AI news, 1 November 2026: one line in robots.txt now speaks for your whole site.

Google and Microsoft published posts on the same day agreeing that a single token in robots.txt should carry your AI-training preference to every provider that honours it. The token is defined in an Internet-Draft written by an engineer at Anthropic. Microsoft attached dates: 15 December 2026, 1 February 2027, 1 March 2027. Separately, Microsoft priced $25 billion of notes running out to 2066, in a quarter where it spent $39.4 billion on property and equipment against $19.4 billion a year earlier. Gemini started editing files inside the chat instead of handing them back as downloads. Two open Swift packages made local speech recognition swappable on Apple hardware. And IBM shipped a model patch that changed two components and left the base model alone.

Every date and figure below was read at the source linked in its own section during this run. Where a claim could not be opened at a primary source, it was cut rather than softened, and the cuts are listed near the end. AcquihireTech writes this roundup for Indian service businesses doing ₹10L+ a month, so each story goes through the same filter: what does this change about how you run the business on Monday?

1. Google and Microsoft agreed on one robots.txt line for AI training

On Friday 30 October 2026, Google's Search Advocate John Mueller posted Supporting a shared AI training control for websites on the Search Central blog. The same day, Microsoft published Your content, your choice on the Bing blog. Both point at the same mechanism: a shared "policy token" that a site names in robots.txt, so that one group of rules expresses an AI-training preference to every provider that agrees to read it, instead of a separate arrangement per company.

The token is __usage_ai-training, and it is specified in an Internet-Draft titled Machine-Readable Training Preference Signal, authored by Kevin Kelley of Anthropic, dated 30 October 2026 and set to expire on 3 May 2027. It is submitted to the IETF's AI Preferences working group. The syntax reuses robots.txt exactly as it already works:

User-Agent: __usage_ai-training
Disallow: /
Allow: /press-kit/

That example expresses a preference against AI training everywhere except /press-kit/, where it expresses a preference in favour. Four rules in the draft are worth knowing before you write anything. A User-Agent: * group must not be read as matching the policy token, so your existing wildcard rules do not quietly become a training preference. A robots.txt file with no __usage_ai-training group expresses no preference at all, which is not the same as refusing. The token is not a crawling directive, so it does not stop anyone fetching the page. And private agreements can override whatever you publish this way.

Google's position is specific. If your robots.txt uses the policy token to opt out of training and does not separately name Google-Extended, Google plans to interpret that as though you had disallowed Google-Extended, the control it has offered since 2023. Company-specific controls keep working alongside the shared signal, so you can set a broad preference and then tailor one provider. Mueller's post also states that sites which specifically allow Google-Extended remain eligible to appear in responses and receive clicks through grounding in Gemini Apps, and notes that the proposal is consistent with the EU Code of Practice for General Purpose AI model providers, which already treats robots.txt as a primary method for publishers to express machine-readable preferences. On timing, Google says only that it aims to support the control along with other AI providers in 2027.

Microsoft's post carries the actual calendar, and it separates two things most people conflate. AI Training covers using your content to change a model's learned parameters. You set it site-wide in robots.txt with the shared token, or per page with <meta name="Bingbot" content="ai-training:disallow"> or the equivalent X-Robots-Tag; the page-level setting overrides robots.txt. AI Grounding covers using your content as input to an AI-generated answer, and Microsoft is explicit that indexing and ranking for ordinary search is not grounding. Grounding is page-level only. There is no site-wide grounding switch in robots.txt. Its options are Allow, which is the default, ai-grounding:disallowextended, which permits grounding inside Microsoft-operated surfaces such as AI Mode, AI Overview, Copilot and Microsoft 365 while excluding third-party grounding services like Web IQ and the Foundry grounding API, and ai-grounding:disallow, which refuses grounding outright. Both default to Allow, and Microsoft states that neither setting affects whether you appear in Bing search results.

The dates, in the order they arrive. 15 December 2026: temporary grounding preferences open in Bing Webmaster Tools, for sites that are using NoArchive or NoCache as of that date. 1 February 2027: Microsoft begins honouring the new preferences as it crawls. 1 March 2027: the deadline to publish page-level grounding preferences and record a temporary preference in Bing Webmaster Tools, after which NoArchive and NoCache stop being read as training or grounding preferences at all. 30 September 2027: temporary grounding preferences expire, or earlier if a page is recrawled and a published preference is found. There is no temporary setting for training, only for grounding.

