United Nations Office Geneva located in the Palais des Nations building in Switzerland. Shutterstock

United Nations Office Geneva located in the Palais des Nations building in Switzerland. 

The Three Temptations Facing the UN's First Global AI Dialogue in Geneva

Broken Chair or Chaise Cassee is a wooden sculpture by the Swiss artist Daniel Berset and Louis Geneve outside the United Nations in Geneva, Switzerland. Shutterstock

 is a wooden sculpture by the Swiss artist Daniel Berset and Louis Geneve outside the United Nations in Geneva, Switzerland. 

In Geneva, the World Can Anchor AI Governance in Free Expression

The Capitol is seen in Washington, Monday, June 29, 2026. (AP Photo/J. Scott Applewhite)

Bipartisan Smorgasbord of Children’s Online Safety Legislation Passes the House

Collage-style illustration of two hands linking pinky fingers in a pinky promise against a cream paper-textured background. The hands are in black and white, with one wearing a red sleeve and the other a red sleeve with red nail polish. Decorative geometric shapes surround them: yellow circles, dark red rings, a grid pattern, parallel lines, and a cluster of navy dots.

Trust Issues Could Make or Break Agentic Commerce

The image shows a silhouette of a person, likely a woman, sitting on the floor in a dimly lit room. She is sitting near a large window with sunlight streaming in, creating a strong contrast between the light outside and the darker interior. The woman appears to be holding a phone or tablet, looking at the screen while leaning her arms on her knees. The overall mood of the image is reflective, perhaps suggesting a moment of contemplation or solitude. The setting seems to be indoors, with part of the room's furniture visible on the right-hand side.

Human Rights Experts Should Engage in Age Assurance Standards

The United Nations Office Geneva in the Palais des Nations. Shutterstock

The United Nations Office Geneva in the Palais des Nations. 

The World Doesn't Need Another UN AI Declaration. It Needs Architecture.

European officials release the “Tech Sovereignty Package” on June 3, 2026. Source

European officials release the “Tech Sovereignty Package” on June 3, 2026. 

Does Europe Really Have a Plan for Tech Sovereignty?

A neural network comes out of the top of an ivory tower, above a crowd of people's heads (shown in green to symbolise grass roots). Some of them are reaching up to try and take some control and pull the net down to them. Watercolour illustration.

How AI Keeps Europe Hooked on US Cloud

People hold signs in opposition of xAI's use of gas turbines at their facility during a meeting of the Memphis and Shelby County Air Pollution Control Board, Monday, Dec. 15, 2025, in Memphis, Tenn. (AP Photo/George Walker IV)

Inside SELC's Clean Air Case Against xAI in Memphis

The undying sun hangs in the sky, as people gather around signal towers, working through their digital devices.

Digital Nomads Across Time by Yutong Liu & Digit. 

Imagining Broadband Policy of, by, and for the People

An abstract background meant to depict pixel errors and glitches.

Nine Things Platforms Could Do Now to Help Blunt Political Violence in the United States

US Commerce Secretary Howard Lutnick listens to President Donald Trump’s speech during a news conference at the G7 summit, Wednesday, June 17, 2026, in Evian-les-Bains, France. (Jeanne Accorsini/SIPA via AP Images)

Commerce Eased Its Block on Anthropic's Mythos, But Major Questions Remain

(L-R) Congressional candidate Claire Valdez, Congressional candidate Brad Lander, Mayor Zohran Mamdani, and Congressional candidate Darializa Avila Chevalier raise their hands during the 'Our Team, Our Year' Get Out The Vote (GOTV) rally for local candidates ahead of this week's Primary Elections at the Kings Theatre in the Brooklyn borough of New York City, NY, June 18, 2026. (Photo by Anthony Behar/SipaUSA via AP Images)

Populist Candidates Need More Tech Policy—Reforming Tenant Screening Is A Good Start

The New UK Prime Minister Inherits a Social Media Ban. The Real Goal Is Safety by Design.

