AI for Intelligence

Google and Microsoft adopt a shared signal to block AI training

Google and Microsoft will read one opt-out from AI training

What happened. On 30 October Google said it plans to support a shared "policy token" in robots.txt, the file that tells automated crawlers what they may do on a site (Mueller, 2026). A site that refuses training with this token will be treated as if it had blocked Google-Extended, Google's own training control. Google aims to support it in 2027. The token comes from the IETF AI Preferences working group, an internet standards body. OpenAI announced the same commitment the same day (OpenAI, 2026b).

Microsoft Bing will honour the shared token from 1 February 2027 and adds a separate page-level setting for grounding, the use of a page to support an AI-generated answer (Madhavan, 2026). Sites can allow grounding, refuse it, or limit it to Microsoft's own products such as Copilot. After 1 March 2027 Microsoft will stop reading its older NoArchive and NoCache tags as AI preferences.

Why it matters. Publishers can now refuse training while still being cited in AI answers, or the reverse. Microsoft warns that refusing grounding reduces brand visibility in AI experiences, and that neither choice guarantees traffic or payment.

What to do. Ask your web team to decide a training and a grounding policy for your sites before 1 March 2027.

The UK competition regulator investigates Meta's crawl terms

What happened. On 30 October the UK Competition and Markets Authority opened an investigation into Meta's commercial terms for businesses (Competition and Markets Authority, 2026). The terms would let Meta collect content from the websites of businesses that use its platforms and use it for its own products, including AI features, without giving them a choice. Meta paused the rollout on 28 October. The regulator will assess whether Meta holds a dominant position and abuses it under the Competition Act 1998.

Why it matters. A robots.txt opt-out means little if a platform's terms of service can override it. The outcome will show how far a platform can make access to its customers conditional on using their content.

OpenAI gives approved labs a biology model that refuses less

What happened. On 30 October OpenAI published the system card for GPT-Rosalind-6, a model for biology and drug discovery (OpenAI, 2026a). It uses the same weights as OpenAI's flagship reasoning model, with no extra training. What changes is the refusal policy: on a biorisk knowledge test it declined 9.4% of questions, against 29.2% for the flagship. Access is limited to approved organisations, which must pass business verification and apply qualified human review before outputs inform experiments or decisions.

Why it matters. The same model now behaves differently depending on who the customer is. What your organisation can do with AI depends on the approvals it holds, as well as the model it picks.

Apple brings Alibaba's Qwen into Siri in China

What happened. On 31 October Apple said Apple Intelligence features will reach users in China, with testing from 4 November and a public release with iOS 27.3 in December (Apple, 2026). Apple's own models run the core features. Through a partnership with Alibaba, Siri can pass some requests to Qwen, and Writing Tools can use Qwen to create text from scratch. Users sign in with a Qwen account or with Sign in with Apple.

Why it matters. A global platform entered a regulated market by building a local model into its assistant. Companies selling AI products in China should expect to partner rather than ship alone.

OpenAI cuts its funding target by a third

What happened. OpenAI is now seeking up to $20 billion in its current round, down from about $30 billion, CNBC confirmed on 31 October (Capoot & Kolodny, 2026). The Financial Times reported a post-money valuation of $1.25 trillion and said some investors reduced their allocations. The round is still open. The cut followed a week in which investors questioned whether big technology firms' AI spending will pay off.

Why it matters. Investor caution has reached the largest AI company. Supplier plans that rest on promised computing capacity deserve a closer look.

An open model predicts which pages AI search engines will cite

What happened. Two researchers tested how well models can predict which sources an AI answer engine will cite, using 229 queries answered by ten engines that saw the same candidate pages (Bajemon & Rochet, 2026). The engines agreed with each other at a rank correlation of 0.704. A small open model they trained and released, GEO Citation Ranker 4B, reached 0.542 to 0.549, statistically level with much larger models.

Why it matters. Companies can estimate how often AI answers will cite their pages, or a competitor's, without paying to query each engine.

Sources

  • Apple. (2026, October 31). Apple 智能即将登陆中国 [Apple Intelligence is coming to China]. Apple Newsroom. https://www.apple.com.cn/newsroom/2026/10/apple-intelligence-features-are-coming-to-users-in-china/
  • Bajemon, E., & Rochet, A.-L. (2026). Predicting what answer engines cite: Open decision models, rerankers and a distilled 4B student on ten engines under controlled retrieval (arXiv:2610.38090). arXiv. https://arxiv.org/abs/2610.38090
  • Capoot, A., & Kolodny, L. (2026, October 31). OpenAI pares fundraising goal, now seeking up to $20B in latest round. CNBC. https://www.cnbc.com/2026/10/31/openai-fundraising-20-billion-valuation.html
  • Competition and Markets Authority. (2026, October 30). CMA opens investigation into Meta's commercial terms. GOV.UK. https://www.gov.uk/government/news/cma-opens-investigation-into-metas-commercial-terms
  • Madhavan, K. (2026, October 30). Your content, your choice: Clearer preferences for training and grounding. Microsoft Bing Blogs. https://blogs.bing.com/search/2026/10/your-content-your-choice-clearer-preferences-for-training-and-grounding/
  • Mueller, J. (2026, October 30). Supporting a shared AI training control for websites. Google Search Central Blog. https://developers.google.com/search/blog/2026/10/ai-training-control
  • OpenAI. (2026a, October 30). GPT-Rosalind-6 system card addendum. OpenAI Deployment Safety Hub. https://deploymentsafety.openai.com/gpt-rosalind-6
  • OpenAI. (2026b, October 30). Making it easier for websites to express AI training preferences. https://openai.com/index/helping-websites-express-ai-training-preferences/