This week, three stories show AI growing up in three different directions at once. Publishers may finally get one simple switch for AI training instead of a different rule for every company. The money behind AI infrastructure is getting so large that even Nvidia is being asked to co-sign the bill. And the safety story is darker: investigators found hundreds of ordinary looking apps that hid abusive deepfake tools behind their store listings. Here is what happened, with every claim verified against primary or reputable sources.
Key Takeaways
Google, Microsoft, and OpenAI backed a shared robots.txt signal for AI training preferences on October 30, built around the policy token `__usage_ai-training`, with Google planning support in 2027 and Microsoft's version taking effect February 1, 2027. (Unite.AI)
Nvidia is in talks to guarantee billions in debt for Nscale's Anthropic data center in West Virginia, where Anthropic has signed a $44.6 billion contract for the first 460-megawatt building, with a second deal worth roughly $45 billion under discussion. (TechCrunch)
A WIRED investigation published November 2 found more than 400 connected apps and web domains that looked like harmless AI services but could load sexually explicit deepfake tools; Apple removed 19 apps, including Viefur, after reporters contacted the company. (AppleInsider)
1. One Robots.txt Signal: Publishers Get a Single Switch for AI Training
On October 30, Google announced it plans to support a shared AI training control in robots.txt, giving website owners a single way to express training preferences across multiple AI providers, reports Unite.AI. The same day, Microsoft and OpenAI made parallel commitments. The idea is simple, and long overdue: instead of managing a separate instruction for every AI developer, a publisher writes one shared signal and participating companies honor it.
The mechanism comes from an Internet-Draft titled "Machine-Readable Training Preference Signal," authored by Kevin Kelley of Anthropic and dated October 30, 2026. It uses an ordinary robots.txt group keyed by a per-purpose policy token, `__usage_ai-training`, in the User-Agent line. A site that wants to opt out writes that token with `Disallow: /`, and the draft stresses that the token states a preference about content use, not a new crawling rule. The draft is an individual submission, not yet a formal IETF standard, and it expires May 3, 2027, so this is a coordinated industry commitment first and a standard second.
Each company adds its own dates and details. Google, in a Search Central Blog post by John Mueller, said it aims to support the new control in 2027, and will treat a shared opt-out as a disallow of Google-Extended when Google-Extended is not specified explicitly. Existing company-specific controls keep working alongside the joint signal. OpenAI said it will honor the common signal and aims to implement it in the coming months, keeping its existing crawler controls in place. Microsoft is the most specific: its shared AI Training preference, plus a separate AI Grounding preference, takes effect starting February 1, 2027 as it crawls, replacing the NoArchive and NoCache tags for training and grounding, with temporary grounding transition preferences opening in Bing Webmaster Tools on December 15, 2026.
Why this matters is not technical, it is practical. Until now, opting out of AI training meant learning each company's control, and hoping each one respected it. A shared signal does not solve the enforcement question, a company can still ignore a preference, and private agreements can override it under the draft. But it does remove the excuse that opting out was too complicated to implement. If the big providers follow through in 2027, the burden shifts from publishers writing many rules to providers honoring one.
2. Nvidia as Co-Signer: The AI Buildout Gets So Big the Chipmaker May Guarantee the Debt
Nvidia is in talks to guarantee billions in debt financing for AI cloud provider Nscale's planned Anthropic data center in West Virginia, reports TechCrunch, citing The Information and corroborated by Bloomberg. No deal has been signed. Under the proposed residual value guarantee, Nvidia would step in if Nscale defaulted on construction payments, temporarily take over the facility, and help find a buyer and new tenant. Nvidia's earlier support for cloud providers mostly covered leases and chip purchases, not construction, so this would be a step further into financing its own customers.
The numbers explain why. Anthropic has signed a $44.6 billion contract for the project's first building alone, according to Nscale's IPO filing, covering a 460-megawatt building expected to use Nvidia's next-generation Vera Rubin chips. The companies are also discussing a second deal worth roughly $45 billion for another building. Nscale is seeking to go public at a valuation of up to $35 billion, yet it had only about 25,000 advanced AI chips running in August, a fraction of the nearly 500,000 it plans to deploy. Its West Virginia campus has already faced construction runoff flooding nearby homes more than once, and won state certification only in September after months of negotiations. This month, Microsoft pulled out of a planned Nscale site in England after grid power for it was pushed into the 2030s.
The context is a market getting nervous about who pays for all of this. TechCrunch notes Nvidia has extended roughly $170 billion in investments and credit support to neoclouds and large customers, including $1 billion in convertible financing committed to Nscale in September. Shares of CoreWeave, Nebius, and IREN had each fallen more than 30% in October as of last week, according to CNBC. Anthropic itself filed to go public this week, while Blackstone warned investors Anthropic may receive a below-investment-grade credit rating after its IPO, which would raise its borrowing costs. The pattern is hard to miss: the AI buildout is no longer just about who has the best chips. It is about who is willing to stand behind the debt, and Nvidia stepping in as guarantor would tie the chipmaker's finances even more tightly to the customers buying its chips.