Tear gas surrounds federal law enforcement officers as they leave a scene after a shooting on Wednesday, Jan. 14, 2026, in Minneapolis. (AP Photo/John Locher, File)

Trump’s Crackdown on Dissent is Targeting Signal—and Zines

Split-screen image depicting contrast. On the left, a Rubik’s cube displays blue, red, and yellow sides, with a LinkedIn 'cr' AI-generation icon nearby. The background is divided between discarded electronic waste on a street and data centre cables, some connecting to the cube. On the right, the entire scene is repeated with inverted colors.

This is not a(i) game! by Laura Sofia Martinez Agudelo / 

When AI Agents Fail, People Ask the Wrong Question About Why

Harvard University graduates, top, celebrate during commencement exercises on the school's campus, Thursday, May 28, 2026, in Cambridge, Mass. (AP Photo/Steven Senne)

For Students, the Process of 'Becoming' is the Challenge No Chatbot Can Solve

Illustration of a person standing with their back facing the viewer, surrounded by large stylized eyes watching them from above. The person is in a spotlight formed by overlapping blue and gray geometric shapes, conveying a feeling of being observed or surveilled.

Who Really Controls Your Digital Likeness in the Age of AI Wearables? Not You.

Paper collage. A photograph of the Atomium in Brussels, cut into diagonal strips and reassembled over a grid of numbers.

Seeing More — Seeing Less by Anna Riepe & FARI / 

Google and Apple’s Anti-DMA Lobbying Strategy Goes All-in on Security and Privacy

A child holds up a smartphone, obscuring their face.

They Said I'd Feel Different About Free Speech as a Parent. They Were Wrong.

On July 6–7, governments, technology companies, researchers, civil society organizations, and international institutions will gather in Geneva for the 

first United Nations Global Dialogue on Artificial Intelligence

 of AI governance, safety, innovation, and international cooperation. But an even more important question has received far less attention: what kind of geopolitical moment is this Dialogue taking place in, and how should that reality shape the conversation?

That question matters more than the agenda itself.

The Global Dialogue arrives at a paradoxical moment. Never has international cooperation on AI been more urgent; yet rarely have the geopolitical conditions for cooperation been less favorable. AI has become the defining technology of our age at precisely the moment when the international system is moving away from collaboration and towards strategic competition, industrial policy and transactional diplomacy. This tension, and not AI itself, should dominate the discussions in Geneva.

History offers a useful lesson. Geneva has often served as the place where dialogue began before trust existed. When Ronald Reagan and Mikhail Gorbachev met there for the first time in 1985, the Cold War was entering one of its most dangerous phases. Their 

 produced no grand agreement and resolved few immediate disputes. Its significance lay elsewhere. It demonstrated that dialogue need not wait for geopolitical rivalry to subside. On the contrary, sustained dialogue became necessary because rivalry had become too dangerous to manage in silence.

AI presents a similar challenge. Unlike previous waves of technological innovation, AI is no longer discussed primarily in terms of productivity or economic growth; it has become strategic infrastructure. Governments increasingly view frontier models, advanced semiconductors, compute capacity, cloud infrastructure and data ecosystems as instruments of national power. AI policy is now inseparable from national security, industrial competitiveness and geopolitical influence.

This shift is understandable. No responsible government can ignore the implications AI will have for economic resilience, military capability, cybersecurity or scientific leadership. The problem is not that AI has become strategic. The problem is that the strategy of individual nations increasingly appears to be foreclosing the possibility of global governance.

This matters because the pace of AI development continues to accelerate while the institutions capable of governing its risks struggle to keep up. The imperative for cooperation has therefore become stronger precisely as the incentives for cooperation have weakened. Governments are investing in sovereign AI capabilities while technology partnerships are increasingly shaped by geopolitical alignment. Export controls, investment restrictions and competition over critical supply chains have become normal instruments of statecraft.

None of these developments are temporary. They reflect a structural shift in international politics. The challenge for Geneva is therefore not to reverse this trend because it cannot. Rather, it is to demonstrate that meaningful cooperation remains possible within it. That requires resisting three temptations.

The first is to avoid talking about security.

