AI Governance Review May 2026 Top 10
Here are the Top 10 stories that I found particularly noteworthy in May 2026.
🚫 Top 1 – The Nudification Ban
Of all the changes contained in the EU AI Act’s simplification package, one made headlines far beyond the usual AI governance audience: the ban on ‘nudification’ apps – tools that generate non-consensual intimate images using AI.
Several high-profile incidents in 2025 and early 2026 – involving schools, public figures, and private individuals – had kept the issue in public view. The political agreement of 7 May 2026 formalised what many had argued was already prohibited under existing rules, but added explicit clarity and visibility.
This specific detail demonstrates a pattern: concrete harms, identifiable victims, and public outrage move AI regulation faster than abstract risk frameworks ever will. The nudification ban is a reminder that the EU AI Act can be a living response to what AI does to real people.
🔗 EU agrees to simplify AI rules and ban nudification apps
🔧 Top 2 – The EU AI Act Gets a Makeover
The nudification ban was just the most visible element of a larger package. Three interconnected developments – all part of the Digital Omnibus simplification agenda – matter directly for deployers and providers.
The political agreement set new implementation timelines: high-risk AI in biometrics, critical infrastructure, education, employment, and border control from 2 December 2027; systems integrated into regulated products from 2 August 2028. The logic: let technical standards and compliance tooling catch up before obligations bite.
Two sets of draft guidelines were also published: one on Article 50 transparency obligations (chatbots, emotion recognition, disclosure requirements) and one on high-risk AI classification. The classification guidance is particularly welcome – it addresses the uncertainty that has made AI inventory work so frustrating in practice.
The ‘simplification for competitiveness’ framing deserves scrutiny. Reducing bureaucratic redundancy is legitimate. Reducing substantive protection is not. The line between the two is where the real debate lies.
🔗 Draft Guidelines on Article 50 Transparency Obligations
🔗 Draft Guidelines on High-Risk AI Classification
💼 Top 3 – The Job Apocalypse Debate
Ezra Klein’s May op-ed in the New York Times pushed back on the AI-will-destroy-all-jobs narrative – not by dismissing the concern, but by presenting the strongest counter-evidence: GDP growth remains unspectacular, unemployment is low, and even automation-exposed sectors are still hiring.
His argument: AI is likely a general-purpose technology whose economic effects unfold over decades, unevenly, sector by sector – consistent with the pattern of previous transformative technologies. Not a sudden cliff. A long slope.
For governance practitioners, the value of this piece is not the conclusion but the reframing: away from apocalypse narratives and towards the harder questions of who benefits, who adjusts, at what pace, and with what policy support. Those questions are less dramatic – and more actionable.
🔗 Why the A.I. Job Apocalypse (Probably) Won’t Happen — NYT
⚖️ Top 4 – Claude for the Legal Industry
Anthropic launched Claude for the Legal Industry in May – 20+ MCP connectors linking Claude to core legal software, plus 12 plugins tailored to specific practice areas. A significant step in the industrialisation of legal AI, and a catalyst for questions the profession has been deferring.
Liability: when AI-assisted legal work goes wrong, who is responsible – the lawyer, the firm, the vendor? Professional ethics: bar associations are still developing guidance on AI use; competence obligations, confidentiality, and supervision duties are all in tension with the efficiency promise. Access to justice: if AI reduces the cost of legal work, does that democratise access – or primarily benefit those who can already afford sophisticated services?
🔗 Claude for the Legal Industry
🌍 Top 5 – Little Attention but SUCH a Large Step: The First Legally Binding International AI Treaty Ratified by the EU
The Council of Europe’s Framework Convention on Artificial Intelligence – the first legally binding international AI treaty – reached another milestone in May with the EU’s ratification.
The Convention covers the full AI lifecycle and requires parties to align with core principles: human dignity, non-discrimination, privacy, transparency, accountability, reliability, and safe innovation. It establishes procedural rights for affected persons, mandates risk and impact assessments, and allows for bans or moratoria on specific AI applications.
What it adds beyond the EU AI Act: geographic and institutional scope. Negotiated with 46 CoE member states plus observer states including Canada, Japan, the US and others, it creates a monitoring mechanism – the Conference of the Parties – to assess implementation over time.
🔗 The Framework Convention on Artificial Intelligence — Council of Europe
🔬 Top 6 – Can We Govern AI Together?
