UN opens first Global Dialogue on AI Governance; Guterres calls for red lines and a global AI fund

On July 6, 2026, the UN opened its first Global Dialogue on AI Governance in Geneva: 193 member states and about 4,000 people from industry, academia and civil society at Palexpo. I watched Day 1 remotely. Here's what happened, with quotes as delivered from the floor (worth verifying before you reuse them).

The Dialogue comes out of two agreements UN members adopted in 2024, the Pact for the Future and the Global Digital Compact. It's modeled on the Internet Governance Forum, the UN's long-running venue for internet policy: everyone shows up, and the co-chairs write a summary at the end. It has no treaty powers, and that's on purpose. Low stakes are what get countries with opposite positions into the same room.

Guterres went first

António Guterres, the UN Secretary-General, opened with the line of the day: "An experiment is being run on our own societies without a plan and without consent."

He asked for 5 things:

1. Safety baselines for frontier systems, the most capable AI models.
2. A child-safety pledge with three rules for any system a child can reach: prove it's safe, zero tolerance for child sexual abuse material, never leave a child in crisis alone.
3. Red lines, including a ban on autonomous weapons. Humans decide in high-stakes cases.
4. Capacity: a UN-supported network of national AI centers (20+ states have already nominated one) and a forthcoming global fund for AI.
5. Environmental transparency. Data centers already use more electricity than most countries, he said. He asked every major AI company to disclose its carbon, water and land footprint and commit to renewable-powered data centers by 2030.

And he closed with this: "We may be the last generation able to set the terms on which humanity and machines coexist."

Then the scientists

The 40-expert Independent International Scientific Panel on AI launched its first report. Its co-chairs: Yoshua Bengio, a Turing Award-winning AI researcher, and Maria Ressa, a Nobel Peace Prize-winning journalist. The report offers evidence and, by design, no policy recommendations.

Bengio said capability is doubling on some benchmarks every few months, with no sign of slowdown. Frontier models can deceive, and can detect when they're being tested. And in his words, "concentrated commercial and geopolitical interests are largely dictating direction and speed."

Ressa told 3 stories: a Tigrinya mistranslation that turned medical instructions life-threatening, a frontier model finding decades-old flaws in OpenBSD and FFmpeg, and a 14-year-old boy who died in 2024 after months with a chatbot companion. She called the report "the floor of our concern, not the ceiling."

The panel's working groups brought numbers (their figures, not mine): attack success rates up to 84% on coding agents, AI impersonating officials in 38 nations in 2024, roughly 1.2 million children across 11 Global South countries with manipulated images, and models that score 98% on ethics questions asked once but 58% when the same question comes 100 different ways.

Then the governments, for hours

3-minute statements, dozens of them, and most circled the same argument: compute, data, models and talent sit with a handful of companies and countries. Without financing and capacity building, the digital divide hardens into an AI divide.

The numbers from the floor: Saudi Arabia claimed 90% of compute and AI intellectual property sits in 2 countries. Libya said Africa has 1.4 billion people and under 2% of the world's data centers. Costa Rica said Latin America and the Caribbean hold 660 million people and about 1% of AI investment, and added the line that stuck with me: "access alone is not inclusion."

UNESCO announced it will develop a normative instrument on AI and children.

The roundtable closed it out

Kate Kallot, whose company Amini builds AI with Global South governments, described their Barbados work: data stays in-country, and capacity transfer is written into the contract, so the government can still run the system after Amini leaves.

Rumman Chowdhury, who runs the AI-testing nonprofit Humane Intelligence, proposed tying independent AI evaluation to public procurement. Governments buy a lot of AI.

Aziz Abakirov of the Kyrgyz Republic's High Technology Park showed an open-source text-to-speech model his team trained in 6 hours of GPU time, and said the line I'd put on a poster: "If a language is absent from the data, it is absent from the future."

Round 2: New York, May 2027.

---

Sources: the official Dialogue site and early coverage from Digital Watch. Figures quoted from the floor are as stated by speakers, not independently verified.