Security has become politically uncomfortable in international AI discussions because it immediately exposes competing national interests. Yet pretending security does not belong in AI governance would render the Dialogue largely irrelevant. Questions surrounding military applications, cyber resilience, frontier model security, critical infrastructure and supply chains are central to how governments understand AI today. Ignoring them will not make them disappear; it will simply ensure that the most consequential conversations take place elsewhere. A stark reminder of this reality is the recent US 

 regarding Anthropic, where an export control directive forced the company to pull its powerful Mythos and Fable models offline worldwide over government cybersecurity fears.

The second temptation is the opposite: allowing security to become the 

This would be an equally serious mistake. Once every governance question becomes a national security question, cooperation inevitably narrows, transparency becomes more difficult and technical collaboration becomes more constrained. Governance risks becoming little more than negotiation between competing strategic blocs. That would undermine one of the Dialogue's most important ambitions: ensuring that different governance approaches remain interoperable even where they are not identical.

Interoperability is often described as a technical objective, but it is rapidly becoming a geopolitical one. Countries will not converge around a single model of AI regulation, nor should they. Different legal systems and political traditions will inevitably produce different approaches. The challenge is ensuring that these systems remain sufficiently compatible to support scientific collaboration, trade, innovation and cross-border trust. Fragmentation may appear strategically advantageous in the short term. In the long term, it risks creating competing technological spheres with diminishing capacity to cooperate on shared risks.

The same logic applies to human rights. Too often, human rights are treated as the normative chapter of AI governance rather than one of its strategic foundations. Yet societies will ultimately judge AI not only by what it can do, but by whether it remains accountable to people. Transparency, due process, non-discrimination and meaningful human oversight are not obstacles to innovation; they are essential to the legitimacy that durable governance requires.

The third temptation, and perhaps the most consequential, is to treat the Global South primarily as the object of geopolitical competition.

Much of today's AI diplomacy is framed through the lens of 

, particularly between the United States and China. As a result, developing countries are too often viewed as constituencies to be persuaded or partners to be recruited into competing technological ecosystems.

This fundamentally misunderstands the politics of AI. Countries across Africa, Latin America, Southeast Asia and other developing regions are not merely choosing between competing governance models. Many are asking different questions altogether. How can AI support economic development? Where will compute infrastructure come from? How can local languages be represented in foundation models? How can countries build the skills, institutions and regulatory capacity necessary to participate meaningfully in the AI economy?

These are not secondary concerns. They are central to whether AI governance will ultimately be viewed as legitimate and inclusive.

 centered on infrastructure, investment and capacity-building across much of the developing world. Whether these initiatives ultimately strengthen global cooperation or reinforce geopolitical fragmentation remains an open question. What is clear, however, is that influence today is built as much through practical engagement as through diplomatic rhetoric.

Other actors should take note. The Global South should arrive in Geneva not as diplomatic terrain over which influence is exercised, but as an equal participant whose priorities help shape the conversation itself. Listening may prove more consequential than persuading.

No one should expect the first Global Dialogue to produce consensus on AI governance. That was never its purpose. Its success should instead be measured by something both more modest and more significant: whether governments remain willing to engage one another on a technology that is increasingly viewed through the lens of strategic competition. Geneva has demonstrated before that dialogue does not require trust to begin. It requires recognition that the costs of refusing to talk have become greater than the risks of engagement. That may be the most important lesson for the AI debate.

In a few days, the world's AI governance community will gather in Geneva. They should arrive prepared to discuss algorithms, models and regulation; but they should leave having answered a more fundamental question. Can the international system still cooperate on transformative technologies before geopolitical competition hardens into permanent fragmentation?

The answer will not determine the future of AI. It may, however, determine whether AI becomes another force accelerating geopolitical division or one that reminds us why international cooperation remains indispensable.

Global Dialogue on Artificial Intelligence Governance

 in Geneva July 6-7, and the stakes are high: what rights will anchor the future of how the world governs AI? Generative AI systems mediate, filter, and generate the information we encounter every day. They are, at their core, a technology of expression and access to information. How we govern them will have downstream effects on what people can say, seek, and know. Geneva is where we can get that right.

