AI News Oct 31, 2026

On October 30, OpenAI, Google, and Microsoft each described a shared robots.txt signal for how their systems may use a site's pages to train generative models. The same day, OpenAI said ChatGPT will give less detail on weapons and past mass attacks, and it promised a Community Safety Report for early 2027. Amazon's third-quarter recap put AWS growth at 39.3 percent, with a $578 billion backlog and custom-chip commitments that have multiplied. Google's Project Suncatcher satellite sent back its first answer from a data-center-class TPU in orbit.

In the news

Training preferences

A site will be able to state, once, how participating companies may use its pages for foundational model training. OpenAI, Google, and Microsoft all point at the same IETF Internet-Draft, draft-kelley-aipref-training-preference-signal¹. OpenAI also links a companion vocabulary draft, says it is joining other working-group participants, and asks other providers to honor the signal. None of the three posts says the control deletes material already inside a trained model, and none of the dates they published treats the signal as switched on today.

OpenAI says it will honor the common signal in the coming months and will publish implementation details later. The new line does not replace the crawler controls OpenAI already runs, and those stay in force². Google's support target is 2027, with updates after that. If a site uses the policy token to opt out of training and does not set Google-Extended itself, Google plans to read that as a Google-Extended disallow.

A site that specifically allows Google-Extended stays eligible to appear in Gemini Apps answers and to receive clicks through grounding, and company-specific controls keep working beside the joint signal. Google dates Google-Extended to 2023 and notes that the EU Code of Practice for general-purpose AI model providers, under the EU AI Act, already treats robots.txt as a primary machine-readable preference³.

Microsoft is adopting the shared AI Training preference and adding a separate AI Grounding preference. Training can be set in robots.txt for a site or a path, and a page can set it too, while grounding is page-level only. The robots.txt examples use the user-agent __usage_ai-training, with Disallow for a path or the whole site and an optional Allow. A page can send ai-training:disallow in a Bingbot meta tag or an X-Robots-Tag, and grounding uses ai-grounding:disallow or ai-grounding:disallowextended⁴.

Allow is the default for both preferences when a site publishes nothing. A grounding Disallow means the page is not used to inform generated answers. DisallowExtended limits grounding to Microsoft-operated experiences, which the post lists as AI Mode, AI Overview, Copilot, and Microsoft 365, and it excludes grounding Microsoft provides to third parties, including Web IQ and API access in Foundry. Training and grounding are independent, so a site can refuse training and still allow grounding. Neither setting changes eligibility for ordinary Bing results, and neither promises inclusion, traffic, or payment.

An Allow does not override a robots.txt rule that already blocks Bingbot from crawling. Microsoft says today's Bing search syndication API still returns links for ordinary search and is not changed by the grounding preference. Honoring begins as Microsoft crawls, starting February 1, 2027, and on March 1, 2027, NoArchive and NoCache stop being read as a training or grounding preference.

Sites that want to keep a grounding restriction have to publish the new on-page preference and record a temporary one in Bing Webmaster Tools before March 1. That temporary control opens on December 15, 2026, and only for sites already using NoArchive or NoCache on that date. It expires on September 30, 2027, or sooner if a recrawl finds the new tag. If nothing temporary is on file by March 1, pages that still carry the old tags default to Allow for grounding until they are recrawled. There is no temporary training control in Webmaster Tools, and once Microsoft detects the new preferences on a site it disregards leftover NoArchive and NoCache tags there.

The same morning, OpenAI published excerpts from briefs supporting fair use in the New York Times lawsuit, including a government excerpt arguing that narrowing fair use to exclude training of OpenAI's models would conflict with basic copyright principles. The post does not say those briefs were filed on October 30, and it does not report a ruling⁵.

Community safety

OpenAI said it made more than 200 law-enforcement referrals in the third quarter of 2026 for imminent threats of violence. The figure counts referrals, not plots stopped. The company promised a Community Safety Report for early 2027, covering the second half of 2026, and then every six months, explaining how its systems detect, investigate, and respond, with anonymized examples, mitigations, policy changes, and data. The October 30 post is that promise, not the report itself⁶.