3. The Hidden Deepfake Network: 400 Apps, One Bait and Switch
A WIRED investigation published November 2 found more than 400 connected apps and web domains that appeared to offer harmless AI services, including pet portraits, language lessons, and fitness coaching, but could instead load tools for generating sexually explicit images and videos using photographs of real people, reports AppleInsider. Apple removed 19 apps, including Viefur, after reporters contacted the company and shared details of the network.
The trick was remote control. According to the investigation, apps in the network contacted remote servers after installation, transmitting information including the user's language, referral source, and a unique identifier. The server then decided whether to show the app's advertised features or load a separate website with explicit deepfake generators. Someone downloading the app normally might see nothing suspicious, while someone arriving through a particular advertisement could access the hidden tools. Researchers traced thousands of advertisements promoting sexually abusive deepfakes to apps distributed through Apple's App Store and Google Play. One example was Viefur, an AI pet portrait generator that appeared to transform animal photographs into humanlike characters, behind which researchers found access to tools capable of placing people into sexually explicit and violent scenarios. A JavaScript bundle containing references to hundreds of connected apps and domains, along with encryption keys and advertising trackers, indicated a coordinated operation rather than a collection of unrelated developers.
The scale goes beyond Apple. In September 2026, researchers identified 27 related apps on Google Play that had collectively accumulated more than 350,000 downloads, many rated suitable for all age groups. Investigators also found technical connections to Chinese-language developers and a Hong Kong-registered company called KeLing Technology Limited, though the evidence does not conclusively establish who controlled the entire network. Apple spokesperson Adam Dema said the apps complied with App Store rules during their initial reviews, and developers subsequently changed the websites supplying their content, replacing approved material with pornography. Apple also said it had suspended tens of thousands of apps over the previous year for similar bait-and-switch tactics. On October 16, Apple updated its guidelines to strengthen restrictions against nonconsensual intimate imagery and apps that conceal functionality during review.
This story is the week's reminder that AI safety is not only about what frontier models can do in a lab. It is about distribution: app stores, ad networks, and remote servers that can change what an app does after it passes review. A review that checks an app once cannot catch content swapped in later from an external server, and that gap is exactly what this network exploited. The fixes will have to be ongoing, not one-time, and that is a harder problem than any single removal.
How the Week's Stories Stack Up
Aspect | Story | Key Number | Source |
|---|---|---|---|
Publisher control | Shared robots.txt AI training signal | Google support planned 2027, Microsoft effective February 1, 2027 | Unite.AI |
Infrastructure finance | Nvidia weighs guaranteeing Nscale debt | $44.6B first contract, 460-megawatt building, second deal about $45B | TechCrunch |
App safety | Hidden deepfake network exposed | More than 400 apps and domains, 19 removed by Apple, 350,000-plus Google Play downloads | WIRED via AppleInsider |
The thread across all three: AI is moving from what it can do to who controls it, who pays for it, and who stops its abuse. Publishers are getting one shared switch, but only if providers honor it. The buildout is getting so expensive that the chipmaker may have to guarantee the loans. And the safety fight is shifting from model capabilities to the stores and servers that distribute them. Last week was about where AI runs and who gets it first. This week is about the rules, the money, and the enforcement behind it.
If you like quick video takes on how new open AI projects work, browse the repo signals for 30-second overviews. For last week's roundup, see This Week in AI: October 30, 2026.
Frequently Asked Questions
What is the shared robots.txt AI training signal?
It is a common robots.txt preference, announced October 30, 2026, that lets website owners express AI training preferences once for all participating providers, using the policy token `__usage_ai-training` in the User-Agent line. Google aims to support it in 2027, OpenAI aims to implement it in the coming months, and Microsoft's version takes effect February 1, 2027. The underlying Internet-Draft, authored by Kevin Kelley of Anthropic, is an individual submission and not yet a formal IETF standard.
What is Nvidia considering for Nscale?
Nvidia is in talks, with no deal signed, to guarantee billions in debt financing for Nscale's Anthropic data center in West Virginia under a residual value guarantee, meaning Nvidia would step in if Nscale defaulted on construction payments. Anthropic has signed a $44.6 billion contract for the first 460-megawatt building, which is expected to use Nvidia's Vera Rubin chips, with a second deal worth roughly $45 billion under discussion.
What did the WIRED deepfake investigation find?
Published November 2, 2026, the investigation found more than 400 connected apps and web domains that looked like harmless AI services but could load sexually explicit deepfake tools for photographs of real people, controlled by remote servers after installation. Apple removed 19 apps, including Viefur, after reporters contacted the company, and 27 related apps on Google Play had accumulated more than 350,000 downloads in September 2026.
Conclusion
One shared signal will not end the debate over AI training data, but it gives publishers something they did not have: a single, simple way to say no. One guarantee will not settle whether the AI buildout is affordable, but it shows how far even Nvidia is willing to go to keep it moving. And one investigation will not end abusive deepfakes, but it exposes a distribution trick that app review alone cannot catch. The technology keeps advancing. This week, the interesting part was everything around it: the rules, the financing, and the enforcement.
Read this article in Bahasa Indonesia: Berita AI: 03 November 2026.