The past several years produced exceptional progress on examining the interactions between generative AI and rights. The 

 in February 2025 produced a declaration signed by 58 sovereign states that anchored AI governance in human rights and sustainability. Notably, the 

 to join — a fracture in rights-focused alignment by AI powerhouses that contributes to the urgency in Geneva.

Despite that setback, this rights-centered commitment was reinforced months later at the regional organizational level. In October 2025, the UN Special Rapporteur on freedom of expression, alongside counterparts from the Organization for Security and Co-operation in Europe (OSCE), the Organization of American States (OAS), and the African Commission on Human and Peoples’ Rights (ACHPR), 

 affirming that freedom of expression must be embedded throughout the AI lifecycle, from design and training to deployment.

These expert-level commitments reflected obligations towards freedom of expression and access to information UN Member States had already adopted by consensus in 2024 through the Global Digital Compact. The Compact committed governments to protecting freedom of expression and access to information when mitigating potential harms and to refrain from information restrictions inconsistent with international law, a meaningful commitment in a universal instrument. It was the Global Digital Compact that established the upcoming Global Dialogue on AI Governance. Even the earliest intergovernmental standard on AI had already embedded this logic, and the 

 conditioned information integrity measures on respect for freedom of expression.

Together, these instruments establish a baseline that governments have already agreed to. Geneva can build on and strengthen this baseline, ensuring that free expression and access to information are robustly protected in the AI era.

The New Delhi Declaration: an opportunity not taken

 was a missed opportunity for human rights and, in particular, freedom of expression. The international community fell short of building AI governance on a human rights foundation. The 

, endorsed by more than 90 countries and international organizations, called for “industry-led voluntary measures” and described AI’s purpose as advancing “social good,” without accountability mechanisms or rights standards attached to that phrase.

 found “innovation,” “growth,” “efficiency,” and “productivity” dominating the program, while “human rights,” “accountability,” and “surveillance” were largely absent from high-level plenary titles.

The Dialogue's human rights cluster and its independent scientific panel provide the structural anchors that the last summit’s voluntary framework deliberately avoided. This gives Geneva the opportunity to correct what was left behind in New Delhi by reaffirming the integral role of human rights in AI governance.

Why freedom of expression Is central to AI

Generative AI systems now reach over a billion people globally, and information-seeking 

 of generative AI in 2025 across six countries (Argentina, Denmark, France, Japan, the UK, and the US). Whether or not you directly use a chatbot, you likely receive an AI-generated summary every time you search online through services like Google AI Overviews or Bing Copilot Search. That reach makes AI governance inseparable from the governance of expression itself.

 on which to build AI governance in the new era. Article 19 of the International Covenant on Civil and Political Rights, the main instrument protecting freedom of expression globally, 

 to “seek, receive and impart information and ideas […] regardless of frontiers” through any media. This forward-looking wording provides a solid basis to protect AI-generated expression and information that the international community can rely on.

Article 19 should be top of mind to all those participating in the Global Dialogue this summer. Governments have often invoked information accuracy and public order to restrict sociopolitical expression online, endangering political speech, government criticism, journalism, and advocacy. While information integrity is a genuine concern, there is a danger in invoking it without the necessary counterbalance of Article 19's strict requirements in AI governance. This tripartite test 

 any restriction must be provided for in law, pursue a legitimate aim, and be necessary to protect a legitimate aim.

Article 19 is also a helpful tool because it enables policymakers to tackle real harms, like non-consensual intimate imagery and child sexual abuse material, while protecting the tenets of democracy.

