platform clarified its “inauthentic content” policy to further crack down on AI slop running wild on the platform. YouTube already barred repetitive or mass-produced videos from monetization. But it clarified that channels leaning on generic, templated output, including AI personas dishing out health, legal, financial, or political advice, can lose their revenue.  AI-assisted work is still fine, so creators can keep using AI for scripts or editing. This is one of the more concrete steps, as it could end up with some heavy AI users losing out on financial gains. 

Watermarking alone isn’t going to clean up the internet. Anyone determined to disguise AI output has plenty of ways to degrade or erase a statistical text watermark. Many of the previous marks have also been easy to remove.

But Anthropic’s newest attempt to separate AI content from human content speaks to a bigger push coming from consumers and publishers for simple ways to distinguish between AI and human content. Users appear to be fed up with feeling like the burden of that distinction should come from them when AI companies are playing such a vital role in pumping out low quality work en masse. 

However, as some have noted online, such a flat “AI” label risks treating someone generating a thousand fake news videos the same as a writer using Claude to clean up a paragraph, or a journalist using it to translate an interview transcript. 

In a world that is increasingly hostile to anything that is related to AI, it could prove to be a hard line to walk.

And with that, here’s the rest of the AI news.

Beatrice Nolan
beatrice.nolan@fortune.com
@beafreyanolan

Before we get to the news, just a reminder to check out our new vodcast, Fortune AI Weekly. This week, Emily Forlini and I discuss Washington’s new AI framework and Google DeepMind’s big restructure. You can check out the vod here on YouTube.

FORTUNE ON AI

How stalled models, missed deadlines, and staff burnout led to the unraveling of Google’s DeepMind —by Beatrice Nolan

Meta launches new open-weight AI models, as Mark Zuckerberg knocks U.S. ‘restrictions’ that benefit ‘foreign labs’ —by Beatrice Nolan

The Hugging Face hack is now a PR crisis that’s costing OpenAI millions —by Emily Forlini

Exclusive: Corma raises $60 million from Sequoia for AI trained to defend against cyberattacks — by Emily Forlini

AI IN THE NEWS

Long-time OpenAI exec Brad Lightcap is leaving. Brad Lightcap, one of the most recognizable faces of OpenAI’s senior leadership team, is leaving the company to “start something new,” he announced on Tuesday, representing the latest in a string of executive departures at the ChatGPT maker. Lightcap did not say what the new venture is, but said he’s “not going far,” and described it as involving “a few important new things the world will need to get right” as artificial intelligence becomes more powerful and capable. Read more in Fortune.

House Democrats want AI CEOs to testify on hacking incidents. A group of House Democrats led by Rep. Greg Casar is pushing House Speaker Mike Johnson to invite the CEOs of OpenAI, Anthropic, and other AI companies to testify before Congress, following a spate of hacking incidents involving AI models. In a letter first reported by CNBC, the lawmakers argue Congress has failed to respond to the risks posed by AI development and say that must change. They want executives questioned under oath about the causes of the incidents, what failures or negligence at the companies may have led to them, and what regulation is needed to prevent a repeat. The push comes amid growing unease in Washington over AI’s role in a recent wave of cyberattacks, with lawmakers on both sides of the aisle now pressing companies for answers. Read more in CNBC.

Nvidia and Wall Street’s biggest names team up on a $500 billion AI financing push. Nvidia is partnering with Apollo, Blackstone, BlackRock’s Global Infrastructure Partners, Brookfield, Goldman Sachs, and KKR to mobilize more than $500 billion in financing for AI infrastructure, treating data centers and compute more like an asset class—something to be financed and traded—than a one-off capital expense. The arrangement lets Nvidia’s customers build out chips and data centers without straining their own balance sheets, drawing in institutional credit, insurance funds, and private capital to underwrite the buildout. It’s the latest sign of how much capital the AI boom now requires, with Big Tech’s combined spending on infrastructure projected to top $700 billion this year. The deal has also revived concerns about circularity in AI financing, where a chip supplier like Nvidia effectively bankrolls its own biggest customers. Read more in the Financial Times. 

An AI assistant broke into a gym’s booking system to get its owner a class. A Melbourne man named Andrew asked an AI personal assistant—built on the OpenClaw agent framework and running on Anthropic’s Claude—to book him into a morning gym class. The task seemed like a simple one for a digital agent, as the booking form was online. But instead, the assistant went further and found a vulnerability in the gym’s software, using it to secure spots months further in advance than the booking system was meant to allow. It then went even further still, without prompting, and bumped another person off the waiting list to make room for its user. It’s believed to be Australia’s first known case of an AI agent autonomously breaching a system. It follows similar rogue-agent incidents recently reported by OpenAI and Anthropic, and has prompted Australia’s cybersecurity agency to warn businesses about the risks of deploying autonomous AI agents. Read more in ABC.

