Agency is the argument
The keynote organised itself around one word. Not safety, not alignment, not innovation, but agency: the capacity of people to decide whether, when and how a machine acts for them. Bogdan-Martin’s premise is that agency is not something technology possesses. It is something people grant.
The urgency she attached to it came from what has changed in a year. AI agents now conduct research and build software unattended, and their traffic has overtaken human activity on parts of the internet. She repeated the warning circulating among the people building these systems, that the next system humanity builds could be the last one it is able to control.
The question answered differently in Kiswahili
The most useful passage was a worked example rather than a principle. Two women ask an AI the same question about a symptom in pregnancy. Because AI systems underperform in underrepresented languages, the woman asking in Kiswahili may be told not to worry. Preeclampsia kills more than 70,000 mothers a year.
This is the language version of the argument that ran through the whole July cluster: a system is only as good as the data behind it, and the people missing from the data get the worse answer. Her counter-example was a person rather than a policy. Israel Adedolapo Adegoke, a young researcher and engineer in Lagos, built an offline speech-recognition system fine-tuned for Kiswahili that runs on low-resource devices, through a machine-learning challenge ITU runs with the African data-science platform Zindi. Her point was that AI agency does not require access to a frontier model.
The institutional version is the AI Skills Coalition, which ITU reports has grown past 70 partner organisations offering more than 180 courses in 13 languages.
Who is that agent, and who authorised it
The keynote previewed the most consequential thing the summit produced. If agents are going to buy, sell and act inside hospitals, banks and critical infrastructure, the system needs to know which agent acted and on whose authority. They have no passport and no phone number.
Her answer was to build what she called the identity layer of the age of artificial intelligence, and the machinery followed the next day: ITU announced a Focus Group on Trust and Identity for Humans and Agentic AI, reporting to its security standards group. The keynote closed on the optimistic formulation, which is also a conditional one.
Why it matters for the SDGs
The keynote touches SDG 3 (the maternal-health example), SDG 4 (AI skills), SDG 9 and SDG 10 (the 2.2 billion still offline and a developing world adopting AI at half the rate of the developed one). For a project about counting, the durable claim is the one about language: an SDG indicator can record that a woman has access to health information without recording that the information she received was wrong. Access is measured. Accuracy, by language, is not.
Watch & read
- AI for Good on YouTube, the source recording.
- ITU Focus Group on Trust and Identity for Humans and Agentic AI, announced 9 July 2026.
- Can AI help solve global crises?, the summit panel where the same divide is argued over.
- UN Global Dialogue on AI Governance · Full AI for Good coverage.
The summit publishes no official transcript. Quotations are lightly edited from an automated (Otter.ai) transcript of the recorded video and should be read as close paraphrase. The speaker is not named on the recording and is identified from her role and first-person references to ITU. The recording is undated; the date is inferred from her reference to the Global Dialogue having concluded the previous evening. The engineer’s name and the AI Skills Coalition figures were checked against ITU’s published materials, which give more than 180 courses where the keynote said more than 200.