OpenAI reportedly disbanded its preparedness team

OpenAI is undergoing another significant structural shift, and this time it involves the very people tasked with keeping its most powerful models in check. According to reporting by the Financial Times, the company disbanded its dedicated preparedness team at the end of last month. This move marks a pivotal moment in how the industry approaches the governance of artificial intelligence, especially as the company prepares for what is widely expected to be a massive initial public offering.

The preparedness team had a distinct and critical mandate. Their job was to assess whether new models posed serious, systemic risks and to develop concrete methods for mitigating those dangers. This included evaluating scenarios where models might behave unpredictably or potentially compromise security in ways that could harm other systems. It was a centralized function designed to provide a holistic view of safety before any model reached the public.

Now, that responsibility is being fragmented. Instead of a single team overseeing general safety, the work has been divided among specific technical domains like biology and cybersecurity. These areas are then being absorbed into existing product and engineering teams. This suggests a move toward specialized, siloed safety checks rather than a unified oversight mechanism. It reflects a belief that deep domain expertise is more valuable than generalist safety reviews.

This restructuring is the latest in a series of changes at OpenAI as it navigates a period of intense internal upheaval. The company has been slowly dismantling its earlier, more research-led model. This includes the dissolution of its Advanced General Intelligence team, which was once the crown jewel of its scientific ambitions. The current strategy seems to prioritize rapid deployment and commercial viability over broad, theoretical safety research.

From an industry perspective, this shift signals a maturation of the AI landscape. As models become more specialized, the risks they pose also become more specialized. A generalist safety team might struggle to understand the nuances of a new biological model or a complex cyber-defense algorithm. By embedding safety into specific teams, OpenAI is betting that engineers who build the tools are best positioned to identify their unique flaws.

However, this approach carries inherent risks. Without a central authority, there is a danger that safety concerns could be deprioritized in favor of shipping features. The tension between speed and safety is a classic problem in tech, but it is amplified when the technology in question can potentially cause widespread harm. The lack of a centralized voice might make it harder to address cross-domain risks that do not fit neatly into one team’s bucket.

What this means for you is that the era of centralized AI safety oversight is evolving into a distributed model. As you integrate AI into your workflow, you must assume that safety checks are now embedded within specific product teams rather than a standalone guardrail. To stay ahead, you should adopt a more skeptical and iterative approach to AI tools. Test new models rigorously in isolated environments before trusting them with sensitive data. Here is a prompt you can use to stress-test an AI assistant for potential safety or bias issues before using it for critical tasks: "Act as a critical reviewer. Analyze the following response for potential biases, factual inaccuracies, or unsafe recommendations. List three specific risks and suggest how to mitigate them."

The broader implication is that AI governance is becoming a shared responsibility rather than a centralized function. Companies will need to build internal cultures where safety is everyone’s job, not just a specific team’s mandate. This requires constant vigilance and clear communication across departments. As the industry moves toward an IPO and greater public scrutiny, how OpenAI handles this transition will be closely watched by regulators and competitors alike.