At the UN’s AI for Good Summit, the Race to Define ‘Responsible AI’ Collides with Reality

The United Nations’ AI for Good Summit in Geneva showed a sharp divide between lofty promises and practical concerns: governments, nonprofits and tech companies gathered this week to pitch artificial intelligence as a tool for solving global problems, even as critics warned that the same systems can deepen inequality, weaken rights and concentrate power.

Held by the UN International Telecommunication Union and now in its 10th year, the summit unfolded across a sprawling convention center near Geneva airport with robot demos, hardware displays, coding sessions and panel discussions on everything from humanitarian aid to digital standards. Beneath the spectacle, the central argument was not whether AI should be used in society, but who gets to decide what “for good” actually means.

What the UN AI for Good Summit was trying to do

The summit’s core mission was straightforward: bring together public institutions, private companies, researchers and civil society groups to discuss how AI can support development, public health, education and disaster response without amplifying harm.

Doreen Bogdan-Martin, secretary-general of the ITU, framed that ambition in keynote remarks that emphasized AI’s promise to help address hunger, disease and climate change. But she also acknowledged that the technology is now under scrutiny, not only because of what it can do, but because of the risks it is creating as it spreads faster than the rules around it.

The event was designed as both a showcase and a policy forum. Attendees moved between live technical demonstrations, workshops, panels, and a networking space outfitted with unusual seating gadgets and attention-grabbing displays. That format reflected the summit’s larger challenge: turning a sweeping global slogan into something measurable, enforceable and useful.

Why the summit felt different from the usual AI conference circuit

The answer is that this was not another industry event focused mainly on product launches, fundraising or enterprise adoption. The Geneva summit took place against a backdrop of escalating geopolitical fights over chips, frontier models and AI regulation, and it kept returning to the question of whether the benefits of AI are being shared fairly across countries and communities.

In Washington, lawmakers have been hearing warnings about superintelligence and national security. In parallel, the White House has used export controls to shape access to advanced chips. Meanwhile, China is weighing how open its own open-weight models should remain. Against that climate, the UN forum highlighted a different concern: that too much of the world may be shut out of AI altogether.

That concern was visible throughout the summit’s discussions. Speakers repeatedly linked access to compute, data, hardware and technical standards with broader questions of development. If a country cannot afford the infrastructure required to run frontier systems, then AI’s benefits may remain concentrated in a narrow group of wealthy markets and major technology firms.

How the debate over compute became a development issue

The answer is that compute is no longer just a technical resource; at the summit, it was described as a form of development infrastructure.

Panels focused on the widening gap between countries that can build and deploy large AI systems and those that cannot. Speakers argued that access to hardware, cloud infrastructure and efficient model architectures will shape who can participate in the AI economy and who will be forced to depend on foreign platforms.

One of the clearest themes was that smaller countries and under-resourced institutions should not be expected to rely entirely on giant, English-dominant models hosted in distant data centers. Instead, participants said, there is a strong case for smaller, locally tuned language models that can run on cheaper hardware and better reflect local languages and needs.

Syed Munir Khasru, chairman of the Institute for Policy, Advocacy, and Governance, argued that if AI for good is to mean compute for all, policymakers must treat computing access as part of development infrastructure rather than as a purely technological question.

That framing captured the summit’s deeper political edge. The issue was not simply whether AI could be made safe. It was whether people outside the US-China-Europe power centers would be allowed to help shape the tools, standards and markets that define AI’s future.

Who is being left out of the AI economy?

The answer, according to many summit speakers, is everyone who lacks money, infrastructure or leverage over the major AI supply chain.

Global debates over AI access usually sound abstract, but the summit kept bringing them back to concrete exclusions: limited access to chips, dependence on proprietary clouds, language bias in model design and public-sector procurement choices that lock institutions into expensive systems they cannot fully audit.

That problem matters because access is increasingly tied to power. A country or agency that cannot inspect its own digital stack may not be able to control how sensitive data is processed, where it is stored or which company can change the tools it depends on. As AI systems get embedded in government, health, education and humanitarian workflows, those dependencies can become hard to unwind.