The UN Office for Digital and Emerging Technologies launched the AI Governance for Humanity Lab – a new initiative premised on a straightforward observation: AI governance decisions are increasingly consequential for all countries, yet too often shaped without universal participation, shared evidence, or common standards.
The Lab sits alongside the Independent International Scientific Panel on AI and the Global Dialogue on AI Governance, both established through UN General Assembly Resolution A/RES/79/325. A practitioner survey – aimed at mapping the gap between AI governance policy intent and implementation reality – closed yesterday. That intersection is exactly where the most interesting tensions live, and I am hoping the results are published soon.
🔗 AI Governance for Humanity Lab — UN Office for Digital and Emerging Technologies
🤝 Top 7 – AI as a Teammate, Not Just a Tool
Capgemini colleagues published a Harvard Business Review piece on what they call Human-AI Team Chemistry – a capability that, their five-month study shows, does not emerge naturally and must be deliberately built through three practices: engage with AI as a team (not just individual users); leverage its role fluidity by assigning it multiple personas to deepen discussion; and maintain collective ownership by jointly shaping prompts, debating outputs, and critically evaluating contributions.
Reading this made me reflect on how much I am, in my own AI work, grounded in the human side of the “human AI chemistry” equation. It also made me think: treating AI as a separate meeting attendee does seem immediately convincing for (re)establishing discipline – but it risks sliding into anthropomorphisation. There is a real difference between using AI well in a team context and treating it as a colleague with agency. For me this is a recommendation requiring a priorly established rather high level of AI Literacy: Not every team will have the critical literacy to hold the agency line.
🔗 It’s Hard to Use AI as a Team. These 3 Practices Can Help. — HBR
🛡️ Top 8 – Sayash Kapoor and Arvind Narayanan on the Cost of Resilience
The authors of AI as Normal Technology – responded to Derek Thompson’s call for ‘extraordinary’ government intervention on AI risks. Their counter: extraordinary intervention is costly, brittle, and often less effective than the alternative: investing in societal resilience.
Nonproliferation relies on a single chokepoint; resilience distributes defences across society. Sclerosis in government and the political ease of targeting AI companies make extraordinary intervention feel appealing – even when proportionate, adaptive approaches would serve better.
The resilience lens is underused in AI governance. The field tends to oscillate between ‘regulate harder’ and ‘let it run’. Building adaptive capacity in institutions, regulatory design, and civil society is unglamorous – but it may prove more durable than any single intervention and clearly provides an interesting alternative perspective.
🔗 Do AI Risks Require Extraordinary Government Intervention? — Knight First Amendment Institute
✝️ Top 9 – The Pope Weighs In
Who would have thought I would be quoting the Pope in an AI governance review? On 15 May, Pope Leo XIV issued Magnifica Humanitas – on “safeguarding the human person in the time of artificial intelligence”.
Beyond being a well-structured and genuinely readable document, the encyclical captures something surprisingly difficult to articulate in technical governance discourse:
“We cannot consider AI to be morally neutral. In reality, every technical tool embodies choices and priorities through what it measures, ignores and optimizes, and how it classifies people and situations.”
This is not a new insight for those in AI ethics – but it is rarely stated so clearly, or with such reach.
🔗 Magnifica Humanitas — Encyclical Letter of Pope Leo XIV (15 May 2026)
🤖 Top 10 – The WEF Agentic AI Playbook
The cherry on the cake of the month of May: the World Economic Forum published AI Agents in Action: A Playbook for Trusted Adoption, Authorization and Scaling on 26 May – co-authored by Capgemini France colleagues.
The Playbook introduces the Agent Capability and Authorization Profile (ACAP) – a deployment-level governance and authorization instrument for agentic systems. The core problem it addresses: organizations deploying multiple agents that share a foundational model face systemic vulnerabilities across their entire agent estate. ACAP provides a structured way to define, enforce, and monitor the conditions under which agents are authorized to act.
Given how much we currently talk about Agentic AI Governance, this is a welcome and timely contribution. Practical, grounded, and directly applicable. Merci @my colleagues from Paris!
🔗 AI Agents in Action: A Playbook for Trusted Adoption, Authorization and Scaling — WEF
That’s a wrap for May 2026. Happy to discuss any of these — the governance landscape moves fast and thinking through it together makes it more navigable.
— Jana Schöneborn | AI Governance Review | May 2026
