← Back to the front page

AI News 9 November 2026

Norway plans to require licences for data centres, while protesters across Europe are calling for a pause in construction. We also look at how AI is changing job ads, why a Norwegian committee wants to phase out short take-home exams, and how Google uses AI to find quantum programs with machine-checked proofs.
12 storiesUpdated
Regjeringen / Norwegian Government

Norway plans data centre licences as Statnett pauses new grid reservations

On 9 November, Norway's government began work on a licensing scheme for new and existing data centres. At the same time, grid operator Statnett paused new grid capacity reservations for data centres. The government wants to set conditions for building and operating facilities, while Statnett needs to ensure the grid can handle their connection.

Read moreShow less

Statnett says data centres with a combined power demand of 800 MW are connected, while 4,200 MW has been reserved and 9,100 MW is in the capacity queue. The reservations show how much grid capacity has been set aside. The queue shows how much additional capacity projects have requested.

Statnett says large, rapid swings in power demand can make the grid harder to operate. Facilities that already meet the necessary technical requirements can still connect. Statnett expects to lift the temporary pause by 1 August 2027.

The government will consider requirements for digital security, preparedness and access to AI computing capacity. It will also look at how large new electricity users could cover more of the costs they impose on the grid. The licensing scheme is still being developed.

Stop Dirty Data Centres

Protesters across Europe call for a pause in data centre construction

Protesters in around 15 European countries rallied against data centre construction on 6 and 7 November, according to Le Monde. In Amsterdam, activists entered the construction site of a future Microsoft data centre and work stopped for the day. The protests were part of the Stop Dirty Data Centres campaign.

Read moreShow less

The protests targeted specific projects. In Zaragoza, people marched in a region where several large cloud companies are expanding. In Warsaw, a local association delivered a petition to the speaker of the lower house of parliament, calling for a halt to the construction and expansion of very large data centres.

The campaign is asking elected officials to pause construction and give residents more say. Concerns vary by location. Near a planned facility north of London, a representative of local residents cited noise, diesel generators and water use for cooling.

Kunnskapsdepartementet / Ministry of Education and Research

Norwegian AI committee calls for phasing out short take-home exams

A Norwegian government-appointed committee recommends that universities and colleges phase out short take-home exams and give all instructors the skills to use AI. In its report, delivered on 6 November, the committee argues that a submitted text alone provides weaker evidence of what a student can do when AI tools are readily available.

Read moreShow less

The committee proposes supervised oral or written exams in place of short take-home exams. Students should still be able to work on substantial projects such as master's theses over time. The committee suggests adding a component such as an oral defence and counting both parts towards the grade.

In teaching, the committee recommends using AI where it helps students learn. Tools could provide feedback during an assignment or explain concepts, for example. That requires both students and instructors to be able to judge the quality of the answers.

Microsoft Research / Stanford University

Microsoft and Stanford track AI across 50,000 job ads

Microsoft and Stanford researchers analysed 50,000 US job ads from 33 companies. Tasks that often appear in Copilot and Claude usage data tended to be mentioned less often over time. Where the ads still mentioned those tasks, they more often described responsibilities involving AI.

Read moreShow less

The researchers extracted responsibilities from job ads and mapped them to standard categories from the US Department of Labor. They then used data on what people did with Copilot and Claude to estimate how exposed each type of work was to AI.

One part of the analysis compared nearly 23,000 ads from 20 companies that had enough postings in both 2023 and 2025. The strength of the association depended on how the researchers counted tasks. Some of the decline was related to how employers described and grouped responsibilities.

The researchers also found that AI appeared more often in descriptions of the company, its products or desired qualifications. An ad that mentions AI may therefore tell readers more about how the employer presents itself than about what the employee will do.

The study is a preprint and measures the wording of job ads. It can show which tasks employers emphasise, while the causes of the changes and their effects on hiring remain open questions.

Suno

Elvis and Eminem rights holders sue Suno

Rights holders for music associated with Elvis Presley and Eminem filed two separate lawsuits against AI music service Suno in California on 6 November. They allege Suno infringed copyright in their works. The Presley plaintiffs also allege the service used his voice without permission.

Read moreShow less

The Presley plaintiffs cite both federal copyright law and Tennessee's ELVIS Act, which protects a person's voice. That raises two questions about the generated music. Who holds rights in the material, and who can authorise imitation of the voice?

Music publishers Eight Mile Style and Martin Affiliated filed a list of works and examples of audio they say infringes their rights. Suno told The Verge that it blocks prompts naming specific artists or songs and uses third-party services to screen uploaded audio and lyrics.

Associated Press

AP says US AI safety centre was asked to report to the White House

Staff at the US AI safety centre CAISSI were asked to report informally to a White House office, according to two former officials who spoke to the Associated Press. The centre sits within the Commerce Department and works with industry to test advanced AI systems.

Read moreShow less

According to AP, the centre lacks a permanent director. In August, Congress began asking why it had stopped publishing reports on the safety of certain models. Public reports let researchers and policymakers see what was tested and which weaknesses those tests exposed.