Why rights protections matter in practice

Over the last three years, it has become clear that the importance of protecting freedom of expression and access to information extends beyond theory. The examples of free speech abuses under vague safety arguments have been too numerous.

 describing the prime minister’s policies in terms some experts characterized as fascist, the government accused Google of violating IT Rules and issued a formal advisory requiring pre-clearance for AI model launches. The message to AI developers was unmistakable: outputs that displease the government create legal exposure. India’s 

three-hour removal rule for AI-generated content

, introduced in early 2026, compounded that pressure. Platforms must now remove flagged unlawful content within three hours of a government notification. The rule draws no distinction between political satire, journalism, and genuinely harmful material, and it reaches political opponents, journalists, ordinary users, and civil society organizations alike. India’s 

 saw direct content blocking orders more than double to over 24,000 in 2025. The AI removal rule extends that infrastructure into a new medium. However, the use of public order as grounds for restricting AI outputs is not unique to India.

 to Grok content in July 2025 after the xAI chatbot generated outputs deemed insulting to President Recep Tayyip Erdoğan. The stated rationale was to protect public order, precisely the kind of rationale that Article 19’s proportionality requirements exist to constrain.

These pressures are not confined to governments with established records of censorship. In the US, 

 in AI have complicated a principled international stance on freedom of expression. The European Union’s 

 contain systemic risk provisions broad enough to function as speech restrictions, with national regulators already 

 for producing “biased” political outputs, which is a pattern worth keeping an eye on.

Most consequentially for Geneva, China's approach is not merely domestic. In July of last year, China released its 

, invoking “public good” and “safety” as governing principles. Unsurprisingly, China’s domestic AI regime requires adherence to 

 and restricts outputs that may challenge state authority or social order. This framework is being actively promoted in the same multilateral forums whose outcomes will determine whether freedom of expression is protected in global AI governance.

UNESCO Recommendation on the Ethics of AI

 of 2021 provides a template on what rights-based AI governance can look like, providing inspiration for what the results of the Global Dialogue may look like. The UNESCO Recommendation requires any limitation on rights to follow the tripartite test (legality, legitimacy, and proportionality), bars AI systems from being used for social scoring or mass surveillance, and requires member states to ensure that AI actors respect rights in the AI lifecycle. This represents a model worth returning to. Building on its foundation, the Global Dialogue can go further: holding freedom of expression and access to information as an operative standard and applying the Article 19 tripartite test for AI content governance.

 human rights cluster and the Independent International Scientific Panel can operationalize commitments that UN Member States have already made. As we argued in our 

submission to the Global Dialogue’s call for input

The Co-Chairs' Summary and any working group mandates adopted at Geneva should name freedom of expression and access to information explicitly — closely following the Global Digital Compact's commitments — with civil society and independent experts participating substantively in the process.

The Scientific Panel should assess the latest capabilities of AI-enabled surveillance and content authenticity technology in order to fully assess the potential harms to freedom of expression, alongside privacy.

The Dialogue should affirm that Article 19’s requirements of legality, legitimacy, and necessity are consistently held throughout the conversation around AI governance, particularly regarding information integrity measures and any speech restrictions.

AI is the information technology of the future, and it deserves governance that gives careful consideration to rights protection. Geneva is where rights-respecting best practices can be made global, breaching national and regional divides.

On Monday, the United States House of Representatives 

 the Kids Internet and Digital Safety (KIDS) Act, a sprawling collection of measures aimed at mitigating online harms to minors, in a 267-117 vote. The bill, H.R. 7757, will now head to the Senate, where it reportedly faces 

Last week, House Energy and Commerce Committee Chair Brett Guthrie (R-KY) and Ranking Member Frank Pallone (D-NJ) 

 that they had agreed on new text for the legislation, which is anchored around a version of Kids Online Safety Act (KOSA) and the Children and Teens’ Online Privacy Protection Act (COPPA 2.0), two bills that have been batted around by both houses of Congress in various forms for more than four years. A version of COPPA 2.0 

 an earlier version of the KIDS Act in March, shortly after the bill was introduced. However, Guthrie and Pallone appear to have decided to find a text that they could both endorse, whether on principle or, especially in light of the Republicans’ paper-thin majority, to shore up the legislation’s support in the event of lobbying from Big Tech.

