.png){kind=logo}
May 2026
This article was originally published in Decoding, our monthly briefing on the latest trends in government technology. Sign up here to receive future editions directly in your inbox.
.png){kind=logo}
Artificial intelligence is reshaping how governments operate, how public services are delivered, and how democratic accountability is exercised. Most public sector organisations are somewhere on the journey from pilot to implementation – under pressure to move faster, adopt more, and demonstrate results.
This edition of Decoding asks a question that lies beneath most of the decisions being made: what do we actually mean when we say AI? The answer matters for governance, for procurement, and for the kind of future public institutions are building toward. A recent Foreign Policy analysis argues that governments cannot coordinate on AI internationally simply because they do not agree on what it is.
In this edition, you’ll read about:
While many countries have called for some form of international coordination on AI, few have achieved anything more than voluntary commitments. The conversation on international AI governance remains deeply fractured. A recent Foreign Policy analysis argues that the most underappreciated reason for this is not political disagreement but something more fundamental: governments cannot coordinate on AI because they do not agree on what AI actually is.
The definitional problem is the most visible layer: some people use the term artificial intelligence to refer to tools like ChatGPT and other large language models (LLMs), while others use it to describe algorithms that have been quietly embedded in public services for years. Others still use the term to refer to superintelligent systems that exceed human capabilities, but that do not yet exist. Trying to govern AI without resolving this is like trying to regulate vehicles without distinguishing between a bicycle and a truck, as computer scientists Arvind Narayanan and Sayash Kapoor have put it. The ambiguity is not just academic; it shapes public discourse about AI and ultimately affects discussions about which specific technologies get regulated, by whom, and in what way.
The deeper disagreement is about pace and scale. Some, including leading AI developers, believe transformative, potentially civilisational capabilities could arrive within years. Others expect AI to have a significant but slower, more sector-specific impact, playing out over decades rather than overnight. Both camps agree AI will matter, but they disagree fundamentally on when and how much.
The disagreement about pace combines with a second question: how much of our own AI infrastructure do we actually control? The United States and China together account for around 90 per cent of global AI compute and most of the world's leading models. Most other governments are, to varying degrees, dependent on one or both of them.
The Foreign Policy analysis maps these two questions onto a framework: on one axis, how a government perceives the speed and scale of AI's transformation; on the other, how self-sufficient it believes its domestic AI capabilities to be. Where a country sits on those two axes, the piece argues, explains more about its AI policy choices than political ideology or national values. How a government sees AI determines how dependent it is willing to be – and on whom.
The decisions being made right now at the global and national levels about AI procurement, infrastructure, and regulation all contain implicit answers to these questions. The assumptions are rarely made explicit, but making them visible is one of the more important things public sector leaders can do before signing the next AI contract. The Foreign Policy analysis concludes that until governments converge on these differences, the prospects for meaningful international coordination are dim, and it describes the current trajectory as a dying international ecosystem for multilateral AI governance. Whether that assessment proves correct remains to be seen.
On 7 May, the European Parliament and the Council reached a political agreement on the AI Act Omnibus, a significant simplification of the original AI Act, which the Commission proposed only five months ago.
The agreement is more than an administrative simplification. It reflects the EU’s view of what AI is and how fast it is moving. The EU is treating AI as an important technology that requires careful governance, as an infrastructure to be regulated deliberately, not a race to be won.
The most significant change is a clearer implementation timeline. High-risk AI systems, covering biometrics, critical infrastructure, education, employment, migration, asylum and border control, will face binding rules from December 2027. Systems built into products such as lifts or toys follow on 2 August 2028. The sequencing gives time for technical standards and guidance to be in place before the rules take effect.
The definition of what counts as high-risk has also narrowed. Only AI systems where failure could cause genuine harm to health or safety face the heaviest requirements; tools that help users or improve performance no longer automatically fall under the full framework.
One change has drawn more scrutiny. Machinery has been carved out of the AI Act entirely and placed under its own sector-specific regulation, a move welcomed by manufacturers but viewed by some MEPs as the beginning of regulatory fragmentation.
The Omnibus is widely covered as a simplification story, and it is, but beneath the technical adjustments lies a harder question that the agreement does not resolve: can any law keep pace with AI? The core framework remains intact, but the deal sets a precedent: the rulebook can be reopened. The next test is whether the AI Office and member states enforce what remains, or whether delays and carve-outs quietly hollow out the law's initial ambitions.
The Danish Government, Local Government Denmark and Danish Regions have, through the Digital Taskforce for Artificial Intelligence, agreed to roll out three large-scale AI projects across the state, municipalities and regions from 2026. The three projects cover automated documentation for caseworkers, nurses, and doctors; digital assistants for public-sector employees; and AI-powered guidance interfaces for citizens and businesses. The vision is to deploy AI where it creates value: freeing up time, improving quality, and strengthening the public sector's capacity to serve citizens.
The governance model is a three-way agreement among the government, municipalities, and regions, explicitly grounded in existing positive experiences in the public sector. The vision is that initiatives will eventually reach all relevant public authorities, but rollout begins in specific areas, with implementation communities that municipalities can apply to join. Applications to participate in the AI-assisted documentation project in the employment area opened on 7 May 2026, with an application deadline of 8 June 2026.
