2.2 billion people are still offline. A new UN commission wants to fix that. 🌿

Roughly a quarter of humanity, some 2.2 billion people, is still offline, cut off from the tools the rest of us spend our days arguing about. On July 2, a new body launched to keep that gap from widening. The AI for Good Global Commission, convened by the UN's ITU and co-chaired by Rwandan President Paul Kagame and Salesforce chief Marc Benioff, brings more than 40 founding members to one table ahead of the AI for Good Global Summit in Geneva.

What makes the roster unusual is who is on it. Heads of state from countries like Estonia, Namibia, Nigeria, and Singapore sit alongside the executives who actually build the technology, including Amazon's Andy Jassy, NVIDIA's Jensen Huang, Microsoft's Brad Smith, and Anthropic's Jack Clark. The mandate is narrow and concrete: write global standards that bridge digital divides so AI's economic gains reach developing countries rather than pooling in a handful of wealthy ones.

Most AI governance efforts gather either the builders or the regulators, rarely both, and seldom the countries deploying AI in public services on tight budgets. Rwanda's co-chairship signals that developing nations intend to help write the rules, not just receive them.

Commissions can become talking shops. The real test is whether the standards it produces get adopted where they matter, and whether 40 big names translate into funding and access for the billions who are not yet in the room.

Source: ITU, Axios

Eco Wave Power, working with NVIDIA, is using AI-driven digital twins to turn ocean waves into clean electricity, with a pilot at the Port of Los Angeles powering a data centre without drawing from the grid. The AI predicts when waves will be strongest and schedules compute around them. Wave energy is predictable and available on every coastline, a rare fresh angle on AI's energy problem.

Source: NVIDIA Blog

On July 1, the Department of Health and Human Services published its first AI strategy, aligning the FDA, CDC, NIH, and CMS around shared rules for patient safety, bias testing, and interoperability. The point is coherence. A clinical AI tool can now move from research to approval to reimbursement along one path, instead of navigating three separate and sometimes conflicting regimes.

Source: DistilINFO

Project Glasswing, a defensive security effort from Anthropic with AWS, Microsoft, and Google, scanned 1,000 open-source projects in its first month and flagged 23,019 real vulnerabilities, with a 90.6% true-positive rate on independent sampling. One was a credential-leaking flaw in the widely used Squid proxy that had survived 29 years of human audits. For once, the AI and cybersecurity story favours the defenders.

Source: Anthropic

What responsible government AI actually looks like

On June 29, California did something bigger than a routine technology purchase. It made Anthropic's Claude available to every state agency, and to any city or county that opts in, at a 50% discount through a central portal, bundled with free training for a workforce of roughly 300,000 and hands-on help from Anthropic engineers. Claude is already live at the DMV to cut wait times, at the Department of Health Care Services to support Medicaid case workers, and inside state cybersecurity teams scanning and patching code. A homegrown assistant called Poppy is heading for statewide rollout after a pilot across 67 departments. It is the largest US state-government AI deployment yet, and it is a useful case study precisely because government is where the stakes are legible: real people, real services, real accountability.

Our PM + Risk Manager lens

The telling design choice is that California bought a capability, not just a licence. Centralised procurement through one portal with transparent pricing solves a genuine adoption problem, since individual agencies rarely have the leverage or expertise to negotiate and deploy AI well. Bundling training and workflow consultation with access is the difference between shelfware and daily use. For anyone selling into large, risk-averse organisations, the lesson holds: the product is the tool plus the path to using it safely. Distribution and enablement, not just model quality, decide whether anything actually changes on the ground.

Concentrating a whole state's AI adoption on a single vendor carries obvious risks: lock-in, dependence, and one point of failure for services 39 million people rely on. The things worth watching are whether the portal stays genuinely multi-vendor over time, whether the state can audit and explain automated decisions that affect benefits or enforcement, and whether "human in the loop" survives the next budget crunch. The reassuring signals are the framing (assistive, not replacement), the pairing with a workforce-disruption plan, and California's existing procurement and equity guidelines. The unresolved one is transparency at scale.

The next 12 to 24 months

Expect other states to copy the shared-services model, because it is the fastest way to move public workers from banned-or-ignored to trained-and-supervised. The states that do this well will publish outcomes, wait times, error rates, and worker feedback, not just press releases. Honest optimism here means judging the deal by one question: does a real person's trip to the DMV or the Medicaid office actually get easier?

Source: Governor of California, TechCrunch

Mental health support is scarce almost everywhere, and nearly absent for adolescent girls in rural India. Wysa's Dreamkit takes a "phygital" approach: physical workbooks co-designed with local partners for classroom use, paired with clinically backed AI tools that offer private, on-demand support, delivered in Marathi rather than English. It has reached more than 250 girls' schools in rural Maharashtra, and Imperial College London is running a study on its effectiveness. What makes it a Good AI story is the restraint. It is not a chatbot pretending to be a therapist, but a carefully designed programme that puts AI where a counsellor cannot be, in a language students actually speak.

→ Read more: ImpactAlpha

Who should decide how AI reaches the communities that don't have it yet?

Hit reply. We read every response.