For every user, ChatGPT will give less detailed reconstructions of past mass attacks and of the people who carried them out, including weapons and equipment, how they were obtained, and comparisons of scale. The limit applies regardless of age or stated reason. People can still ask for help understanding the event, the effect on victims and communities, warning signs, and prevention. The change follows a May note about short factual summaries of earlier warning signs. When recent chats show signs of possible violence, ChatGPT is supposed to answer more conservatively about information that could be misused.

These safeguards begin rolling out over the next few weeks, and OpenAI does not describe them as already on for every account. The same stricter limit applies when the system estimates that the user is under 18, even with no sign of violent intent. When a conversation shows someone fixating on violence, or on seriously harming another person, ChatGPT will be able to end the chat, and repeated endings may pause the account for a time. OpenAI says it is still working with mental-health and violence-prevention experts on earlier automated intervention, so the auto-end and the pause are in progress rather than a finished switch.

For users under 18, a further limit covers practical information about weapons and security systems even when the question looks harmless. Existing rules already bar help planning violence and help finding or obtaining weapons. Teens can still ask about prevention, de-escalation, support, and how to reach a trusted adult. Age prediction continues to place estimated under-18 users in the teen experience, and an adult who lands there can confirm age through Persona.

OpenAI is working toward a parent notification if a linked teen account is paused for repeatedly discussing violence. Parents of linked accounts already receive some safety alerts, and a July change already tells them if an account is deactivated for a violence-policy violation.

AWS and custom chips

Amazon announced third-quarter results on October 29. In the recaps published after that announcement, AWS grew 39.3 percent year over year, accelerating for a sixth straight quarter, which Amazon calls the fastest growth in nearly five years. The backlog was $578 billion, growing at a triple-digit rate and described as well diversified. Andy Jassy said the company is installing an AI accelerator about every 14 seconds, a line from the call rather than a separate audited count⁷⁸.

On Amazon Bedrock, Fortune 500 companies spent nearly seven times as much year over year, and total token volume was nearly 17 times higher. Jassy called Bedrock the fastest-growing multi-billion-dollar service in AWS history. He said OpenAI models were added recently, and that Bedrock-managed Kimi, GLM, and SpaceX models had just been added, without dates for those listings. Amazon says inputs and outputs on Bedrock are not shared with model providers and are not used to train base models. Inference already accounts for well over half of AWS AI revenue and is still growing, with no dollar split given.

AgentCore has tens of thousands of customers, and in the quarter they deployed twice as many agents with it as in the prior three quarters combined. Those figures describe a product customers are already running.

The custom-chip business continued to grow at a triple-digit rate year over year. Revenue commitments for Graviton and for Trainium each increased more than 20 times from a year earlier. Those figures are purchase commitments, not a report that either chip's revenue grew twentyfold. Customers using the custom silicon saved more than $2 billion in the quarter, Amazon said⁹.

Graviton is described as up to 30 to 40 percent better on price-performance than the alternatives in the comparison, on the stated ground that post-training, reinforcement learning, and agent tool use mostly run on CPUs. After the Graviton 5 launch, backlog for prior-generation Graviton instances more than doubled from the previous quarter, and the posts do not date that launch. Short-term capacity for earlier Trainium instances is selling at a premium to the first half of the year. Customers have asked to use Trainium outside AWS, and Amazon says it intends to make that available late in 2027.

Jassy also said Amazon had just open-sourced TorchNeuron, the software it uses with Trainium. The earnings posts mention the release and do not describe the code, a license, or a download.

An answer from orbit

On October 30, Travis Beals, senior director of Paradigms of Intelligence at Google Research, wrote that a Project Suncatcher prototype had returned its first model output. Google calls the result the first AI response from a data-center-class TPU in orbit. The satellite reached orbit on October 1, and the new result is the onboard answer¹⁰.

During a ground-station pass, the team uplinked a prompt to Gemma 4, an open model running on the satellite, and asked it to write a haiku about being in space. The downlink, as the satellite passed Svalbard, Norway, opens with the line "Blue marble below."¹⁰ They then tried space-themed limericks and prompts about the payload's status.

Google places the test in "this week" and does not give an hour. The next measurements are how the TPUs behave under thermal stress and radiation, and the milestone named after that is a pair of interconnected satellites in 2027. The post is a research note about one prototype. It is not a cloud region, it is not a Gemma 4 service launch, and nobody can rent orbital TPU capacity from it.