AI training: a shared signal | 31 October

In one minute

Four providers support `__usage_ai-training` in `robots.txt`; Microsoft separates training from using content to support generated answers.
The CMA is investigating Meta, and Norway opens a consultation until 11 December. These are separate proceedings, without automatic effects on current Norwegian law.
AWS grows 39.3%, but Amazon records negative free cash flow over twelve months. Google runs Gemma 4 in orbit, and Afi estimates Spanish workers' exposure to AI.

Main story

CONTENT · TRAINING · WEB PREFERENCES

Announcements and proposal published on 30 October 2026.

One preference for several providers, with gradual implementation

Kevin Kelley of Anthropic submitted the individual Internet-Draft Machine-Readable Training Preference Signal. It uses the `__usage_ai-training` token in `robots.txt`, the file through which a website communicates instructions to crawlers. Preferences can cover the whole site or particular paths. It is not an approved IETF standard. [1][2]

Each company has its own schedule:

- Anthropic: has started implementation but still needs to extend support across its services. It gives no completion date. [2] - OpenAI: aims to implement support in the coming months and retain its existing crawler controls. [3] - Google: its announcement targets 2027; its documentation specifies early 2027. Provider-specific controls, such as Google-Extended, will continue to work. [4][5] - Microsoft: will honour the new preferences as it crawls pages from 1 February 2027. [6]

Training a model and supporting an answer are different uses

Microsoft adds an independent control for grounding: using web content when producing an answer. This is set at page level, while the shared training preference can cover a site or path. Eligibility for Bing search results remains unchanged, although restricting grounding reduces visibility in generative experiences. [6]

Why it matters. A publisher needs to assess desired visibility and acceptable uses separately. The signal simplifies communication among participants, but does not itself block downloads or create a licence or automatic entitlement to payment. [1][6]

Other news

1. The UK investigates Meta's terms for using businesses' content

COMPETITION · UNITED KINGDOM

Investigation announced on 30 October.

The Competition and Markets Authority will examine whether Meta holds a dominant position and whether its commercial terms constitute abuse under the Competition Act 1998. It is concerned that businesses and publishers dependent on its services must accept collection of content from their websites for Meta's products, including AI features, without meaningful choice. [7]

The changes were planned for 30 October, postponed to 20 November, and Meta announced a pause on 28 October, pending consultation with industry groups. The CMA has not found an infringement. [7]

This follows the content dispute we covered on 29 October: according to MBL, an interested party, Oslo District Court had temporarily stopped the terms for Norwegian publishers. These are separate proceedings. [8]

2. Norway consults on the AI Digital Omnibus until 11 December

REGULATION · NORWAY · EEA

Consultation opened on 30 October.

The Ministry of Digitalisation and Public Governance is consulting on Regulation (EU) 2026/1744. It seeks to simplify compliance for smaller businesses, clarify requirements and expand AI testing. The consultation supplements the Norwegian process launched on 30 June 2025. Responses are due by 11 December 2026. [9][10]

The official memorandum states that the Omnibus entered into force in the EU on 27 July 2026. It explains that high-risk requirements under Annex III will apply from 2 December 2027, and those under Annex I from 2 August 2028. These are EU dates; the consultation does not itself amend current Norwegian law. [11]

3. Amazon accelerates in the cloud as markets punish AI costs

BUSINESS · INVESTMENT · MARKETS

Results on 29 October; market close on 30 October.

AWS reported $46.0 billion in third-quarter revenue, up 39.3% year on year. Amazon's group operating income was $28.1 billion. Free cash flow for the twelve months ending 30 September was −$42.4 billion, compared with +$14.8 billion in the comparable period. The company attributes the deterioration primarily to AI infrastructure investment. [13]

On Friday, the Nasdaq Composite fell 2.86% and Intel 28.6%, according to Reuters' closing report. The agency links the technology selloff to disappointing results and doubts about investment returns. Revenue growth can coexist with cash consumption; a market decline does not by itself establish a bubble. [14]

4. Google runs Gemma 4 on a TPU aboard a satellite

RESEARCH · COMPUTING IN SPACE

Announced on 30 October.

Google reports a first response from Gemma 4 generated by a TPU, a specialised AI processor, in orbit. The Project Suncatcher satellite launched on 1 October. The response reached the ground during a pass near Svalbard. [15]

The company plans to test resistance to radiation and thermal stress, and to experiment with two interconnected satellites in 2027. This is an experimental demonstration reported by Google; it does not establish that space-based data centres are already competitive on cost, energy or reliability. Assessing that possibility will require evidence of sustained operation as well as an initial response. [15]

5. Afi estimates that 11.2 million Spanish workers have high AI exposure

EMPLOYMENT · SKILLS · SPAIN

Study presented on 29 October.

Fundación Afi Emilio Ontiveros examines the mismatch between education and employment. Using 2025 data from Spain's EPA labour force survey, it finds that 51.4% of workers have jobs matching their educational level, 37.6% are overqualified and 10.9% underqualified. [16][17]

The study estimates that 11.2 million workers in 2024 held occupations with above-average AI exposure. It also considers up to 0.4 percentage points of additional potential growth. Exposure does not mean job losses, and that growth is a scenario, not an observed effect. The practical lesson concerns organisation: introducing tools requires adapting tasks and making use of existing skills. [16]

Briefs

OpenAI collects arguments about fair use. On 30 October, the company published extracts from third-party briefs supporting training on copyrighted works. These are positions selected by a party to the litigation, not a judgment; fair use belongs to US law. [18]

Community Safety Report. OpenAI announces its first report for early 2027, covering the second half of 2026, with updates every six months. It also announces restrictions on information about mass attacks and additional protections for minors. These are company measures and commitments, without an independent effectiveness assessment in this announcement. [19]

The signal in the noise

Control is divided among three decisions: which content can be used, under which rules, and with what return. Publishers negotiate uses, authorities examine terms, and businesses fund infrastructure and reorganise work. Today's editorial point is the connection: a more capable model does not itself settle the rights over its resources or the value it produces.

Spain and Norway

Spain. Afi places the use of existing skills at the centre of the debate. Separately, Microsoft Spain president Paco Salcedo called for dialogue on the draft royal decree on data centres on 30 October, according to Europa Press. This is a business position, not the approval of legislation. [16][21]

Norway and the EEA. EFTA continues to list the AI Act, Regulation (EU) 2024/1689, as under scrutiny for incorporation into the EEA Agreement. Norway's government is still preparing the national law and is now consulting on the Omnibus. EU adoption and application in Norway must be distinguished. [9][11][12]

On the radar

- Microsoft controls. Sites using NoArchive or NoCache as of 15 December 2026 should review the transition before 1 March 2027. To maintain grounding restrictions, Microsoft asks for page-level preferences and a temporary preference in Bing Webmaster Tools. [6] - AI in space. The experiments planned for 2027 should show what connecting satellites can contribute. [15] - Business inventories. Wiz publishes a commercial survey of 100 senior IT and security decision-makers about AI asset oversight. A topic to follow, without extrapolating its findings to all businesses. [20]