AI for Intelligence

EU court adviser says chatbot answers can breach publishers' rights

The signal

An EU court adviser sets two tests for AI answers built from news articles

What happened. On 3 November Advocate General Maciej Szpunar gave his opinion in Case C-250/25, the first generative AI case to reach the Court of Justice of the European Union (Court of Justice of the European Union, 2026). Like Company, a Hungarian web publisher, sued Google over answers from Gemini that drew on its articles live from the web. Szpunar set two tests. The first is whether the chatbot reached the article unlawfully, for example by going around a paywall or a block on scraping. The second is whether the answer reuses the article's wording so closely that the reader no longer needs the original. He called the questions about model training hypothetical in this case, but added that training on press content is, as a rule, allowed under the EU exception for text and data mining.

Why it matters. Facts stay free; the publisher's wording does not. If the judges follow the opinion, any tool that summarises live news, including an internal research assistant, carries legal risk when it bypasses access controls or replaces the source. The provider is liable when that reuse is built into its business model, or when it has no credible measures to stop it. The opinion is not binding, and the judgment will come later.

What to do. Ask whoever runs your news-monitoring or research tool whether it can reach paywalled content and whether its summaries reproduce whole passages.

Governors who want data-centre pauses won in Texas, Ohio and Georgia

What happened. Democrats won the governor's races in Texas, Ohio and Georgia on 3 November, and all three winners campaigned for moratoria or tighter rules on data centres, Bloomberg reports (Doan et al., 2026). Texas governor-elect Gina Hinojosa has promised to halt approvals until the state legislature passes new rules. SemiAnalysis, a research firm that tracks chips and data centres, says Texas holds nearly 30% of the US data-centre capacity due to come online in 2028.

Why it matters. The price of AI computing depends on how fast new capacity gets built. A pause in the state with the largest share would slow supply in 2027 and 2028, and slower supply tends to keep cloud AI prices high. Local opposition to data centres was visible for months before the vote, the kind of early sign I describe in where I look for weak signals.

What to do. Before you sign a multi-year AI or cloud contract, ask the provider where the capacity behind it is being built.

Tencent's Hy4 now reads images, and its output costs $2.50 per million tokens

What happened. Tencent released Hy4 on 4 November, the first model in its Hy line that understands images as well as text (Tencent Hy, 2026). It is available on Tencent's own cloud and on OpenRouter, a service that offers many models through one account, at $0.834 per million input tokens and $2.501 per million output tokens. Tokens are the small chunks of text that models are billed by. Tencent says Hy4 is among the strongest open-weight models, those whose files anyone can download, but still trails the best closed models on research-level maths and long coding tasks.

Why it matters. Google's Gemini 3.8 Flash costs $3.75 per million output tokens at its introductory rate (Google, n.d.-b), so Hy4 is about a third cheaper on output. Tencent plans to publish the weights with a system card on its safety tests, but has given no date, so for now Hy4 runs only on Tencent's and its partners' servers.

Try this today

Turn a scanned PDF into editable text and tables for free, with no sign-up

What it is: OXR is an open model from SoMark, with 1.3 billion parameters, that turns PDFs and document images into Markdown or JSON, the structured text formats AI tools read, and records where each element sits on the page.

Who it is for: Analysts and researchers who copy tables out of annual reports, patents, filings or scanned archives.

Access: Free. The online demo needs no registration and no API key; the code is on GitHub, with 3,308 stars on 4 November 2026.

First step: Upload one public report with a difficult table and compare the result with what you get from copy and paste.

Caution: The demo shows no terms for how uploaded files are handled, so use public documents only. SoMark will keep the free service online "until we can no longer afford to run it" (SoMark Team, 2026). Its benchmark rankings are its own. Commercial use is free only for organisations under $2 million in funding and under $2 million in annual revenue; above that, you need a separate licence.

(SoMark Team, 2026)

Run Gemini at low thinking first: it solved 50 of 50 tasks in Google's test

What it is: A Google Cloud engineer ran 50 coding-agent tasks and found that Gemini 3.8 Flash at its low thinking level, a setting that shortens how long the model reasons before it answers, solved all 50 (Villani, 2026). Capping the answer length saved nothing on easy tasks and broke 7 of the 25 hard ones.

Who it is for: Product owners and engineering leads who pay for the Gemini API in their products or internal agents.

Access: The Gemini API has a free tier and a paid tier; Gemini 3.8 Flash costs $0.75 per million input tokens and $3.75 per million output tokens at its introductory price until the end of the year (Google, n.d.-a; Google, n.d.-b).

First step: Ask the person who runs your Gemini integration to set the thinking level to low on one recurring task for a week, then compare the results with the default, medium.