The summit’s discussions also reflected a growing concern that AI governance often assumes wealthy regulators and well-resourced institutions. That leaves many regions with weak bargaining power when it comes to licensing, procurement and standards-setting.

AI language bias remains part of the problem

One of the recurring criticisms was that many large language models still work best in English, which makes them less useful in the places where access to technology is already uneven.

That matters because AI systems do not become globally useful simply by being available online. They need local linguistic support, practical affordability and deployment models that match the realities of public agencies, schools, clinics and small businesses. Without that, the benefits of AI remain skewed toward rich, English-speaking markets.

What critics said about the tech industry’s role

The answer is that some attendees were openly skeptical of the private sector’s influence over humanitarian and public-interest technology.

On the sidelines of the summit, Giulio Coppi of the advocacy group Access Now criticized the long-running tendency of public and humanitarian organizations to rely on large technology companies without enough transparency or accountability. He argued that the sector needs to move past what he called an age of innocence and stop treating major tech vendors as trusted partners by default.

Coppi warned that decades of expensive, publicly funded technology deals have often produced systems that are difficult to explain, difficult to verify and constantly changing under the hood.

His criticism was part of a broader warning about opacity. In many cases, governments and aid organizations buy complex stacks of software and services that they do not fully understand, then struggle to answer basic questions about data processing, vendor updates or long-term control.

That dynamic is especially fraught in the AI era. When models are updated rapidly and systems are built from layers of third-party services, it becomes harder to know what exactly is running, what it is trained on and how it may affect the people who rely on it.

How human rights got pulled into the technical conversation

The answer is that summit participants argued human rights can no longer be treated as a separate policy issue once they are built into standards, procurement and system design.

Gilles Thonet, deputy secretary-general of the International Electrotechnical Commission, said that engineers have historically assumed questions about human rights belong to lawyers or policymakers. At the summit, that assumption was challenged directly.

His point was echoed by other standards experts who said the most important decisions are often made long before a system reaches the public. They happen in technical specifications, product procurement, interoperability choices and the hidden architecture that governs how systems are built and deployed.

Anja Kaspersen, director of global markets development for frontier and critical technologies at IEEE, said those invisible layers are where much of AI’s real power is concentrated. In her view, the industry needs a “middleware” layer that can translate broad human-rights principles into technical enforcement mechanisms that can actually be checked and audited.

That kind of thinking points to a major shift in AI governance. Instead of relying only on broad principles or voluntary commitments, speakers at the summit argued for operational tools that can be embedded into systems, tested and enforced.

Why impact assessments are not enough on their own

The answer is that impact assessments often become paperwork unless they are tied to real oversight.

Jeremy Ng, counsel for AI and the digital economy at the World Bank, said AI impact assessments need to be more than symbolic exercises. His warning was that such reviews can easily turn into “governance theater” if they are used mainly to show regulators or donors that an organization has checked the right boxes.

That critique resonated across the summit, where many speakers seemed to agree that good intentions are not a substitute for enforceable accountability. The challenge is designing processes that force institutions to answer concrete questions: Who benefits? Who is harmed? Who can appeal? Who can inspect the system?

What happened on the summit floor?

The answer is that the event mixed serious policy debate with a near-festival atmosphere of gadgets, demonstrations and visual spectacle.

Attendees encountered live coding sessions, AI refresher courses, and hands-on displays meant to show how the technology can be applied across sectors. There were also more eccentric touches, including a networking installation that resembled a rotating seating platform and a convention-floor environment where headphones and simultaneous audio streams made the event feel like a hybrid of trade show and immersive exhibit.

Among the most eye-catching displays were Tesla Cybertrucks, a UN rescue helicopter and a variety of humanoid and robot demonstrations that drew crowds across the hall. The pace of the exhibits created a contrast with the slower-moving policy talks upstairs and in conference rooms, where delegates tried to negotiate language about governance, fairness and responsibility.

How did politics shape the mood at the summit?

The answer is that politics were impossible to separate from the technology conversation.

