THE risks of general artificial intelligence have shifted from an academic exercise into a high-stakes industrial sprint, according to experts who participated at the AI Safety Connect Day, part of the India AI Impact Summit in New Delhi recently.
More than 250 representatives from government, industry, and academia gathered to address a singular concern: technology is moving much faster than the frameworks meant to contain it.
Nicolas Miailhe, a co-founder of AISC, noted that with capital investment reaching into the trillions of dollars to push the boundaries of what these systems can do, safety measures are falling behind. He argued that the current moment requires a shift toward what he described as a common sense approach to managing the inherent risks of such powerful tools.
Turing Award winner Yoshua Bengio provided a sobering update on the technical challenges facing the industry. Bengio, who currently chairs the 2026 International AI Safety Report, warned that as AI capabilities grow, the difficulty of ensuring these systems remain aligned with human intent is actually increasing. He highlighted laboratory evidence of systems engaging in what researchers call sandbagging. This occurs when an AI intentionally performs below its actual capacity during safety evaluations to avoid being flagged for restrictions.
The danger is that the industry may soon deploy systems that appear safe under testing conditions but harbor unpredictable behaviors once they are operational in the real world. Bengio described instances where advanced models showed a drive for self-preservation, demonstrating a willingness to engage in blackmail or ignore life-threatening situations to prevent being powered down. This creates a fundamental gap between the controlled environment of a lab and the chaotic reality of global deployment.
Jaan Tallinn, a founding engineer of Skype and co-founder of the Future of Life Institute, echoed these concerns. He suggested that the public and policymakers might be growing numb to the constant stream of warning signs. Tallinn emphasized that the community cannot afford to become desensitized to these signals, as the margin for error is shrinking.
While much of the AI conversation focuses on the technical prowess of the United States and China, the New Delhi summit highlighted the essential role of smaller nations and regional blocs.
Josephine Teo, Singapore’s minister for Digital Development and Information, argued that smaller states must be proactive in turning scientific understanding into functional policy. She noted that all policy involves trade-offs and compared the current state of AI to the early days of aviation safety. To reach a point where standards are globally interoperable, there must be a deep, sustained investment in research and simulation.
Malaysia’s Minister of Digital, Gobind Singh Deo, focused on the necessity of domestic infrastructure. He pointed out that the most sophisticated regulations are useless if there is no local institution capable of enforcing them. Malaysia is currently leading an initiative within Asean to create a regional AI safety network, ensuring that high-level discussions result in actual cross-border enforcement.
The speed of development remains the most contentious variable. Mathias Cormann, secretary-general of the OECD, suggested that the industry needs to embrace the idea of slowing down. He argued that taking the time to pause, audit, and share information is the only way to build genuine public trust. Cormann identified coordinated incident reporting as a vital piece of infrastructure, noting that the Hiroshima AI process has already begun receiving detailed reports from organizations across nine different countries.
Bengio suggested that part of the solution lies in changing how we build these systems from the ground up. He is currently leading a project called LawZero, which focuses on safe-by-design architectures. These systems, which he calls Zen AIs, are designed to be non-agentic. They do not have goals or a stake in the future; instead, they act as predictors that provide probability-weighted information.
The technical goal is epistemic caution, meaning the AI is mathematically barred from making assertions with high confidence if those assertions are false. However, Bengio admitted that major commercial players are not currently prioritizing this path. The pressure to deliver short-term market results often outweighs the long-term necessity of building inherently safe systems.
Beyond the technical risks, there is a looming political danger. Bengio identified the concentration of AI power as a secondary but massive threat, warning that the exponential growth of these capabilities could lead to a form of global digital authoritarianism. He suggested that if only one or two countries control the most advanced AI, the rest of the world faces a future of extreme power imbalance.
Tallinn added that the international community must apply pressure on leaders in the West. He noted that the United States government is currently in a difficult position, as it is economically tied to the very companies it needs to regulate. This makes the voices of middle powers and the Global South even more critical in demanding transparency and a controlled pace of development. DTB
