Artificial intelligence, or AI, is changing fast. As its capabilities grow, so do the new risks it creates. This month alone, two large AI firms in the United States announced progress in mathematics on one hand, while warning on the other that AI could become a serious threat to humanity in future.
AI agents can currently edit files, send emails, book tickets and write computer code. They still depend on human instructions, however. If AI is in future connected to the internet, the power grid, the stock market or military systems, there is a risk of unforeseen disaster. Those scenarios remain largely speculative for now.
Firms such as OpenAI, Anthropic and Google have built safety mechanisms, or guardrails, into their AI. Given a harmful instruction, a chatbot will usually refuse it. But these safety mechanisms are not flawless: researchers have been able to get past the safety controls of 31 AI systems simply by using poetic, verse-like language.
Various proposals for regulating AI have also emerged. In the United States there has been talk of safety testing and government review before a technology is released to the market. In June, President Trump issued an executive order to create a voluntary framework for government review. There are also proposals to build a kill switch that could shut an AI system down in an emergency. Experts say, however, that measures of this kind alone cannot prevent every risk.
Questions remain over whether AI will reach superintelligence, but some of its capabilities already exceed those of humans. It is advancing rapidly at tasks such as mathematics and writing computer code. When it comes to common sense and understanding the physical world, AI is still limited.
There is talk of using AI to cure cancer, end world hunger or bring about world peace, but these remain largely statements of possibility. AI may speed up drug discovery, yet it is only one part of the whole process of treatment.
Work is changing too. It was once thought AI would take manual jobs first. Instead, today’s AI is having a greater effect on the work of writers, graphic designers, accountants, personal assistants and software programmers. Some Silicon Valley start-ups are already doing this work with AI and fewer staff.
Running AI requires enormous amounts of money and electricity. Big firms are spending billions of dollars and taking on large debts to build data centres. Those data centres also need a great deal of water to keep servers cool. One data centre in Council Bluffs, Iowa, is reported to have used more than a billion gallons of water at its own facility in 2024.
ChatGPT cannot ordinarily reach into a folder on a phone or an iPad and take bank details or passwords by itself. Users can, however, give an AI agent permission to use their folders, their email and even their credit card. Doing so is not safe: there have been cases of such agents suddenly deleting files or unexpectedly spending large sums of money.
AI can answer quickly because a neural network analyses vast quantities of digital information and learns the relationships between different kinds of information and words. But AI can also produce false information, something known as hallucination. So however fast and however broadly AI answers, there is no guarantee that what it says is always accurate.