Denmark’s National Centre for AI in Society, CAISA, has published two research briefs on whether public-sector AI tools are working for everyone, and whether the transparency obligations being written into EU law are sufficient.
The first brief reviews the research literature on public sector chatbot implementation and finds a familiar pattern: the tools work better for some than for others. Citizens who are younger, well-educated, and digitally confident report positive experiences and higher levels of trust. For citizens with disabilities or complex life situations, the same tools can introduce new friction and deepen existing inequalities in access to public services. The brief also flags something less discussed: chatbots do not simply remove work, they redistribute it. Drifted tasks and new administrative burdens on specialist staff are documented side effects of implementations that look efficient on paper. The Digital Taskforce for AI's ambition is to deploy AI where it creates value, and the CAISA brief asks whose value it creates, how it is measured, and what the fallback is when the tool does not serve a citizen well.
The second brief, published on the same day as the AI Act Omnibus agreement, addresses one of its transparency provisions: the requirement that AI-generated content be marked and disclosed to users. Produced by ten leading European scholars, it concludes that labelling AI-generated content is necessary but almost certainly not sufficient. Research consistently shows that people struggle to distinguish AI-generated from human-generated content, and that transparency labels alone are unlikely to restore trust or meaningfully empower people to act on that information. The debate on whether to label is, the brief argues, settled. The urgent question is how to label effectively, and what governance infrastructure needs to surround it: organisational accountability, enforcement mechanisms, and ongoing research as the technology evolves. For public sector organisations deploying generative AI in citizen-facing services, the EU's Code of Practice on AI content labelling is due in its final form in June 2026, ahead of the binding watermarking obligation, which takes effect on 2 December 2026. The question of how to signal AI involvement to citizens is no longer optional, and the research suggests the answer requires more than a label.
🇪🇺 EU: Largest-ever Digital Markets Act fine expected for Google
The European Commission is preparing to fine Google a triple-digit-million-euro sum for breaching the Digital Markets Act (DMA). The decision, expected before the summer break, would be the largest penalty imposed under the DMA to date. The investigation, launched in March 2025, concerns Google favouring its own services in search results. The Commission has indicated it is primarily seeking compliance rather than punishment, but has signalled it will not wait indefinitely. Google has criticised the DMA's impact on its search product, describing changes made under the regulation as "the biggest downgrade in the product's history."
🇳🇱 Netherlands: Microsoft shares civil servants' names with the US government
Microsoft has reportedly shared the names of Dutch civil servants working at two regulatory agencies, the Authority for Consumers and Markets and the Dutch Data Protection Authority, with the US House of Representatives. The officials were working on implementing the Digital Services Act. Emails, meeting minutes, and invitations were shared without redaction of names. Under the US Cloud Act, American technology companies are required to comply with government data requests regardless of where the data is stored. The Dutch cabinet described the disclosure as extremely worrying, and the State Secretary for Digital Economy and Sovereignty has raised the matter with the US ambassador. The incident is the clearest illustration yet of the gap between European digital sovereignty ambitions and the legal obligations of US cloud providers.
🇩🇪 Germany: Open source developers brought into standards bodies
Germany's Sovereign Tech Agency has launched a pilot programme that funds 10 open-source maintainers to participate in technical standards development at ISO, W3C and IETF. The initiative addresses a structural gap: three-quarters of open source maintainers rely on technical standards, but few participate in writing them. The Agency, which has invested €37.3 million in foundational open-source projects since 2022, frames the programme as part of Germany's broader push to reduce dependence on US technology providers.
🇬🇧 UK: AI adoption stuck between pilot and integration
Around 65 per cent of UK public sector staff are experimenting with AI, but only around 30 per cent have fully integrated it into their workflows, according to a May 2026 analysis by FSP. 41 per cent of public sector employees feel unprepared or unsupported in using AI tools, and 37 per cent of leaders say they would use AI more with clearer training and safeguards in place. The findings suggest the primary barrier to AI adoption in the UK government is not access to tools but workforce readiness and change management.
🇺🇸 US: Erin Brockovich launches crowdsourced data centre impact map
Activist Erin Brockovich has launched a crowdsourced map overlaying major operational and planned hyperscale AI data centres across the US with community-submitted concerns. The map has received more than 2,700 reports, with water usage, energy consumption and health as the top concerns. The initiative comes as major tech companies are collectively expected to spend at least $700 billion on AI infrastructure in 2026, while the Trump administration has taken a largely hands-off approach to regulation. As Europe debates how to measure and regulate the environmental footprint of data centres, this offers an early illustration of what unregulated expansion looks like in practice in the US.
🇻🇦The Vatican: Pope Leo XIV releases encyclical on AI and human dignity
On 25 May, Pope Leo XIV released Magnifica Humanitas, a 42,000-word encyclical on safeguarding the human person in the time of artificial intelligence. The document draws an explicit parallel to Leo XIII's Rerum Novarum of 1891, which shaped labour law and social policy in response to the Industrial Revolution. It addresses human dignity, labour displacement, autonomous weapons, surveillance, and the risk that AI reshapes how we value human life. Addressed to Catholics and non-Catholics alike, it is the most significant institutional voice to enter the AI governance debate this year.
For questions, comments, or suggestions regarding this article, please contact Emilia Lindén Guíñez.
Enjoyed this edition of Decoding? Subscribe here to receive future insights on digital public services directly in your inbox.
.png){kind=logo}
