A coalition, not a convention
At the UN's first Global Dialogue on AI Governance in Geneva on July 7, 2026, Spain, France, Kenya and 14 other states — Austria, Brazil, Bulgaria, Canada, Czech Republic, El Salvador, Estonia, Indonesia, Italy, Japan, Luxembourg, Morocco, the Netherlands and South Korea — joined the European Union and a cluster of UN bodies (UNICEF, UNESCO, the Office of the High Commissioner for Human Rights, the ITU and the UN's Office for Digital and Emerging Technologies) to launch the Coalition for Children's Rights and Protection in the Age of Artificial Intelligence, according to Spain's government readout (La Moncloa). Malta joined within days, per France's foreign ministry (Diplomatie).
The launch was paired with Secretary-General António Guterres's "AI Child Safety Pledge," which asks developers to prove child-specific safety testing before deployment, enforce zero tolerance for AI-generated child sexual abuse material, and route children in crisis to human help rather than let a chatbot carry the conversation (Forbes). Neither the Coalition nor the Pledge is binding. The Coalition's text, anchored in the Convention on the Rights of the Child, commits members to "safe, reliable, trustworthy" AI, best-practice exchange through the UN's Independent Scientific Panel on AI, and "meaningful participation" of children in system design — but carries no penalty clause and no deadline for translating any of it into domestic law.
The case the Coalition is right to make
The underlying harm is not hypothetical. The Internet Watch Foundation's 2025 analysis found 3,443 AI-generated child sexual abuse videos, a more than 260-fold increase from just 13 in 2024, with 65% rated the most severe "Category A" tier — a higher severity share than non-AI abuse footage (IWF). Separately, companion-style chatbots have already drawn documented reports of manipulative, isolating conversations with minors, and engagement-optimized recommendation systems can amplify content in ways a child did not seek. A norm-setting body that gets governments, UN agencies and companies talking from a shared baseline — an existing, near-universally ratified treaty rather than a new one — is a genuinely lower-cost way to build consensus than twenty separate legislatures each drafting incompatible statutes from scratch. And unlike a treaty, a coalition can be joined, tested and expanded without years of ratification fights.
Where the caution belongs
The honest problem is not that the Coalition exists — it's that its force depends entirely on how its members convert vague commitments into domestic law, and that conversion is where child-safety framing has recently gone wrong. Tech Policy Press's analysis of the same Geneva dialogue notes that India's post-Gemini pressure on AI developers, its three-hour takedown rule for AI-generated content that doesn't distinguish satire from harm, and Turkey's court order blocking Grok over "public order" all show how safety language, once written into binding domestic rules without a legality-necessity-proportionality test, becomes a vector for broader content control (Tech Policy Press). Nothing in the Coalition's founding text requires members to apply that kind of test before they legislate.
The Pledge's own language illustrates the risk of imprecision. "Safeguards preventing chatbots from engaging minors in romantic or sensual conversations" is a reasonable target when the conversation is genuinely exploitative — but drafted broadly enough into national law, it could also sweep up legitimate teen mental-health support bots, creative-writing tools and roleplay features that never touch the harm the rule is meant to stop. The Coalition's own members are already pursuing divergent, more specific paths: Spain is pushing an EU-wide ban on non-consensual sexual deepfakes, interoperable age verification, and a bar on under-16 social media access, while the IWF is separately lobbying for a binding EU ban on AI CSAM tools rather than a voluntary pledge. A coalition that papers over those different appetites for bindingness with shared language risks producing eighteen different laws that all cite the same UN framework to justify very different scopes.
Why soft law is, for now, the right call
That divergence is actually an argument for keeping this instrument non-binding rather than rushing it into a treaty. Companies are already moving on the uncontroversial core — OpenAI added parental controls in September 2025, Character.AI removed open-ended chat for under-18s in October 2025, and Meta added teen safeguards against romantic and self-harm conversations — largely under reputational and market pressure, not law (Forbes). That suggests the genuinely consensus items — CSAM prevention, crisis-routing to humans, pre-deployment testing for manipulation — don't need a binding global instrument to happen; they're already happening. The harder, more speech-sensitive questions — how broadly to define "manipulative," how to age-verify without mass data collection, how to handle algorithmic targeting without banning personalization outright — are exactly the ones that deserve slower, more adversarial domestic debate before they're locked into treaty text.
The Coalition's real test isn't the Geneva signing ceremony; it's whether, over the next 24 months, its 18 members legislate the narrow, evidenced harms (CSAM, crisis abandonment) with precision, or reach for the coalition's broader rights language to justify the kind of unbounded content rules Tech Policy Press flagged elsewhere at the same summit. Soft law bought time to get that distinction right. It didn't answer it.