Caution: The test was a single run of 50 short prompts with simple grading. On the free tier, Google uses your content to improve its products.

(Villani, 2026)

Run open models such as DeepSeek or Gemma in a private EU instance that keeps no prompts

What it is: Model Vault, Cohere's managed hosting for a single customer, now runs open models from other developers, such as DeepSeek V4.1 Flash and Google's Gemma 4, next to Cohere's own, and by default it retains neither prompts nor outputs (Cohere, 2026).

Who it is for: IT and compliance leads in banks, insurers, hospitals and public bodies that cannot send data to a shared AI service.

Access: Paid, priced per GPU-hour by type of chip, with promotional rates until 31 March 2027 on Cohere's pricing page; reserved capacity through sales (Cohere, n.d.).

First step: Ask your IT lead to name one AI use case that is stuck because the data cannot leave your control.

Caution: Model Vault runs in the US, the EU and Canada. The Encrypted tier, which keeps data encrypted even while the model works on it, is live only in the US, and it does not include DeepSeek.

(Cohere, 2026)

Worth watching

China says it will speed up a comprehensive AI law

On 3 November Wang Ruomeng of the National Development and Reform Commission, China's economic planner, said China will speed up work on foundational AI legislation and widen its rules on AI's core inputs and key uses (Qin, 2026). No draft or timetable has been published; a draft AI law has been on the State Council's legislative plan since June 2023.

Investors put $1 billion on agents that find why systems fail

Traversal, whose agents investigate outages in production software, raised $120 million at a $1 billion valuation on 3 November (Traversal, 2026). It says monthly investigations grew from 513 to nearly 73,000 this year. Watch whether incident diagnosis becomes the next job companies hand to agents after code writing; no date is attached. Yesterday's edition covered Anthropic telling investors its AI may soon speed up its own research.

The CERNY.AI read

Today's items are about what AI reads and what it runs on. The EU opinion says the wording of a news article belongs to its publisher even when the facts do not, and the US vote says the power and land for new data centres now depend on local politics. I read the Szpunar tests as a design brief for anyone who builds research tools: respect access controls, report the facts and send the reader to the source. Organisations in competitive and patent intelligence should audit their monitoring tools against those two tests now, before a judgment makes it urgent. Setting up that audit is part of my competitive intelligence consulting.

One question for your next meeting

If a publisher asked us tomorrow how our AI tools use its articles, could we show which sources they reach, how, and how much of the wording ends up in our reports? The CI maturity quick check shows whether your team has that answer documented.

Sources

  • Cohere. (n.d.). Pricing. https://cohere.com/pricing
  • Cohere. (2026, November 3). The world's leading open models, fully under your control. https://cohere.com/blog/open-source-model-vault
  • Court of Justice of the European Union. (2026, November 3). Advocate General Szpunar: The Court should find that the retrieval and reuse, in real time, of online press publications by AI chatbots in order to generate answers to user queries may, in certain circumstances, infringe the rights of press publishers (Press release No 155/26). https://curia.europa.eu/site/upload/docs/application/pdf/2026-11/cp260155en.pdf
  • Doan, L., Ford, B., & Eckhouse, B. (2026, November 4). AI data centre boom is set for upheaval after US midterm election rout. Business Standard. https://www.business-standard.com/world-news/ai-data-centre-boom-is-set-for-upheaval-after-us-midterm-election-rout-126110400718_1.html
  • Google. (n.d.-a). Gemini Developer API pricing. Google AI for Developers. https://ai.google.dev/gemini-api/docs/pricing
  • Google. (n.d.-b). What's new in Gemini 3.8 Flash. Google AI for Developers. https://ai.google.dev/gemini-api/docs/generate-content/whats-new-gemini-3.8
  • Qin, J. (2026, November 4). China to accelerate work on comprehensive AI legislation. Caixin Global. https://www.caixinglobal.com/2026-11-04/china-to-accelerate-work-on-comprehensive-ai-legislation-102500765.html
  • SoMark Team. (2026). OXR [Computer software]. GitHub. https://github.com/SoMarkAI/OXR
  • Tencent Hy. (2026, November 4). Introducing Hy4. https://hy.tencent.ai/research/hy4?langVersion=en
  • Traversal. (2026, November 3). We've raised $120M to close the outer loop of the SDLC with a superintelligent AI SRE. https://traversal.cv/blog/traversal-series-b-120-million-ai-sre
  • Villani, S. (2026, November 3). Cost per task: Using cheap classifiers to route across Gemini models and thinking levels. Google Developers Blog. https://developers.googleblog.com/en/cost-per-task-using-cheap-classifiers-to-route-across-gemini-models-and-thinking-levels/