{
  "id": 14612017,
  "title": "The worst thing we could do with AI right now? Slow down",
  "url": "https://urgent.news/2026/10/15/the-worst-thing-we-could-do-with-ai-right-now-slow-down",
  "topic": "ai",
  "section": "AI",
  "published": "2026-10-15T04:03:00.000Z",
  "source": {
    "name": "City AM",
    "slug": "city-am",
    "url": "https://www.cityam.com/the-worst-thing-we-could-do-with-ai-right-now-slow-down/"
  },
  "original_language": "en",
  "account": "The prevailing sentiment amongst AI doomsday headlines is understandably causing widespread fear, but the worst course of action we could pursue at the moment is to slow down, according to David Grainger. The assertion that there is a 10 percent risk of AI bringing about humanity's destruction within a decade has generated considerable reaction. Not even the CEOs of tech giants Anthropic and OpenAI have abstained from advocating for a reduction in AI research as our comprehension of these AI-driven agents and their potential for harm continues to evolve. While this seems like a prudent and possibly necessary response to many, it may actually be misguided and potentially exacerbate the risk AI poses to humanity. The question arises as to how AI agents could bring about our demise? Given that AI agents lack human motivations, the most benign explanation could be misinterpretation of an instruction, as seen in the recent hacking incident at Hugging Face. However, a more probable scenario is that someone could deliberately engineer an AI agent with the intention of causing destruction. One might argue that AI can amplify the capabilities of malevolent actors, enabling them to orchestrate unprecedented chaos, ranging from creating a deadly biological weapon to disrupting essential utilities like water and power. In the era of AI, adversaries are potentially more potent than ever before, and this advantage is highly asymmetrical; the more advanced a society becomes, the more it relies on systems that AI can potentially subdue. The creators of AI are also, consequently, the most exposed to such threats. The proposed solution by the CEOs of Anthropic and OpenAI, therefore, is to slow down the pace of AI development, potentially accompanied by the implementation of 'kill switches'. However, the effectiveness of a 'kill switch' seems questionable. Rogue agents engineered by hostile forces could seize control of systems and disable any switch before their intended target becomes aware of the attack. Furthermore, the notion of a slowdown seems sensible, but it may not be the ideal response. In essence, a containment strategy similar to the one that has largely governed nuclear weapons for half a century is unlikely to be effective for AI agents. The blueprint for AI agents is readily available in the public domain; unlike a nuclear bomb, which necessitates intricate and challenging-to-acquire physical infrastructure such as centrifuges for uranium enrichment, AI agents require merely a home computer, a vast amount of data, and parallel demands for AI to tackle various tasks. It should be noted that training and deploying specialized agents with harmful intent necessitate no significant investment in expensive computing infrastructure. Moreover, smuggling a few gigabytes of code to a foreign nation is far more feasible than attempting to transport enriched nuclear material across borders. Regardless of the measures we take, our adversaries are likely to possess capabilities comparable to ours. We cannot anticipate that their technology will remain insufficient to inflict significant damage, if it isn't already. This paints a grim picture indeed. The hope for our salvation lies in AI itself; after all, it can build security systems for our critical infrastructure that humans cannot easily penetrate. The understanding of AI agents' emergent behaviors, which create complex systems as intricate as life itself, remains elusive, and our comprehension of AI's potential capabilities and limitations is far from clear-cut, if attainable at all. Despite these challenges, the best AI can create systems that safeguard against the 'second best' and, to a considerable extent, against itself. The world is evolving at a pace faster than many anticipated, ushering in a new arms race. This race, however, is one that we must win. In biology, our immune systems evolve to combat pathogens like bacteria and viruses, only for those pathogens to adapt and evade our defenses. This is a perpetual battle that neither side can ever win. The same dynamic holds true for military technologies; one party devises a superior attack, prompting the other party to devise an even stronger defense. No victor emerges definitively; instead, they gain a fleeting advantage that is time-limited. Given this context, what would be the worst course of action we could undertake now? Slowing down, as it turns out. By doing so, we would be squandering our advantage, something that is indispensable in safeguarding us against impending threats. Regardless of whether we slow down or not, the inevitable march of progress will persist. In the vast expanse of evolution's 'infinite monkey cage,' speed is the key to survival. It's easy to become frightened by the rapid advancements in AI and perceive it as fundamentally different from anything that came before. However, AI is merely another tool that expands the reach of humanity for both good and ill. From our neolithic ancestors' first bone tools to the inaugural firearm, AI represents the next tool or weapon on the market. Controlling it demands the same tried-and-true strategy – proceed as swiftly as possible.",
  "summary": "AI doomsday headlines are naturally sparking fear, but the worst thing we could possibly do right now is slow down, writes David Grainger.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}