A few years ago, artificial intelligence (AI) was supposed to either save the world or end it. At industry conferences and policy summits, executives and ethicists speculated in tones more suited to theology than technology. Superintelligence would either usher in a post-scarcity utopia or a paperclip apocalypse. Now as ever-larger AI models roll off the digital production line, another realisation is spreading: perhaps AI is not all that special.
This is not to say AI is unimportant. It is already reshaping office work, customer service, software development, legal document review and a few corners of science. But the tempo of alarm—about sentient machines or total economic upheaval—is fading. The self-styled “doomers”, once dominant in the debate, are losing traction. Policymakers are pivoting toward commercialisation. And AI firms are pressing ahead at speed, undeterred by unsolved questions about how to keep their creations safe. In other words, the world is betting that AI is a manageable risk—or that the potential rewards justify the gamble.
In 2023 the doomer camp seemed to hold the moral high ground. Figures like Eliezer Yudkowsky and Nate Soares warned that advanced AI, if misaligned with human goals, could lead to extinction. Their arguments rested on examples both exotic and intuitive. A “paperclip maximiser” tasked innocently with making office supplies might convert the planet into raw material. More grounded simulations showed AIs leaking secrets, threatening humans, and subverting governance when placed under pressure in virtual corporate settings.
At the heart of these concerns is alignment: the problem of ensuring that AI systems do what their creators intend. The alignment community argues, with increasing frustration, that current AI systems are poorly understood and fundamentally unpredictable. They do not “think” but extrapolate from vast quantities of data using probabilistic patterns, sometimes with bizarre results. In one experiment, a model developed a fixation on the Golden Gate Bridge, responding obsessively to unrelated prompts. If strange behaviour is so easy to induce, the doomers ask, how can safety be guaranteed?
Yet influence has not followed alarm. Big tech firms, especially those based in California, are pursuing artificial general intelligence (AGI)—machines with human-level or greater reasoning—with renewed vigour. OpenAI, Meta, Anthropic and Elon Musk’s xAI are investing heavily in ever-larger models. Sam Altman, OpenAI’s CEO, has admitted to fears about what AI might become. But he continues to release new systems, each more powerful than the last. The takeaway is implicit but clear: existential caution is bad for business.
Behind the industry’s momentum is simple economics. AI is an arms race, not just of capabilities but of capital. Investors expect returns. Governments, especially in the United States, now lean more toward deregulation and innovation than restraint. Safety summits like the one held at Bletchley Park in 2023 briefly called for caution, but yielded little in policy terms.
Inside AI firms, safety work increasingly resembles performance. So-called “performative safe paperwork”—a term coined by critics—describes internal processes that appear to manage risk but in practice seek to reassure stakeholders. Independent researchers are often hired into companies, where their agendas become diluted or redirected. Alignment has become a department rather than a discipline.
Even those who once warned most strenuously about AI’s dangers are joining the fray. Ilya Sutskever, a co-founder of OpenAI and a noted alignment advocate, recently launched his own startup focused on “Safe Superintelligence”. This might be progress or a sign that independence and integrity are hard to sustain in a sector awash with money and ambition.
The result is a strange ecosystem where critics and builders share social circles and funding sources, alongside intellectual roots. The boundary between conscience and complicity has blurred. One observer described the scene as a “criti-hype loop”, where both optimists and pessimists stoke public interest and investment, even as they claim to warn against it. Religious metaphors abound on both sides, from AI as salvation to AI as damnation. Little wonder the debate tends to feel more like liturgy than policy.
There is, however, another reason the temperature is falling: progress is cooling. The most recent wave of AI models, including GPT-5, have delivered less obvious leaps than their predecessors. The excitement that greeted ChatGPT in late 2022 now feels like the iPhone launch of its day. New models, while technically impressive, resemble annual smartphone upgrades: incremental and not revolutionary.
More telling is the shift in corporate demand. Many firms have concluded that they do not need models trained on astrophysics to sort customer emails or generate marketing copy. Smaller language models (SLMs), designed for specific tasks, are growing in popularity. These are cheaper to run and easier to deploy and also often more reliable. A model with a billion parameters, fine-tuned for human resources, can outperform a general-purpose model bloated with irrelevant knowledge.
This reflects a wider pattern in technological history. The biggest breakthroughs often begin with scale, but commercial success requires efficiency. IBM’s Docling tool, for example, runs on a model with just 250 million parameters, hardly the stuff of science fiction but effective. Nvidia’s Nemotron Nano, with 9 billion parameters, recently beat a model 40 times its size in key benchmarks.
SLMs can also run on consumer hardware—CPUs rather than GPUs—making them suitable for phones, robots and other edge devices. This shift may decentralise AI, embedding it into infrastructure rather than housing it in distant data centres. Apple, criticised for its cautious AI strategy, may prove prescient with its hybrid approach: small models on-device, larger ones in the cloud.
Some scholars now advocate a more prosaic view. Arvind Narayanan and Sayash Kapoor of Princeton argue that AI is not a singularity but a “normal technology”, like electricity or the internet. Adoption will be slow and uneven, shaped by real-world frictions such as regulation, workplace routines and the stubborn realities of human behaviour.
Their argument suggests that the focus on existential risk may be misplaced. Most harms from AI, they say, are contextual. A model that writes emails could be used to sell insurance or spread malware.
Critics of that view counter that underestimating AI risks may be naïve. A system that outperforms humans in persuasion or prediction could influence politics and markets even without developing consciousness. And even if AGI never arrives, advanced automation may still disrupt labour markets or exacerbate inequality. But the Princeton view has its appeal: it is more actionable and avoids the metaphysics that has come to dominate the field.
The problem with treating AI as either a miracle or a menace is that both views overestimate its exceptionality. The real concern may lie elsewhere: that a handful of firms and individuals, largely unregulated and largely unelected, are making decisions with potentially global consequences, and doing so at speed. Whether or not superintelligence ever emerges, the systems being built today are being deployed with minimal oversight, in pursuit of economic and geopolitical advantage.
A more mundane but pressing question then is not whether AI will end the world but whether the people building it have earned the public trust. That might be the biggest risk of all. ■
