The Insider’s Warning: When Safety Takes a Backseat
Verity Harding knows the landscape of artificial intelligence better than most. As a former executive at DeepMind, one of the world’s most prestigious AI research organizations, she stood in the trenches of rapid innovation. But today, her message is stark and urgent. Harding warns that the current trajectory of the AI industry, driven by a frantic arms race and fueled by nationalistic government policies, is heading toward a potential disaster.
In a candid discussion with WIRED, Harding articulated a growing concern among safety researchers: the worst-case scenarios we’ve been modeling are not just theoretical—they are taking shape in real-time. The pressure to deploy powerful models before competitors, combined with a geopolitical mindset that treats AI development as a zero-sum game, is creating an environment where safety protocols are being eroded. For Harding, this is not merely a business risk; it is an existential threat that demands immediate attention.
The Toxicity of the AI Arms Race
At the heart of Harding’s critique is the “AI arms race.” For years, major tech companies have competed to build the most capable models, often prioritizing speed and market dominance over rigorous safety testing. This dynamic creates a race to the bottom. If one company slows down to address safety concerns, a competitor might release a less safe but more advanced model, capturing the market share. This leaves safety-focused teams with little leverage, as they are constantly pressured to keep up with a pace that compromises thorough evaluation.
Lessons from DeepMind
Harding’s perspective is informed by her time at DeepMind. She has spoken about the internal tensions between researchers who want to ensure AI systems are reliable and safe, and business units pushing for rapid commercialization. These aren’t abstract debates; they are daily realities in leading labs. When the goal is to outpace rivals, the nuanced work of alignment, interpretability, and risk assessment can be sidelined as an obstacle to progress. Harding’s departure from DeepMind highlights the difficulty of maintaining a safety-first culture in an environment obsessed with speed.
Nationalism and the Worst-Case Scenario
A critical point Harding raises is the role of government policy, particularly in the United States. She identifies a nationalistic attitude toward AI that is accelerating the dangers. Instead of fostering international cooperation on safety standards or establishing global guardrails, governments are increasingly treating AI as a tool of national security and economic warfare.
This mindset manifests in export controls, massive subsidies for compute infrastructure, and policies designed to give one nation a technological edge over others. Harding argues that this approach is evidence that a worst-case scenario is unfolding. The fear of falling behind geopolitical rivals, particularly China, has driven US policymakers to support rapid AI expansion. While economic competitiveness is important, Harding points out that this urgency often comes at the expense of thorough safety evaluations. When the government incentivizes speed and dominance, it inadvertently signals that safety is secondary to national interest. The focus shifts from “Is this safe?” to “Do we have it before they do?”
The Anthology Alternative
Harding is not just a critic; she is also building a solution. She co-founded Anthology, a company dedicated to developing advanced AI research with a primary focus on safety and scientific rigor. Anthology represents a different philosophy: one that rejects the reckless pace of the arms race in favor of methodical, responsible development. The goal is to prove that you can pursue cutting-edge AI without sacrificing the safeguards that protect society.
Anthology’s approach emphasizes “interpretability”—the ability to understand how AI models make decisions. Harding argues that we cannot trust systems we do not understand. By focusing on the underlying mechanisms of AI, rather than just their outputs, Anthology aims to build models that are inherently safer and more predictable. This scientific rigor stands in contrast to the opaque, black-box models that dominate the current market, offering a path toward AI that is both powerful and controllable.
What Needs to Change?
For Harding and fellow safety advocates, the path forward requires a fundamental shift. The industry needs to decouple AI development from the pressures of geopolitical competition. This means pushing for international agreements that prioritize safety over supremacy. It also means holding companies accountable for the risks their models pose, rather than rewarding them solely for capability benchmarks.
Furthermore, the public and policymakers need to recognize that the “move fast and break things” era of technology is over. With AI, breaking things could mean breaking society. The stakes are too high for trial and error. We need a culture of caution, transparency, and collaboration that transcends national borders and corporate rivalries. Safety cannot be an afterthought; it must be the foundation upon which all future AI systems are built.
Conclusion
Verity Harding’s warning serves as a crucial wake-up call. The AI arms race is not just a business story; it is a matter of global security and human welfare. As governments embrace nationalistic approaches and companies continue to sprint toward more powerful models, the risk of a catastrophic outcome grows. The window to course-correct is narrowing. As models become more autonomous and powerful, the margin for error shrinks. Harding’s call to action is clear: we must prioritize the long-term safety of humanity over short-term gains. This requires a unified effort from researchers, companies, and governments to build a framework where AI development serves the public good, rather than fueling a dangerous competition that could end in disaster. Ignoring these warnings could cost us far more than we can afford to lose.
