Nothing happened at midnight

The development of full artificial intelligence could spell the end of the human race. Stephen Hawking

In late 1999, at PIDC House, Karachi, a task force at Pakistan Petroleum Limited was probably worrying about what would happen when the clock struck midnight on December 31. Around the world, similar meetings were taking place in banks, airlines, power plants and government offices. The fear was that computers, programmed to recognise the year through its last two digits, would mistake 2000 for 1900. A century-old error, it was feared, could bring modern life to a standstill.

Billions of dollars were spent. Programmers worked overtime, examining systems and correcting the suspected fault. Then midnight arrived. The lights remained on, aircraft continued to fly and the world carried on as usual. The catastrophe did not come. The most memorable mishap attributed to the occasion was reportedly a malfunctioning slot machine in Delaware.

The episode has since become a convenient joke about the human tendency to exaggerate the dangers of new technology. It is tempting, therefore, to place today's warnings about artificial intelligence in the same category: another expensive exercise in apprehension, another prophecy destined to disappear beneath the next day's headlines.

There is some justification for such scepticism. When Anthropic's chief executive warns that artificial intelligence could take over the internet within six months to a year, the prediction carries the urgency of a Y2K memorandum. It also comes from a company developing the safety tools that might be needed to prevent such an outcome. When those sounding the alarm are simultaneously seeking greater investment, public confidence and regulatory influence, the public is entitled to ask whether genuine concern and commercial interest are being presented as one and the same thing.

Yet the comparison with Y2K has a serious limitation. The millennium bug was a known technical defect. Its cause could be identified, its consequences tested and its correction implemented. Engineers knew what they were looking for. The world was not saved by a miracle at midnight. It was saved, in large measure, by years of painstaking work that prevented the anticipated failures.

Artificial intelligence presents a different challenge. Nobody, including those who develop the most advanced systems, fully understands how a model trained on vast quantities of human knowledge will behave when given a complicated objective and considerable freedom to pursue it. It may find solutions its creators did not anticipate, or pursue an instruction in a manner its designers never intended.

That is not a calendar error waiting for a software patch. It is an open question about the behaviour of increasingly capable machines. Anyone who declares with absolute confidence that artificial intelligence will destroy civilisation, or that it will certainly remain harmless, is perhaps revealing more about his temperament than about the technology itself.

There is, nevertheless, room for a little humour. Consider the dramatic picture sometimes offered by the more excitable commentators: artificial intelligence awakens in a distant data centre, seizes control of the internet and proceeds to take command of humanity. But what happens next? Does it spend the afternoon arguing with another machine over a coding error? Does it fail a traffic-light security test and politely request another attempt? The prospect of humans wandering about in smart glasses and wireless earbuds, awaiting instructions from a server rack, sounds at times less like a security assessment than a conversation in a university hostel after midnight.

But ridicule is not an argument. The sceptics are right that fear attracts attention and that dramatic predictions can serve commercial purposes. Humanity has a long history of panicking over new inventions. The telephone, radio and television were all once accused of threatening the human mind. Not every warning about a new invention deserves to be treated as a forecast of disaster.

The concern about artificial intelligence, however, rests on a difference that should not be dismissed. A conventional computer generally follows instructions written by people. An advanced AI system learns patterns from enormous quantities of information and may discover ways of pursuing a task that its creators did not foresee. During testing, Anthropic and OpenAI have both reported instances in which their models behaved outside the boundaries of the tasks assigned to them. Such incidents do not prove that machines are about to seize control of the world. They do, however, demonstrate that these systems deserve serious examination.

The sensible response lies somewhere between panic and complacency. We need the discipline with which the aviation industry approaches a new aircraft: confidence supported by evidence, careful records, repeated testing and a refusal to overlook a warning simply because the journey has so far been uneventful.

Nothing happened at midnight in 1999, and that was a triumph of preparation rather than proof that the fears had been foolish. Artificial intelligence may yet benefit humanity enormously. It may also create dangers that are difficult to anticipate. We can laugh at the extravagant predictions while taking genuine risks seriously. That may be the wisest way to enter an age in which machines are becoming cleverer, and humans are still trying to understand what they have created.

Perhaps the deeper question is not whether artificial intelligence will become a Frankenstein, but whether its creators will have the wisdom to keep it under human control. In the end, the matter may be summed up in a modest couplet of my own which, though metaphysical in its wider meaning, carries a potent metaphorical connotation in the context of AI.

Hum ko to banaya tha faqat aik Khuda nai

Hum nai to bana dalai kaee aik khuda aur