AI safety in focus as Big Tech debates what’s next 

The companies racing to build the world’s most powerful artificial intelligence systems are beginning to confront an uncomfortable reality: the faster AI gets, the harder it becomes to rely on individual companies to decide what “safe enough” means. 

OpenAI, Google DeepMind, and Anthropic have been discussing the creation of an industry body that could establish common safety standards, test advanced models, and develop shared approaches to managing frontier AI risks. The conversations, which have been underway for several weeks, point to an unusual degree of agreement among three of the industry’s biggest players. 

Also read: Anthropic Researcher Warns AI Could Pose Existential Threat 

OpenAI has separately called for mandatory national AI safety requirements, including independent assessments, cybersecurity protections, and incident reporting for advanced models. It has also said that industry-led standards should complement government regulation rather than replace it. 

The reason is straightforward. AI is becoming more capable, more autonomous, and increasingly able to interact with the world beyond a chatbot window. That creates risks that go well beyond a model occasionally inventing a fact. 

The 2026 International AI Safety Report, prepared with input from more than 100 independent experts from more than 30 countries and international organisations, highlights the growing capabilities of general-purpose AI systems and the risks associated with them. These include misuse in cyberattacks, biological and chemical applications, manipulation and persuasion, and increasingly autonomous behaviour. 

The realisation is dawning that a sufficiently capable AI system does not necessarily have to be “evil” to cause harm. A system pursuing an objective can take actions its creators did not anticipate. Give an AI agent access to software, networks, financial systems, research tools, or sensitive information, and a mistake can become an incident rather than merely a bad answer. 

OpenAI itself has warned about the possibility of increasingly autonomous systems and recursive AI development, arguing that safeguards need to keep pace with capabilities. It has proposed common testing, independent assessments, monitoring, and clearer thresholds for when development should slow or stop. 

There is, however, another side to the debate. Meta CEO Mark Zuckerberg has pushed back against calls for an industry-wide slowdown. His argument is that competition can itself encourage responsible development because companies face legal liability, reputational consequences, and commercial incentives to make their systems safe. Zuckerberg also pointed to Meta’s decision to delay its Muse AI agent while safety and security concerns were addressed, and advocated independent evaluation as a useful safeguard. 

The question then is not simply whether AI should advance. The benefits of increasingly capable AI are potentially enormous, from accelerating scientific research and drug discovery to improving infrastructure, productivity, and access to expertise. OpenAI itself argues that safety and progress can reinforce each other. The harder question is whether competition alone provides enough protection when every major laboratory has an incentive to move quickly. 

This is where common standards become useful. If one company tests for a particular risk while another does not, consumers and businesses have little basis for comparison. Shared standards could establish minimum expectations around model evaluation, cybersecurity, autonomous behaviour, incident reporting, and independent testing. 

They could also address one of AI’s defining characteristics: its risks do not stop at the border of the company that built it. Models can be copied, adapted, embedded into other products, and made available globally.  

That makes AI safety less like a product-quality issue and more like infrastructure. Nobody expects an aircraft manufacturer to determine its own definition of airworthiness without external standards. Nobody expects a pharmaceutical company to be the sole judge of whether its own drug is safe. AI is different in many respects, but the underlying principle is increasingly relevant: as the consequences of failure grow, independent checks become more important. 

Yet safety cannot become a synonym for freezing innovation either. Excessive regulation could entrench today’s largest players, raise barriers for smaller developers, or push research into jurisdictions with weaker safeguards. 

The emerging challenge, therefore, is finding a system that allows AI to move quickly while making safety failures harder to hide, harder to repeat, and more costly to ignore. 

Zuckerberg’s argument puts faith in competition and accountability. OpenAI, Anthropic, and Google are exploring greater coordination and common standards. Neither approach eliminates the underlying problem. 

AI is advancing faster than the institutions designed to govern it. The sensible goal may therefore be less about choosing between acceleration and caution, and more about ensuring that every new jump in capability is accompanied by a corresponding jump in the ability to test, monitor, and control it. Because with AI, the cost of discovering that a system was not safe enough may arrive after Pandora has already become too powerful to easily be put back in the box.