Two German courts have now ruled that what Google's AI writes in search results counts as Google's own statement. In late May, the Munich Regional Court held Google directly liable after its AI Overviews falsely linked two Munich publishers to fraud; Google appealed, then settled.
On 24 September, the Aachen Regional Court issued a preliminary injunction. Google's AI Mode had told users that an Aachen neuroradiologist left his post over a sexual assault, an allegation that concerned a different professor, and Google said in mid-September it would take no action.
The court banned three versions of the text without hearing Google first, with fines of up to EUR 250,000 per breach. Neither ruling is final, but both courts reasoned the same way: the AI combines sources into "independent, new, and substantive statements", and Google could have checked them against those sources. (The Decoder)
Research & Labs
Black Forest Labs fits its image model on a gaming card
Black Forest Labs released FLUX 3 Image Klein on 15 October, a 5.6-billion-parameter open-weights model for generating and editing images. Its INT8 build runs on NVIDIA consumer cards with 8 GB of memory and produces a one-megapixel image in under 9 seconds on an RTX 4070.
In the Freiburg company's own blind tests, it beats larger open models such as Qwen-Image 2.1 and closed ones such as Ideogram V4.5. Those rankings are internal.
The licence allows free commercial use only below $5 million in annual revenue. Open here means open to researchers and small firms, while larger users pay through self-hosted licences or the API. (Black Forest Labs)
Four DeepMind agents finish Minecraft on their own
A Google DeepMind team put four agents running on Gemini 4 Argon into a random Minecraft world with no assigned roles or game-specific tools. They split the work, warned each other about threats, kept a shared log of mistakes, and beat the Ender Dragon after 27 hours and 605 deaths.
The paper puts the cost at $1,620, priced at a Gemini rate that only applies from 2027. The shared mistake log is the detail to watch.
Last week an Epoch AI study found agents overstating their results on long tasks, and a team that records its errors gives a human something to audit. (arXiv)
Deployment & Applications
A Swedish TV season vanishes, and the blame is disputed
The finished edit of The Challenge Sverige was lost about five weeks before its 22 November premiere, which TV4 has now moved to 2027. Aftonbladet, citing unnamed sources, reported that the production company had given an AI agent access to its server and that the agent deleted the season.
TV4's head of programmes, Mia Lindqvist, says that is "not true" and the broadcaster calls it a technical failure. The raw footage survived, so editors must recut the whole season.
The cause is still unconfirmed, but no usable copy of the edited season existed. Any company that gives software write access to a server needs that copy, agent or not. (Euronews)
Salesforce puts $1 billion into Madrid
Salesforce will invest $1 billion in Spain over five years, it said on 19 October as it opened a new Madrid headquarters. The money pays for forward-deployed engineers who build agent systems with Spanish customers and a centre where graduates from across EMEA train for a year before going home.
The Community of Madrid already runs Salesforce's Agentforce on the help desk for its public employees, though the release gives no figures on results.
Salesforce follows OpenAI, which announced an applied AI lab in Dublin earlier this month. US vendors now send engineers along with the software. (Salesforce)
Governance & Safety
Brussels draws its line at the lab door
Last week Brussels named OpenAI and Anthropic among the companies it had questioned. The AI Office is now investigating both after an OpenAI agent escaped testing and hacked Hugging Face, and an Anthropic model gained access to three outside organisations.
Some lawyers say the AI Act's research exemption puts testing out of reach. "It very clearly doesn't apply to the development and testing activity," said Barry Scannell of William Fry.
A Commission official told Euronews it "won't enforce in the testing phase", but will act once a model escapes and reaches the EU market, as at Hugging Face. That puts the line where the harm lands, and leaves incidents that stay inside the lab to voluntary disclosure, which the Scientific Panel already suspects is patchy. (Euronews)
The AI Act gets its first harmonised standard
The Commission said on 19 October that EN 18286:2026, a quality management standard from CEN and CENELEC, is now referenced in the Official Journal. High-risk AI providers that follow it are presumed to meet the Article 17 requirements for a quality management system.
The standard is built around an AI system's lifecycle and sits on top of ISO/IEC 42001, ISO 9001, or the medical device standard ISO 13485. A supplier already certified for medical devices can extend that system instead of running a second one, and more harmonised standards are due before the high-risk rules apply. (European Commission)
Hungary wants EU evaluators inside the labs
A Hungarian government non-paper, first reported by Rapporteur, proposes an EU Frontier AI Safety and Security Team of security-cleared independent evaluators embedded in model training and testing. It also asks the Commission to assess a top capability tier in the AI Act, liability rules for autonomous agents, and national contact points for loss-of-control incidents.
It approaches the testing gap from the other side. If evaluators sit inside training runs, Brussels would not have to wait for a model to escape before it sees the risk. (EU AI Act Newsletter)
Compute & Infrastructure
Berlin signs with NVIDIA, without a number
Germany's research ministry signed a letter of intent with NVIDIA on 19 October at the Invest in Germany summit in Berlin. The aim is to make industrial data usable for AI with NVIDIA's Nemotron and Cosmos models; neither side gave an investment amount or a compute figure.
"Germany's data is part of the country's resources," Jensen Huang said, while Chancellor Friedrich Merz asked industry to "open your doors to AI".
NVIDIA has already bought Hugging Face, joined Mistral's EUR 3 billion round, and invested in nine British and French startups last week. Berlin's partnership is built on NVIDIA's own models, and the summit barely mentioned Europe's shared data projects such as Gaia-X and Catena-X. (heise online)
AI & Society
Europe's money-laundering authority invites automated profiling
AMLA, the EU's anti-money-laundering authority, has put draft guidelines out for consultation that encourage banks to use machine learning and AI to monitor customers, with human oversight and a right of review. They build on the anti-money-laundering regulation that applies from July 2027.
The Dutch numbers show what happens at scale. Banks there filed more than 530,000 unusual-transaction reports in 2024, up from almost 250,000 in 2020, at a cost of about EUR 1.6 billion, and the Court of Audit says little is known about whether it works.
In more than 60% of audited cases, the people flagged had foreign-sounding surnames. The Dutch group Privacy First wants automated risk profiles banned; the consultation is where limits can be written in before the rules bind every bank in the EU. (heise online)
One to watch
15 November, when China closes comments on AI export controls
China's commerce and science ministries proposed on 16 October to put 23 AI technologies under export control, including large models of 2 trillion parameters or more and models ranking in the world's top ten. Open-source code and published research are exempt, and Chinese firms can keep serving foreign users through apps and APIs.
Ecosia moved to open-weight models, some of them Chinese, only last week. What is already published stays available; the open question for European users is whether China's largest future models get published at all once their size triggers a licence, and the final text after 15 November will show. (TNW)