Why it matters. Two reasons, and the second is the one with a cost attached. First, this is the first time Google and Microsoft have published, on the same day, in support of the same line of text, which turns an open-ended per-vendor chore into one edit. Reports that other providers will honour it too are plausible and are not primary, so treat the two posts above as what is confirmed. Second, there is a trap inside the transition. If your site has NoArchive or NoCache set from some earlier decision, those tags stop carrying meaning for AI use after 1 March 2027, and pages relying on them default to Allow for grounding until Microsoft recrawls. Nothing breaks loudly. A preference you thought you had expressed simply stops existing, and you find out by not finding out.

What it means for a service business. Decide training and grounding separately, because for most lead-generating sites the commercial answer differs between them. Training is content going into a model you do not control and cannot be cited by. Grounding is how your page gets quoted, and clicked, inside an AI answer, which for a firm that sells on expertise is distribution rather than leakage. Refusing both is a choice some businesses should make and most should think about twice. The concrete steps are small: check whether NoArchive or NoCache appear anywhere in your headers or meta tags today, put 15 December 2026 in a calendar if they do, and decide now whether your service pages, case studies and blog should be groundable. Then add the __usage_ai-training group to robots.txt, because a preference published early costs nothing and a preference published in March 2027 is late by one crawl.

2. Microsoft priced $25 billion of notes, and its quarterly capex doubled

Microsoft filed a pricing term sheet with the SEC under Rule 433, dated 29 October 2026, for ten tranches of senior unsecured notes. Two are floating rate: $1 billion due November 2028 at Compounded SOFR plus 35 basis points, and $1 billion due November 2029 at SOFR plus 53 basis points. Eight are fixed: $2 billion at 4.850% due 2028, $3.25 billion at 5.000% due 2029, $3.25 billion at 5.200% due 2031, $2.5 billion at 5.350% due 2033, $4 billion at 5.550% due 2036, $2 billion at 6.150% due 2046, $4 billion at 6.250% due 2056, and $2 billion at 6.350% due 2066. That totals $25 billion. Trade date was 29 October, settlement 4 November 2026 on a T+4 basis. The notes are rated Aaa by Moody's and AAA by S&P, both stable, and will not be listed on any exchange. J.P. Morgan, UBS and Wells Fargo are joint book-running managers alongside eight others.

The filing does not say the money is for AI. Microsoft's base prospectus language is general corporate purposes, which it lists as working capital, capital expenditures, share repurchases, acquisitions and repaying existing debt. Anyone telling you the bond is earmarked for data centres is reading an intention into a document that does not state one. What the filings do show, in the Form 10-Q for the quarter ended 30 September 2026, filed 28 October, is the scale of what is being funded. Cash additions to property and equipment were $39.367 billion in the quarter, against $19.394 billion in the same quarter a year earlier. Slightly more than double, in twelve months, on a base that was already large.

Why it matters. The number to look at is 6.350% for forty years. A borrower rated AAA by both agencies, with the balance sheet Microsoft has, chose to lock in money until 2066 and pay six and a third per cent for the privilege. That is not a company expecting this to be a short build. It is also a fixed obligation: the coupon has to be paid in 2032 and 2045 whether or not the demand forecast behind the spending holds. Compute pricing is set by people carrying that obligation, which is the opposite of the condition under which prices fall quickly.

What it means for a service business. Stop building margin models that assume inference gets cheaper on a schedule. Per-token prices have fallen a long way and may well keep falling, but the people selling you tokens now have forty-year interest payments in the denominator, and your pricing should not depend on their generosity. Two practical consequences. Keep every model call behind one internal interface, so switching provider is a configuration change rather than a project. And if you have quoted a client a fixed monthly figure that embeds an assumption about token costs twelve months out, go and look at what happens to that engagement's margin if the cost stays flat instead. Flat is the scenario most people have not modelled.