Earlier this year, the KIDS Act was the subject of much 

 (for the version of KOSA that was incorporated into the KIDS Act). There are three main critiques of the earlier KIDS Act, two of which have been addressed in the new version:

: The former version preempted state laws covering similar ground; the new version does not preempt any state law that is more protective. Still, Axios 

 that critics are concerned the bill “would make it more difficult to sue social media companies for design features.”

: The earlier version only expected online services to treat users as minors if they “have actual knowledge” of their age or “acted in willful disregard” in assuming that they are not minors, a standard that was seen as too weak by critics, compared to the Senate KOSA’s “knowledge fairly implied on the basis of objective circumstances.” The version that passed the House employs a “know or should have known” standard, which is likely closer to the Senate’s formulation, for KOSA and some of the other measures. The “willful disregard” standard remains in the SPY Kids Act and Data Broker Disclosures measure.

: Other versions of KOSA, including the Senate’s, include a duty of care for platforms “in the creation and implementation of any design feature to prevent and mitigate” a series of specific harms. In both the March and the June texts for the KIDS Act, this is replaced by a requirement to “establish, implement, maintain, and enforce reasonable policies, practices, and procedures that address” a list of harms, which do not include several mental health-related harms that are present in the Senate’s text.

The other important development is that the KIDS Act now incorporates a version of COPPA 2.0 and a measure that requires data brokers who hold minors’ data to register with the Federal Trade Commission (FTC). Meanwhile, POLITICO 

 that the White House is working with Senator Marsha Blackburn (R-TN) on a new package for the Senate, set to include the NO FAKES Act and the App Store Accountability Act as well as KOSA.

Although KOSA and COPPA 2.0 are undoubtedly the two centerpieces of the package, 12 other pieces of legislation are also part of the Act, many of which were 

 independently. Each of these measures is summarized below. Of particular interest, however, are a number of caveats that apply to the interpretation of all measures in the KIDS Act and respond to the concerns of many critics of tech legislation.

None of the following may result from the Act:

prevention of platform actions to mitigate spam or criminality, or to ensure security

preclusion of the use of encryption or decrypt user communications, or,

“the affirmative collection of any personal information with respect to age that is not already collected in the normal course of business.”

It remains to be seen how compatible in practice these stipulations are with the measures prescribed in the main body of the bill, or if they can appease many of the passionate voices in the child safety, online civil liberties, or tech industry communities.

A single regime is also established for all the measures in the KIDS Act. The FTC is to enforce the act under its “unfair or deceptive acts or practices” processes. State Attorneys General may also bring civil actions, though the FTC can intervene and file appeal petitions in those cases. State suits may not be brought while an FTC-led case is pending for the same defendant.

Websites with more than one-third material deemed “harmful to children,” i.e. pornography, are required to use age verification to block minors from access. This is a version of a statute 

already adopted in many red-leaning states

 to restrict minors’ access to pornography, and 

. This appears to be the only measure in the bill that requires age verification; KOSA and COPPA 2.0 specifically exclude a requirement for age gating or age verification being implied. The prescription for age verification here is vague (“a system or process to determine whether it is more likely than not that a user of a covered platform is a minor”), but the legislation appears to give FTC leeway to create specific standards.

The KIDS Act includes a version of KOSA with a collection of measures to protect children from harms on social media platforms, including:

Covered platforms must take steps to address four specific harms to minors (under 17): severe threats of physical violence; sexual exploitation and abuse; the distribution, sale and use of gambling and drugs (including. alcohol, tobacco, and cannabis); and financial scams. Minors should still be able to actively seek out content and access protective resources.

User tools must be provided and defaulted to the most restrictive setting for minors. These include: contact and privacy restrictions; limits on platform time; and the ability to opt out of personalized recommendations and or from specific categories of content being recommended.

Parental controls must also be provided, and defaulted to the most restrictive setting for under-13s. Minors are to be notified when parental controls are applied.

Direct messaging warnings and protections must be provided for teens.

Platforms must respond to reports of harms to minors within 10 days, or sooner for imminent threats.

Advertising illegal or age-restricted products to minors is prohibited.

A third-party audit is required to assess compliance.

