Forget killer robots: AI's most likely end is a great disappointment

Andrew Rogoyski, writing for Live Science, has a contrarian message for the AI era: the technology is far more likely to end in a great disappointment than in a Hollywood-style war against humanity. The immediate danger is not an awakened digital god but a high-profile debacle that freezes investment before AI's most promising applications can mature. That, Rogoyski argues, is the scenario executives and policymakers should be planning for, not killer robots.

The reason the nightmare scenario is unlikely is almost mundane. Today's frontier AI systems are codependent on fragile infrastructure: electricity, cooling, data centers, specialized chips, networks and technicians. Interrupt the power, restrict the GPUs used for training, disable the cooling or disconnect the network, and in most scenarios the supposedly omnipotent machine stops. Rogoyski adds that a true artificial superintelligence would probably understand this, since destroying its own industrial support system would be self-defeating.

The systems that should worry us most, he says, are not the smartest ones but incomplete intelligences acting on narrow objectives with access to critical infrastructure. The classic thought experiment is Nick Bostrom's paper clip problem, where an AI pursuing a single objective converts the Earth and humanity into paperclips. The question is what we give these systems control over. In the near term, two hazards deserve priority: AI-enabled cyberattacks and long-term economic upheaval.

Cyber disruption is a real and present risk because hospitals, banks, energy networks, logistics companies and governments depend on interconnected digital systems. Recent examples of frontier AI systems hacking organizations, such as the OpenAI/Hugging Face incident, have given currency to the fear. Such attacks can cause extraordinary financial damage without being existential. But the likeliest route by which advanced AI could exert control is not force; it is influence, misleading people, manipulating institutions and redirecting human effort. AI may industrialize human weaknesses, though it did not invent them.

The economic danger is more subtle. Rogoyski warns that executives seduced by demos may dismiss workers long before AI has proved capable of replacing them, cashing in the "AI dividend" too soon. Fluent language output can be confused with dependable labor, shedding institutional knowledge to chase theoretical efficiencies. The vision of the "one-person unicorn," a billion-dollar business built by a person or small group using AI as its workforce, is questionable technically, economically and socially. Organizations are not merely bundles of tasks; they contain accountability, relationships, tacit knowledge and trust.

The political turning point could resemble the 2017 WannaCry ransomware attack, which used NSA-designed code and disrupted the UK's National Health Service. An AI-amplified cyber operation would be worse: services fail, lawsuits multiply, companies collapse, investors retreat. Governments would demand a pause or force labs to redirect resources toward control, auditing and safety. After the panic, the public might ask why we let this happen. This is how an AI winter arrives, as it did in the 1970s and early 1990s.

Rogoyski also points to an uncomfortable incentive: AI labs may be amplifying fear and uncertainty because casting a product as potentially world-ending implies unprecedented power. That narrative is marketing that can sustain investment, defend stock valuations and prepare for IPOs even when business models are uncertain. The more mundane outcome is the great disappointment: advanced AI could prove too expensive and insufficiently useful to continue on its present trajectory. Training and operating frontier systems consume enormous computing time, electricity, water, capital, hardware and human effort.

For founders, operators and investors, the strategic stakes are clear. Don't anchor strategy on superintelligence; build for a world where a cyber shock freezes funding and AI's economics get tested. Preserve institutional knowledge while experimenting, because the companies that treat AI as dependable labor too early may find they cashed the dividend before it was earned. And watch the infrastructure: the machines that cannot run without power, chips and cooling are not gods, they are industrial assets with a very human vulnerability.