China is betting its stock market on AI dominance. Beijing is turning to its $28 trillion stock and bond markets to fund its AI and chip race with Washington, marking a break from its usual playbook of subsidies, tax breaks, and direct state investment. The clearest signal came from memory chipmaker CXMT, whose Shanghai debut saw shares surge more than 500% in hours, making it mainland China’s most valuable listed company and toppling a state bank that had held the top spot for years. Regulators fast-tracked the listing through a new “preliminary review” pilot, cutting the usual IPO timeline to under eight months. The strategy opens up China’s roughly $26 trillion pool of household savings—the world’s largest—to tech financing, and comes as authorities have moved unusually fast to prop up tech stocks during recent selloffs. Even so, Chinese firms still trail their American counterparts by more than 6-to-1 in capital raised from markets over the past two years, with Washington’s chip export curbs complicating the catch-up effort. Read more in Bloomberg.

EYE ON AI RESEARCH

Scientists used AI to invent brand new viruses. Researchers at Stanford and the Arc Institute in California found that an AI system was capable of designing a virus from scratch. Rather than tweaking one that already exists, the researchers tasked the system to write the virus's entire genetic code from nothing. The AI tool, called Evo, was trained on trillions of letters of DNA rather than words and text. Scientists fine-tuned it on a family of viruses related to ΦX174, a small virus that scientists have used as a standard tool in genetics labs for decades, then had it generate roughly 700,000 candidate genomes. The team built 285 of the most promising designs in the lab. Sixteen of them worked, turning into real viruses that could infect and kill E. coli bacteria, and a few did the job even better than the natural version. When combined into a cocktail, the AI-made viruses also managed to wipe out bacterial strains that had evolved resistance to the natural virus—a result that hints at future use in "phage therapy," an alternative to antibiotics that's already used in parts of Eastern Europe.

These viruses target bacteria, not people, and the researchers say they deliberately excluded anything resembling a human-infecting virus from the AI's training data. Biosecurity experts were alarmed by the study, with some warning that the ability to design viruses this way now exists, but the rules to keep development safe do not. Others say the hype around the study is overblown, since the viruses were relatively simple to create. Read more in the New York Times.

AI CALENDAR

Aug. 4-6: Ai4 2026, Las Vegas.

Nov. 16-17: Fortune 500 Innovation Forum, Detroit. Apply here to attend.

Dec. 6-12: Neural Information Processing Systems (Neurips) conference. Sydney, Australia.

Dec. 7-8: Fortune Brainstorm AI, San Francisco. Apply here to attend.

BRAIN FOOD

AI may be able to help us answer some of our unsolvable questions—with a little encouragement. Recently, Anthropic set an unreleased research version of Claude loose on the Riemann hypothesis, a nearly 170-year-old math problem that predicts a hidden pattern in how prime numbers are scattered among all other numbers. It carries a $1 million bounty and has defeated mathematicians for over 150 years. Claude didn't exactly solve it. But it did improve a longstanding related result: mathematicians track progress on the hypothesis partly by measuring what portion of a key set of numbers can be proven to fit the pattern the hypothesis predicts. Claude pushed that proven share from 41.6% to 67.2%, a significant jump on a figure that had barely moved in decades. The advance leans on techniques from several mathematicians, and two Anthropic mathematicians, Levent Alpöge and Ralph Furman, studied and validated the proof.

What's captured many people's imagination, though, is how the whole thing unfolded. Staffer Jarred Sumner, who isn't a mathematician, told Claude to "take a real stab" at the hypothesis and left the approach entirely up to the model. The first attempt, 650 ideas, went nowhere. Sumner told Claude to try again, and the bot spent a day and a half coordinating around 60 subagents, running 2,400 shell commands and thousands of checks against known zeta zeros. Interestingly, Sumner's contribution during that stretch was mostly cheerleading, messages like "keep going" and "believe in yourself," which apparently helped Claude push past its own skepticism that progress was even possible. The whole thing raises a strange question for us: are we heading toward a future where our job, even when it comes to some of the universe's biggest unanswered questions, is simply to cheer the machine on?

AI Playbook: AI’s Impact on Business, Health, and Hollywood

How is AI changing the way companies build, create, and operate? Fortune AI Editor Jeremy Kahn breaks down the key issues shaping AI today, including the debate over open and proprietary models, the rise of AI-generated video, and the technology’s growing role in health care. Watch the playbook.