Prohibits covered platforms from conducting “market or product-focused research” on minors (under 17), unless it is “used solely to improve the privacy, security, transparency, or safety of the covered platform.”

Imposes safety measures on games that facilitate communication between users. Parental action required to allow communications for under-17s. Minors and parents must also have tools to: prevent the minor being recommended as a contact, restrict purchases, and impose time limits. Minors are to be notified when parental controls are applied.

A collection of specific requirements for providers of AI chatbots when engaging with minors, including: AI chatbots may not claim to be licensed professionals; they must disclose that they are not real people both at the initiation of a session or if asked; they must must provide suicide and crisis hotline details; and they must advise the user to take a break after three hours. The providers must also have policies and procedures in place to address the first three harms cited in KOSA.

Directs a study on minors’ use of social media, platforms’ use of personal data, and mental health impacts of social media on minors.

Directs a report on the ability of minors to access fentanyl through social media; relevant social media features to the access of fentanyl; and possible mitigations.

Assessing Safety Tools for Parents and Minors Act

Directs a review of industry efforts to promote minors’ online safety and their effectiveness.

Also included in the section following this language is a provision directing a 4-year longitudinal study on the risks and benefits of chatbots with respect to the mental health of minors.

Replaces a 2008 statute directing an online safety education program (Protecting Children in the 21st Century Act) with a new effort to identify best practices for online safety education, conduct an outreach campaign, and facilitate the sharing of and access to educational resources and information about online harms to minors.

Directs the FTC to develop and make available educational resources regarding safe and responsible use of chatbots by minors.

Establishes a body to coordinate with agencies and stakeholders to study and report on risks, benefits and evidence-based best practices regarding the impacts on minors of a broad range of online services and apps.

A series of amendments to the original 1998 version of COPPA, providing updates for the contemporary internet and increased stringency. The measures include:

Raising the age of a “child” for the purposes of the law to 14 from 13 and adding a teen category (14-18). Teens have some of the same protections as “children,” but the notice and consent that is in the domain of younger users’ parents is given over to teens themselves.

The “actual knowledge” standard is updated, in at least some cases, to the stricter “knows or should have known” standard.

The category of “personal information” protected under the law is updated to include additional types, such as media with image or voice, identifiers like IP addresses and cookies, biometrics, DNA, and geolocation data.

Prohibits the collection, use, and sharing of children or teens’ personal information for individual-specific advertising 

 unless it is necessary for the service provided (or for compliance with federal law). Prohibits the retention of such information for longer than needed.

Parents and teens have a new right of deletion for their child’s / their personal information.

Separate rules are established for services provided through educational institutions, where the notice and consent are generally given over to that institution.

State laws are no longer preempted from imposing stricter standards.

Directs a study on risks to the privacy and mental health of teens who use financial technology products.

Requires that data brokers who process minors’ personal data (as defined by COPPA 2.0) register with FTC and provide information on their customer credentialing practices and security breaches. A searchable directory of registered data brokers is to be published.

On the bill’s passage in the House, Politico 

 that the prospects of it advancing in the other chamber are “dim.” The August recess looms, as does a break ahead of the midterm elections, leaving Senators little time to consider the bill amid the chamber’s many competing priorities.

is published in partnership with the AI Now Institute.

In our previous iteration of this series, we talked about European AI start-ups that operate at the application layer — and why betting on certain versions of applied AI may further entrench Europe’s dependency on a handful of United States firms.

The obvious counterexample to start-ups like the Swedish vibe-coding company Lovable is Germany’s DeepL. Launched in 2017, the company develops its own, proprietary models at a 

 rather than matching the training-data volume Google built up over years of web crawling. DeepL is more than a European AI success story — it 

has outperformed Google Translate in accuracy assessments

 — and it made GDPR compliance, along with processing customer data on its own servers in Germany and Iceland, a core part of its public narrative, and a main selling point for privacy and sovereignty-conscious users.

That was true until April 2026, when the Cologne-based company informed its paying customers that it was entering a partnership with Amazon Web Services (AWS) and would no longer process data exclusively on its own servers. This announcement was publicly 

 as a deliberate decision and, at least implicitly, as a betrayal of the company’s commitment to Europe.

