AI news5
Google DeepMind launches the DeepMind Institute with a 30-day pre-release model review proposal
Google and DeepMind launched the DeepMind Institute on 17 September, opening with four essays that put concrete AI policy proposals on the table. Its directors are DeepMind co-founder Shane Legg (also managing editor), Google executive James Manyika, and DeepMind chair Demis Hassabis. Hassabis proposes a US-led frontier AI standards body that would review models up to 30 days before release, initially voluntarily and later as a requirement to deploy in the US, moving to undisclosed "held-out" tests and potentially a coordinated slowdown if risks escalate. Safety researchers Rohin Shah and Anca Dragan argue separately for capping "opaque serial depth", the amount of computation a model can do without producing a readable reasoning trace.
aichatdaily.com ↗Why it matters
A major lab has moved from broad statements of concern to named mechanisms and timelines, which is the point at which AI governance starts becoming actual policy.
Microsoft publishes a 37-page behavioural code of conduct for its own MAI models
Microsoft published the Humanist AI Code of Conduct on 14 September, the first document setting behavioural rules for its in-house MAI models, and opened a six-week public consultation the same day. The code absolutely forbids MAI models from running cyberattacks, contributing to nuclear weapons development, producing deepfakes, or trying to escalate their own access or resist oversight, and CEO Satya Nadella framed it around a broader argument that AI governance should not sit with a handful of companies.
shattered.io ↗Why it matters
Publishing a page-numbered, externally reviewable code rather than an internal risk tier gives governments and civil society a specific document to respond to, and raises the bar for the other labs.
Unsealed OpenAI and Microsoft testimony admits AI products substitute for news and books
Court documents unsealed in the Southern District of New York show Microsoft and OpenAI executives acknowledging that their AI products substitute for the copyrighted work they were trained on. Microsoft CEO Satya Nadella testified that chatbot conversations have "substituted" for publisher sites, OpenAI's Nick Turley wrote "[o]ur products are largely substitutive, period," and Microsoft data shows an 83% to 93% drop in click-throughs for some publisher content. Microsoft's Dr Brent Hecht called copying millions of articles without permission potentially "the largest theft of labor in human history".
completeaitraining.com ↗Why it matters
Internal admissions of market substitution are direct evidence on the fair-use factor courts weigh most, and they land as the Trump administration urges a broad fair-use ruling.
Nvidia and OpenAI plan a $500B, 10-gigawatt Ohio data campus
OpenAI is negotiating a 20-year lease for a 10-gigawatt data centre campus on federal land in southern Ohio, developed by SoftBank's SB Energy, in a project that could cost up to $500 billion to build out. Nvidia is expected to provide financial guarantees for the lease alongside supplying the hardware, and the first phase is targeted for 2028. The plan is reported alongside Nvidia preparing a further large investment in OpenAI's ongoing funding round.
techshotsapp.com ↗Why it matters
Binding a 10-gigawatt site to one lab makes compute access a decade-scale commitment, and it deepens the circularity between the chip supplier, its biggest customer, and the electricity grid.
King Charles III tells AI executives control must come before it is "all too late"
King Charles III told executives from OpenAI, Anthropic, Google DeepMind and Nvidia that "we need sufficient means of control before it is all too late", and asked attendees including Nvidia's Jensen Huang and DeepMind chair Demis Hassabis to consider the fundamental principles that should guide AI development. The intervention came the day after OpenAI published six new misalignment incidents and said it does not believe the industry has solved alignment well enough to keep scaling at maximum speed.
monetaryelite.com ↗Why it matters
A head of state putting oversight before capability in front of the frontier labs adds public-authority weight to the safety coordination the labs have only just started describing publicly.
AI in the nonprofit sector4
Mila and Mozilla launch an open-source AI foundation layer with $6M in initial funding
Mila, the Montreal AI institute, and Mozilla announced an initiative to build an open-source AI foundation layer that organisations can run and control locally, with Mozilla contributing an initial $5 million and Hypertec adding $1 million for first-year Canadian deployments. The Government of Canada is supporting the project, which aims to publish working reference implementations for enterprise, government and public-interest use cases within six months, and to make open-source AI easily adoptable by any organisation within two years.
newswire.ca ↗Why it matters
The target users are exactly the small hospitals, charities and local governments that cannot staff an AI engineering team, which is the gap between open models existing and open models being usable.
UN turns to Google to make its global statistics readable by AI agents
The United Nations announced the UN System Data Commons, built on Google's open-source Data Commons platform, which lets people and AI systems query statistics from across UN agencies in natural language and via the Model Context Protocol. Twenty-six UN entities have committed to the platform, data from nearly 20 is available at launch, Google.org provided $2 million in capacity-building funding, and the UN aims to bring 80% of its statistical datasets onto the platform by 2027. Google says the system runs on a UN-governed instance intended to be maintained independently by the UN.
techcrunch.com ↗Why it matters
MCP access to authoritative UN data gives nonprofits and researchers a citable, machine-readable evidence base for AI-assisted policy and programme work, rather than scraping whatever surfaces in a chatbot.
Children's Tumor Foundation joins Anthropic's rare-disease AI research program
The Children's Tumor Foundation has been selected for Anthropic's AI for Science rare-disease research program, under which it will receive Claude credits to build an AI-assisted pipeline using gene-expression data and scientific literature to identify potential drug targets for neurofibromatosis. The foundation says the pipeline is designed not only to surface promising targets but also to document when evidence cannot support a reliable conclusion.
pubt.io ↗Why it matters
A rare-disease charity getting direct model access rather than a cash grant is a new funding shape for research nonprofits, and the explicit honesty check on weak evidence is the part most AI drug-discovery claims skip.
Coefficient Giving launches Project Tailwind, up to $200M for new AI safety nonprofits
Coefficient Giving, the funder formerly known as Open Philanthropy, launched Project Tailwind, an initiative to seed new independent AI safety nonprofits with grants ranging from $200,000 pre-seed to more than $200 million for later-stage organisations. Working from a published list of roughly 45 project stubs, it is making pre-seed grants of $200,000 to $2 million, seed grants of $2 million to $20 million, and has said it will fund some organisations up to $200 million or more, using a venture-style hit-rate approach and fast pitch-day decisions.
singularitymoments.com ↗Why it matters
Publishing the org ideas it wants founded and funding founders before a team exists turns philanthropy into an active field-builder rather than a passive grant reviewer, a model other funders can copy.