3. Gemini now edits the file instead of handing it back

On 29 October 2026, Google announced that users can generate, preview and edit documents, PDFs, presentations, spreadsheets and code directly inside the Gemini app. You can expand a file to full screen and type changes into it, highlight a passage to ask a question about it, or ask Gemini to revise it. Google's own example is turning uploaded handwritten lecture notes into a PDF study guide, then editing that guide in place.

Format support is split by platform, and the split matters. PDF preview and editing work on both web and mobile. Google Docs, Google Sheets, slide decks, Markdown, TXT and code are web only. Google says the capability is available to all users, in every language where the Gemini app is offered.

Why it matters. The round trip is where document automation usually dies. A model that produces a file you then download, open in another application, correct, and re-upload has not removed the work; it has moved it and added two file versions. Editing in place removes a specific, boring failure point. That is a smaller claim than a new model and a more reliable one, because the gain does not depend on the output being right the first time.

What it means for a service business. The obvious use is drafting: proposals, scopes, quotes, SOPs, the client-facing summary somebody writes at 9pm. The less obvious point is governance. This is the consumer Gemini app rolling out to all users, which includes your team, and the files they will paste in include client material. If you do not already have a written rule about which client documents may go into a consumer chat application and which may not, this is the week to write one, because the feature makes pasting a whole document more attractive than pasting a paragraph. A one-page rule naming the categories, approved by whoever owns client confidentiality, is more useful than a ban nobody follows.

4. Local speech recognition became swappable on Apple hardware

On 29 October 2026, Hugging Face published Introducing AnySpeechRecognizer and AnyDecisionModel, written by Mattt. They are two open Swift packages. AnySpeechRecognizer implements Apple's SpeechAnalyzer API but lets you choose the engine behind it: Apple's own Speech framework by default, or NVIDIA's Parakeet v3, Parakeet Ultra and Parakeet Unified, or Alibaba's Qwen3-ASR at 0.6B and 1.7B, all running locally through Core ML. It also adds speaker diarization for up to four speakers using Streaming Sortformer 4spk.

The post publishes word error rates on LibriSpeech, measured with the hear CLI on an M4 Max Mac Studio running macOS 27.0.1, on files of three to eight minutes. Parakeet Unified in file mode records 1.43% on test-clean and 2.19% on test-other. Apple's own file transcriber records 2.12% and 3.97%. In live mode Parakeet Unified records 1.45% and 2.42%, which is the more interesting pair, because live transcription usually degrades sharply and here it barely moves. Parakeet v3 in live mode, by contrast, goes to 4.73% on test-clean. Licences for the converted Core ML models differ: Parakeet v3 and Parakeet Ultra are CC BY 4.0, Qwen3-ASR is Apache 2.0, and Parakeet Unified and Sortformer 4spk v2.1 are under the NVIDIA Open Model License. Both packages need Swift 6.2 and macOS 26 or iOS 26 or later; Parakeet, Sortformer and the MLX backend require Apple silicon. The author states the packages were tested on macOS 27 and not on iPhone.

The second package is the more transferable idea. AnyDecisionModel answers yes-or-no, multiple-choice and rating-scale questions by reading the model's output probabilities rather than parsing generated text, returning a typed result with a probability attached. It supports six backends, including MLX, Core AI, llama.cpp and Ollama.

Why it matters. Reading probabilities instead of parsing prose is the correct shape for a routing decision, and almost nobody builds it that way. If you ask a model "is this enquiry a qualified lead, yes or no" and then regex its sentence, you get a brittle string and no confidence. If you read the probability, you get a number you can threshold, log, and tune later when you discover the threshold was wrong. The same change turns an unauditable judgment into something you can chart.

What it means for a service business. Honestly, not much directly. This is Apple-only, the strongest transcription results are English-only, and if your call volume is Hindi, Hinglish or regional languages on a phone line, none of it deploys as written. The architecture is what transfers, in two parts. Keep speech recognition behind one interface so the engine is a swappable component rather than a rewrite, which is exactly the abstraction AnySpeechRecognizer exists to demonstrate. And make every automated routing decision in your pipeline return a score rather than a sentence. A routing step that records a confidence number can be audited later; one that parsed a model's prose leaves you nothing to audit, which is the position AcquihireTech argues from whenever a pipeline needs to be explained to the person who owns its outcome.

