Global experts demand coordinated AI oversight as deepfakes, system failures surge

In New Delhi this week, a room full of technology officials and researchers from across the globe confronted a problem that has no easy answer: artificial intelligence is moving faster than anyone can regulate it, and the damage is already piling up.

The occasion was the AI Impact Summit, where experts gathered to discuss a growing crisis of AI-related incidents—deepfakes that look indistinguishable from reality, misinformation spreading at machine speed, system failures that cascade through critical infrastructure, cybersecurity breaches that exploit AI vulnerabilities. The conversation kept returning to a single, uncomfortable fact: the tools that create these problems are cheap, accessible, and getting better every month.

Marko Grobelnik, who serves as digital champion for Slovenia at the European Commission, laid out the scale plainly. Generative AI has advanced so rapidly in the past six months that creating high-quality deepfakes now costs almost nothing. The technology itself is neutral—but its potential for abuse, especially around elections, is not. Grobelnik pointed to data from the OECD's taxonomy of AI incidents and hazards, which sorts problems into 14 categories. The patterns are telling: before elections, incident reports spike. The technology knows when it matters most.

But the real problem, Grobelnik argued, runs deeper than the incidents themselves. When something goes wrong—when a deepfake spreads, when a system fails, when a breach occurs—nobody actually understands why. Current monitoring and reporting systems lack what he called causality analysis. They count the damage but not the cause. Without knowing why an accident happened, it's nearly impossible to prevent the next one. What's needed, he said, is a complete framework: detection, analysis, reporting, and a feedback loop that actually learns from failure.

Elham Tabassi, a senior fellow at the Brookings Institution and former chief AI advisor at the U.S. National Institute of Standards and Technology, identified where the system is broken. Governments and companies spend enormous energy checking AI systems before they're released into the world. But most of the real damage happens after deployment, in the wild, where systems encounter situations no one predicted. "The tech is moving much faster for policy to keep up with it," she said. "The majority of incidents we have to worry about cannot be reliably predicted before deployment." The evaluation methods themselves are inadequate. No one quite knows how to test these systems the right way.

Tabassi called for standardized reporting frameworks—common definitions of what counts as an incident, what counts as an accident, what data matters for decision-making. If standards are designed with flexibility, she suggested, they could work across different regulatory regimes. The goal isn't to slow down innovation. It's to create a shared language so that when something breaks, everyone learns from it.

Brazil, facing national elections this year, is taking the threat seriously. Hugo Valadares, from the country's Department of Science, Technology and Digital Innovation, described a national AI plan built around five pillars and 54 specific actions, with more being added as the technology evolves. The focus areas include human resources, data sovereignty, and supercomputing capacity. Cybersecurity has recently emerged as an additional priority. Brazil has recently endured what Valadares called an unprecedented attack on its systems, and the country is investing heavily in large-scale AI infrastructure to support research and defense. President Luiz Inácio Lula da Silva is, according to Valadares, deeply concerned about fake news and generative AI in the election cycle ahead. "We need to take care of children in this dangerous environment," Valadares said—a reminder that the stakes extend beyond politics to the basic safety of vulnerable populations.

What emerged from the summit was a consensus without a solution: the world needs coordinated oversight, faster response mechanisms, and standardized ways of reporting and learning from AI failures. But building that infrastructure while the technology itself continues to accelerate remains the central unsolved problem. The experts in the room knew what needed to happen. Whether governments and companies can move fast enough to make it happen is still an open question.