From a strategic perspective, it might seem puzzling: Why would DeepL, a company that has made sovereignty its key selling point, end up partnering with AWS, one of the (perceived) three horsemen of Big Tech?

In this piece, we argue why European AI Sovereignty is intricately connected to its existing dependence on three US cloud providers: Amazon, Microsoft, and Google.

As a result, each of Europe's attempts to build its way out risks deepening the very dependence it aims to cure. By boosting demand for AI, especially at the frontier, it also boosts demand for proprietary models bundled with the dominant cloud providers. By building gigafactories, it trains models that then run on hyperscaler infrastructure. And by building its own compute for inference, it risks subsidizing capacity for the leading US AI firms, the only customers large enough to fill it.

Hyperscalers own the distribution channels

Contrary to public opinion, DeepL’s decision can be understood through the brutal, structural reality of a highly concentrated global AI stack: DeepL’s enterprise customers (think multinationals) want the same service everywhere. AWS operates globally and can provide low-latency, real-time performance that no European cloud providers can match. Especially as the company is 

 to real-time voice translation, the feasibility of serving AI models globally from a data center in Southern 

 has become unfeasible. For an ambitious company with an IPO lined up for this year, loading the balance sheet with debt to facilitate its own data center buildout in a historically tight market would likely have faced resistance. Lastly, given the business need to compete with Google Translate's ubiquitous presence, access to AWS's end user market and distribution channels likely drove the decision: Once DeepL is listed on AWS Marketplace, enterprises can acquire, deploy and manage DeepL directly within their existing AWS infrastructure through a trusted procurement platform.

All of these factors point to the pervasive monopolies of the global computational infrastructure that pull European companies into their orbits, rather than the companies’ commitment to Europe.

Europe's AI industrial strategy since 2025 has been built around expanding the continent’s ability to train frontier models: centralized, frontier-scale training capacity designed to train, post-train, and evaluate the most complex models. This requires highly specialized compute infrastructure, equipped with (often Nvidia’s) latest AI chips. This framing is politically convenient, as it enables the leveraging and pivoting of existing public high-performance computing clusters (EuroHPC) towards the AI race. Recent AI Factories and Gigafactories are descendants of these initiatives, now facing 

 due to emerging funding gaps, elusive demand and difficulties in governance.

But even if these questions were solved, a more serious problem looms – widespread AI adoption doesn't automatically generate demand for one-off gigantic training. It generates demand for serving inference: responding to individual user requests one at a time, as they arrive, with low latency and high availability, like a user typing into a chatbot and expecting an immediate answer. Inference also requires high-capacity (though not the latest) GPUs, but is a fundamentally different kind of infrastructure than training. It is distributed, elastic, close to users, rather than centralized and connected. This leads to what we have 

— the tendency to fixate on training runs that are conducted on public computational clusters while ignoring the private compute ecosystem in which those models will eventually live and produce value.

Why does mainstream demand for AI products benefit hyperscalers

Now, the obvious question for European policymakers is how to ensure that models run on European cloud infrastructure. There are three reasons why even a huge demand for AI inference doesn’t mean that this demand can (currently) benefit European providers.

Proprietary models on hyperscaler infrastructure

What is worth reiterating for a policy audience is that AI models come in two types, and the difference decides who is allowed to run them. With an open-weight model, the developer publishes the model itself for download, so anyone (i.e., a startup, a university, a European cloud provider) can run a copy on their own infrastructure. Alibaba’s Qwen, Meta's Llama and France's Mistral are open-weight. The most capable models remain proprietary: Anthropic's Claude, OpenAI's GPT, Google's Gemini, even as open-weight models like DeepSeek and Qwen have caught up with the frontier tier.

Proprietary model makers distribute their models on the cloud providers they depend on for capital and compute – Amazon AWS, Microsoft Azure, and Google Cloud – largely bypassing European providers. 

Latest

Podcast

Recent

Our content delivered to your inbox

Thank you!

Around the globe