Critics who spoke to AP worry that closer political contact could influence what the centre studies or publishes. Others say it could give the centre more influence. The White House responded to AP by pointing to the president's AI initiatives and left the specific questions about CAISSI unanswered.

Google DeepMind / Google Quantum AI

Google uses AI to find quantum programs with machine-checked proofs

Google DeepMind and Google Quantum AI used language model agents to find quantum programs that require fewer computing resources. Researchers wrote mathematical specifications for each task and used the Lean proof assistant to check whether the proposed programs met them.

Read moreShow less

The researchers built a library of components and specified what a program should calculate, how much error was acceptable and how to measure its cost. The agents searched for ways to combine the components and supplied proofs that Lean could check.

In one example, the researchers found a circuit for computing an inverse sine function with an estimated resource cost of about 1/290 of the baseline in the paper. They report smaller reductions for simulations relevant to fusion, photochemistry and drug metabolism.

The cost measure combines factors including the number of logical qubits a circuit needs and how many operations of a particular kind it uses. The figure describes how demanding the circuit is under the study's assumptions. It does not measure how fast it runs on a quantum computer.

The study is a preprint. Several simulations still require more resources than the first fault-tolerant quantum computers are expected to offer. The concrete contribution is that researchers can compare new circuits with machine-checked proofs of both their behaviour and their cost.

Video Rebirth

HyperFlow H3 Flash gets new model files for ComfyUI

On 9 November, Video Rebirth released INT8 weights that load directly in ComfyUI. They are for HyperFlow H3 Flash, a compressed derivative of MiniMax-H3 that generates video with synchronised stereo audio. You can start from one or two keyframes or use references to guide the clip.

Read moreShow less

The developer recommends the INT8 version for RTX cards with at least 24 GB of video memory and lists a minimum of 16 GB of system RAM for ComfyUI. Smaller GGUF files are an option when video memory is limited. The model card shows measurements on 24 GB and 32 GB cards and says 16 GB cards have yet to be measured.

The weights are distributed under the MiniMax-H3 licence. The model card says downloading or using them in the EU, UK, South Korea and US requires prior approval from MiniMax. The code for running the model is licensed separately.

Asahi Linux

Asahi Linux begins switching from the MIT licence to the GPL

Asahi Linux has begun switching its repositories to GPLv2 or later. The project develops Linux support for Apple Silicon Macs and previously used the MIT licence where possible. Asahi cites cases in which code under development was taken from developers' own forks and published without attribution, sometimes after being rewritten with language models.

Read moreShow less

The change has practical consequences for anyone building on GPL-licensed Asahi code. The GPL allows modification and redistribution, but requires that recipients get access to the source code and receive the same rights when a modified version is distributed.

Nous Research

Nous Research releases a Qwen variant fine-tuned for politically sensitive questions

Nous Research released the weights for Arete-27B on 6 November. The model is a fine-tuned variant of Qwen3.8-27B, trained to give more fact-based answers to politically sensitive questions about China. In the developer's own evaluation, the share of responses it classified as Chinese state propaganda fell from 22% to 0.9%.

Read moreShow less

The evaluation includes 803 questions across seven categories, including freedom of expression. Nous scored factual content separately from whether an answer adopted the Chinese government's account as its own. Quoting an official position and identifying it as such was therefore treated differently from presenting it as the model's own answer.

The figures apply to this question set and Nous's own scoring rules. The base model and Arete were compared with reasoning mode switched off. Nous published the questions, their underlying sources and the criteria used to score the answers.

Artificial Analysis

A new AI benchmark tests whether models know how confident to be

Artificial Analysis launched Decision Index v2 overnight into 9 November Norwegian time. It measures both whether a model makes the right choice and whether its stated confidence matches how often it is right. A model that is confidently wrong can therefore score lower than one that expresses appropriate uncertainty.

Read moreShow less

In the AA-Paperclip document test, the model makes five decisions and extracts ten facts per document. The test covers 1,500 résumés, invoices, receipts and research papers. All 15 answers must be correct for a document to count as fully solved.

Another test checks whether the model sticks to its decision when answer choices are reordered or relabelled. For someone using AI to route cases or select tools, these individual tests give more task-specific information than the aggregate score alone.

Google Cloud

Google announces two TPU systems for sensitive data and long contexts

Google announced two new systems built around its own AI chips on 9 November. TPU 8tx is designed to run and adapt models on sensitive data. TPU 8i Max is designed for tasks where a model needs to keep large amounts of text, code or video available at once.

Read moreShow less

The 8tx uses hardware protections to secure data while workloads run. For the 8i Max, Google put two 8i chips in one package. The aim is to provide more memory capacity and bandwidth for long-context workloads such as coding agents working across large codebases.

A separate part of the announcement covers software customers can already use. Cloud Pathways, which distributes training and inference across TPUs, became generally available on Google Kubernetes Engine in regions that offer supported TPU generations. Google also announced software that can move idle agents' working memory off the chips to free up capacity.