5. IBM patched a model the way you patch software

Also on 29 October 2026, IBM published Granite Switch 4.1.1 on Hugging Face, a patch release to Granite Switch 4.1, which came out on 15 October. The base model did not change. Two components were retrained: a new speech recognition model, granite-speech-5.0-470m-turboctc, roughly 0.5B parameters, doing multilingual speech-to-text across ten languages with punctuation and capitalisation built in, and the Decision adapter in granitelib-core-r1.0. The release is backward compatible with 4.1 and needs no migration. Adapters compose with the base model on a single endpoint and can be swapped without touching it. IBM says an Adapter SDK and tutorials are coming so that developers and third parties can build, version and test their own adapters on a software-like release cycle, and gives no date for that.

Two caveats from the post itself. It claims "improved overall performance" on the Jev Decision Index test set and publishes no numbers for it. And it does not state a licence for the models involved. So the performance claim is IBM's framing, and what is checkable here is the architecture rather than the improvement.

Why it matters. The reason model upgrades frighten people running AI in production is that everything moves at once. A new version arrives, your prompts behave differently, your evaluation set has to be rerun from scratch, and you cannot tell which change caused which regression. If only one adapter changed and the base model is bit-for-bit identical, the blast radius of an upgrade is knowable before you deploy it. That is a release engineering idea rather than a capability one, and it is worth more to an operator than most benchmark deltas.

What it means for a service business. You will almost certainly not run Granite. Take the discipline anyway, because it costs an afternoon. Pin the model version in every integration you have, so that nothing changes underneath you on a vendor's schedule. Keep a small evaluation set, twenty to fifty real examples with the answers you consider correct, and run it before and after any version change. Most businesses running AI in production have neither, which is why "the AI got worse" is a sentence people say and cannot substantiate.

The window, and the stories we cut

The window for this edition is 29 October to 1 November. It is wider than a day because 31 October was a Saturday and genuinely quiet: nothing dated 31 October or 1 November could be opened at a primary source, so nothing from those two days is reported above. Three of the five stories are dated 29 October and none of them appeared in the 31 October edition, which covered Mistral's weights release, OpenAI's Model Spec, Deepgram's acquisition of Daily, Project Suncatcher and Amazon's quarterly filing. Everything above was read at the source linked in its own section during this run: Google's Search Central blog, Microsoft's Bing blog, the IETF draft, Microsoft's Rule 433 term sheet and Form 10-Q on SEC EDGAR, Google's Gemini post, and two Hugging Face blog posts. Where secondary coverage and the primary source disagreed, the primary source is what is written above.

Seven candidates were cut. OpenAI's funding round being trimmed to roughly $20 billion at a $1.25 trillion valuation is the one readers will have seen most, and it is cut for the same reason as last week: openai.com returns HTTP 403 to this pipeline, and no primary OpenAI document could be read. A number that large should not be repeated on the strength of aggregator summaries. The expanded OpenAI and ServiceNow partnership is real and has a primary source in ServiceNow's newsroom, but it is dated 28 October, outside the window, and sits next to the ServiceNow item in the 28 October edition. AWS's Step Functions connector for Bedrock AgentCore Gateway, which exposes state machines to agents as MCP tools, is genuinely useful and appears in AWS's October release notes, but those notes date entries only to the month, so the window test cannot be applied and the story fails it by default rather than on merit. An NVIDIA Nemotron post listed on Hugging Face for 30 October returned 404 to this pipeline. Project Suncatcher's first satellite response is dated 30 October and inside the window, but it was the fourth story in the 31 October edition. A reported S&P move of Meta's outlook to negative over AI spending has no primary S&P document that could be opened. And the widely repeated line that Microsoft's $25 billion is its first major US bond sale since 2017 is absent from the filing itself, so it is not in the section above.

The short version of this edition: grep your site for NoArchive and NoCache today, and if you find either, put 15 December 2026 in your calendar. Add a __usage_ai-training group to robots.txt and decide training and grounding separately rather than together. Write the one-page rule about which client documents may be pasted into a consumer chat app. Pin your model versions and keep twenty real examples you can rerun. And re-run one engagement's margin on the assumption that token prices stay exactly where they are, because somebody is now paying 6.350% until 2066 to supply them.