- The EU AI Act’s full enforcement powers over GPAI model providers took effect August 2, 2026, with fines up to 3% of global annual turnover, meaning an AI vendor’s compliance posture now directly affects every enterprise that deploys their models in the EU market.
- The Global Partnership on AI (GPAI), a 46-member initiative hosted by the OECD, translates its principles into nine concrete projects under its 2026 Work Plan, covering responsible AI scaling, government data sharing and workforce impact research.
Full enforcement of the EU AI Act‘s provisions for General-Purpose AI model providers arrived on August 2, 2026, and with it, the first genuinely binding global AI governance deadline that carries real financial teeth. Fines for non-compliance reach 3% of global annual turnover or €15 million, whichever is higher. Running alongside this hard-law push, the Global Partnership on AI (GPAI) is pursuing a parallel, soft-law track: nine concrete projects designed to turn high-level principles into practical tools that any of its 46 member governments can actually use.
The EU’s Hard Enforcement Reality
The August 2, 2026 enforcement date marks the point at which the EU AI Act shifts from preparation to accountability for GPAI model providers. The European Commission and its AI Office can now conduct technical evaluations, demand compliance and risk-mitigation measures, and restrict or withdraw models from the EU market. GPAI models are defined as those capable of performing a wide range of tasks across multiple domains, processing large volumes of training data and integrating into downstream systems. Providers of models trained on more than 1025 FLOPS, the threshold for systemic-risk designation, must notify the AI Office and comply with stricter documentation, transparency and copyright requirements.
The Act also requires mandatory machine-detectable content marking by that same August 2026 date. For enterprises, the practical implication is direct: their AI vendor’s compliance posture is now coupled to their own. A model pulled from the EU market disrupts every downstream deployment built on it. The enforcement timeline itself has been contested an earlier 16-month delay in the Act’s rollout left compliance teams uncertain about deadlines, which makes the August 2026 date a fixed point organisations should treat as firm.
GPAI’s Foundations
GPAI was proposed by Canada and France at the 2018 G7 summit and formally launched in June 2020. It is hosted by the Organisation for Economic Co-operation and Development (OECD) in Paris and grounded in the OECD’s 2019 AI Principles, which set a human-centric, safety-focused baseline that most major AI governance frameworks now reference. The partnership has grown to 46 members, including individual countries and the European Union, with working groups covering Responsible AI, Data Governance, the Future of Work, and Innovation and Commercialisation.
Three Centres of Expertise anchor the operational work: Montreal, Paris and Tokyo. These function as nationally funded research and project hubs, bringing together governments, industry, academia and civil society to develop guidance that is evidence-based rather than purely aspirational. The deliberate separation from binding regulation is the point, GPAI’s outputs are designed to complement hard law, not compete with it. Where the EU AI Act sets mandatory thresholds GPAI develops the practical tools that help institutions meet them.
Nine Projects, Real Deliverables
The 2026 Work Plan organises GPAI’s output around nine Associated Projects, each developed and delivered by the Centres of Expertise. “Scaling Responsible AI Solutions” is a mentorship programme that supports teams deploying responsible AI frameworks in practice; its third global cohort launched in 2025 and continues through 2026. The “Government Data Sharing Roadmap” provides frameworks and tools for public institutions to share data responsibly for AI innovation, with workshops for GPAI members planned in 2026.
The VIADUCT project (Virtuous Innovative Approaches and Data Use Collaboration for AI Training) targets the persistent barriers to data sharing for AI training, drawing on technology, economics and law to connect data holders with developers. Two new projects, launched on the back of the May 2025 Tokyo Innovation Workshop, address culturally aware datasets and multilingual AI models, areas that matter most to members in the Global South and non-English-speaking regions.
The “AI@Work Labs Network” ties into the Future of Work workstream. A 2025 report mapped the landscape of AI and employment; the network now expands to additional labs in 2026, with collaborative research designed to feed directly into international policy discussions. Taken together, these projects represent GPAI’s bet that concrete, use-case-specific tools will outlast high-level declarations in actual policy influence.
Summits and Declarations
GPAI’s November 2025 Bratislava Summit, held under the theme “Putting the Partnership into action,” produced a draft Bratislava Declaration covering future orientations for 2026. Ministers and experts discussed compute infrastructure, AI investment mobilisation, lifecycle safety and equitable benefit-sharing across development levels. The summit’s framing was deliberately operational, less about principles, more about what GPAI members commit to doing next.
In February 2026, India hosted the AI Impact Summit in New Delhi, where a GPAI Council meeting took place on the margins. That followed the Paris AI Action Summit of February 2025, co-chaired by French President Emmanuel Macron and Indian Prime Minister Narendra Modi, which drew more than 1,000 participants from over 100 countries across themes including Public Service AI, the Future of Work and Global AI Governance. The New Delhi Declaration focused on balancing AI promotion with regulation, a tension that runs through most national AI strategies and that GPAI’s multi-stakeholder model is designed, at least in theory, to help navigate.
The Expert Community in Practice
GPAI’s Multistakeholder Expert Community numbers more than 500 AI specialists across government, industry, academia and civil society. The Tokyo Innovation Workshop in May 2025 brought together over 170 participants to design solution-oriented projects across four themes: AI in the Global South, interoperability of governance frameworks, multilingual and multicultural AI, and open-source AI. Two of the nine 2026 Work Plan projects trace directly back to that workshop’s outputs.
The use-case-based approach is deliberate. GPAI’s expert groups are tasked with developing guidance that is practically implementable, not just technically correct. The Centre in Tokyo, for example, focuses on issues of AI governance interoperability that are particularly acute in Asia-Pacific, where regulatory approaches diverge significantly. Whether the volume of expert input translates into policy adoption across 46 diverse member governments is a separate question, and one GPAI has not yet fully answered.
Challenges Ahead
A May 2025 report from GPAI’s Future of Work Working Group identified what it called the “Big Unknown”: roughly 281 million workers whose employment futures depend on AI-related decisions that have not yet been made. That scale of socio-economic exposure demands governance that moves faster than current international processes typically allow.
The broader challenge for GPAI is one of positioning. The global AI governance calendar is crowded: the EU AI Act, the UN’s advisory body work, bilateral AI agreements and domestic regulatory regimes all compete for policy bandwidth. GPAI’s value lies in its multi-stakeholder model and its non-binding outputs, guidance that governments can adopt without treaty obligations, but that flexibility is also the limitation. Where the EU can compel compliance, GPAI can only persuade. How much of its expert-driven work actually shapes national policy, rather than informing it abstractly, will determine whether the partnership remains relevant as hard-law frameworks like the AI Act increasingly set the terms of the debate.