Outside Geneva, the global AI agenda is already being shaped by trade restrictions, export controls and strategic competition. Those realities were felt inside the summit as well, because every discussion about infrastructure, standards or access ultimately ran into questions of sovereignty and control.

That political tension also surfaced in a more visible form during the event. Pro-Palestine protesters disrupted a keynote by Amazon chief technology officer Werner Vogels, alleging that the company’s technology is being used by Israel against Palestinians. Security eventually removed the activists from the venue.

The disruption underscored how easily AI conferences can become stages for disputes far beyond machine learning or software architecture. As more governments and companies deploy AI in sensitive settings, the line between technical policymaking and broader ethical controversy continues to blur.

What does the new UN commission actually do?

The answer is that the 44-member commission is intended to help guide the “AI for Good” agenda, not to solve every AI problem on its own.

The UN said the group will help shepherd the initiative going forward. It is co-chaired by Rwandan President Paul Kagame and Salesforce CEO Marc Benioff, an arrangement that reflects the summit’s effort to blend political authority, private-sector influence and multilateral coordination.

At a high level, the commission is meant to support consensus-building and help define priorities for future action. That matters because the summit repeatedly returned to the same obstacle: no single government, company or NGO can set the rules for AI alone.

Bogdan-Martin said the future of AI cannot be shaped by one stakeholder acting alone and stressed that builders, policymakers and institutions all have a role in making the technology work for society.

Still, the commission’s existence also points to the limits of summit diplomacy. It can assemble a broad coalition and create momentum, but it cannot by itself resolve the hardest questions: enforcement, investment, market concentration, accountability or international power imbalance.

Why the summit’s biggest question is still unanswered

The answer is that the world still has no shared definition of what “AI for good” means in practice.

That phrase can refer to everything from medical diagnostics to climate modeling to disaster response. But at Geneva, speakers kept returning to the fact that noble language is not enough. A system may be marketed as beneficial while still excluding languages, shifting costs onto poorer countries or making institutions dependent on vendors they cannot audit.

Harvard engineering professor Vijay Janapa Reddi captured that frustration in a blunt assessment: there is plenty of excitement around AI, but the reality often falls short of the hype. He argued that “good” is too vague to serve as a design target on its own.

That bluntness mattered because it exposed the central weakness of the summit’s idealism. Engineers need precise specifications. Policymakers need enforceable standards. Communities need ways to challenge harmful systems. Without those things, “AI for good” can become an attractive slogan rather than a workable framework.

From idealism to implementation

The answer is that the summit’s next test will be whether it can turn principles into procurement rules, standards and measurable outcomes.

That would mean moving beyond broad declarations and into specifics: local-language model support, transparent vendor contracts, public-sector auditing tools, accessible compute, and standards that encode human rights rather than merely referencing them.

It would also require more honest conversation about trade-offs. AI systems that are cheap, fast and widely available may still create dependency or amplify bias. Systems that are carefully governed may be slower to deploy. The point, several participants suggested, is not to eliminate those trade-offs but to make them visible and politically accountable.

The bigger picture: AI governance is becoming infrastructure politics

The answer is that AI governance is no longer just about rules for software; it is about who controls the infrastructure underneath it.

That infrastructure includes chips, cloud platforms, model weights, language coverage, public procurement and the standards that determine how systems interoperate. The summit made clear that each of those layers carries power.

For wealthier nations and the largest technology firms, that may look like a competitive advantage. For smaller states, humanitarian groups and underfunded public agencies, it can look like a sovereignty problem. The more essential AI becomes to basic services, the more expensive it is to remain dependent on systems you cannot fully inspect.

The Geneva summit did not resolve that problem, but it did make it harder to ignore. By putting robot demos, policy experts, development officials, activists and corporate leaders in the same room, it exposed the tension between AI’s futuristic branding and the old-fashioned realities of power, money and control.

In the end, the most striking thing about the conference may have been its pacing. The robots on the floor looked faster than the consensus-building upstairs. That mismatch captured the moment: AI is advancing quickly, but the world is still arguing over who should benefit, who should decide, and what counts as good in the first place.