Scientists issue warning after robot’s rapid learning becomes too complex
Scientists have issued a disturbing warning after an AI robot taught itself how to play chess and then learned moves, that have “never been seen before” in the game’s entire 1500 year history, in just four hours.
After just 240 minutes, the Artificial Intelligence software called AlphaZero had devised a whole new strategy that was far too complex for human opponents and beat the world’s top chess grandmaster.
Leading Oxford academic, Professor Michael Wooldridge warned that this is a perfect example of how AI could go “rogue,” that machines might become so complex that the engineers who create them will no longer understand them or be able to predict how they function.
This has raised fears that technology may be becoming too advanced and could be dangerous should humans “lose control” as its “thinking” goes beyond our comprehension.
Daily Mail reports: AlphaZero, an AI computer program, this month proved itself to be the world’s greatest ever chess champion, thrashing a previous title-holder, another AI system called Stockfish 8, in a 100-game marathon.
So far, so nerdy, and possibly something only chess devotees or computer geeks might get excited about.
But what’s so frighteningly clever about AlphaZero is that it taught itself chess in just four hours. It was simply given the rules and — crucially — instructed to learn how to win by playing against itself.
In doing so, it assimilated hundreds of years of chess knowledge and tactics — but then went on to surpass all previous human invention in the game.
In those 240 minutes of practice, the program not only taught itself how to play but developed tactics that are unbeatably innovative — and revealed its startling ability to trounce human intelligence. Some of its winning moves had never been recorded in the 1,500 years that human brains have pitted wits across the chequered board.
Employing your King as an attacking piece? Unprecedented. But AlphaZero wielded it with merciless self-taught logic.
Garry Kasparov, the grandmaster who was famously defeated by IBM’s supercomputer Deep Blue in 1997 when it was pre-programmed with the best moves, said: ‘The ability of a machine to surpass centuries of human knowledge . . . is a world-changing tool.’
Simon Williams, the English grandmaster, claimed this was ‘one for the history books’ and joked: ‘On December 6, 2017, AlphaZero took over the chess world . . . eventually solving the game and finally enslaving the human race as pets.’
The wider implications are indeed chilling, as I will explain.
AlphaZero was born in London, the brainchild of a UK company called DeepMind, which develops computer programs that learn for themselves. It was bought by Google for £400 million in 2014.
The complex piece of programming that created AlphaZero can be more simply described as an algorithm — a set of mathematical instructions or rules that can work out answers to problems.
The other term for it is a ‘deep machine learning’ tool. The more data that an AI such as AlphaZero processes, the more it teaches itself — by reprogramming itself with the new knowledge.
In this way, its problem-solving powers become stronger all the time, multiplying its intelligence at speeds and scales far beyond the abilities of a human brain. As a result, it is unconstrained by the limits of human thinking, as its success in chess proved.
But the real purpose of such artificial intelligence goes far beyond playing board games against other boxes of silicon chips. It is already starting to make life-or-death decisions in the high-tech world of a cancer diagnosis.
It is being trialed at NHS hospitals in London, including University College London Hospital (UCLH) and Moorfields Eye Hospital.
At UCLH, a system is being developed in which an AI developed by DeepMind will analyze scans of patients with cancers of the head and neck, which afflict more than 11,000 people a year in the UK.
Google experts say the AI should be able to teach itself to read these scans ever quicker and more accurately than any human, so radiation can be more precisely targeted at tumors while minimising damage to healthy tissues in the brain and neck. What currently takes doctors and radiologists four hours could be done in less than an hour.
Meanwhile, at Moorfields, a DeepMind AI will analyze the 3,000 or so high-tech eye scans carried out each week. Currently, only a handful of experts can interpret the results, which may cause delays in treatment. It is believed that AI will be able to identify problem scans faster.
On the surface, it looks like a win-win for patients and the NHS. But there are major issues. The first is privacy — the London hospital trials have involved handing over the scans of more than a million NHS patients to Google.
This is causing alarm among privacy campaigners and academics. Dr. Julia Powles, who works on technology law and policy at Cambridge University, says ‘Google is getting a free pass for swift and broad access into the NHS, on the back of unproven promises of efficiency and innovation’.
Dr. Powles adds: ‘We do not know — and have no power to find out — what Google and DeepMind are really doing with NHS patient data.’
Google has tried to address the criticisms of its project by declaring that all data access will be subject to NHS monitoring, but this is an organization that has long had to contend with allegations of prying into people’s data for commercial advantage.
It faces court action in the UK over claims it unlawfully harvested information from 5.4 million UK users by bypassing privacy settings on their iPhones. The group taking action, called Google You Owe Us, alleges Google placed ‘cookies’ (used to collect information from devices to deliver tailored adverts) on users’ phones without their knowledge or permission.
Google has responded: ‘This is not new. We don’t believe it has any merit and we will contest it.’
But the insertion of a super-intelligent AI into NHS decision-making procedures brings an infinitely more worrying concern.
It is an open secret that the NHS effectively rations access to care — through waiting lists, bed numbers and limiting the availability of drugs and treatments — as it will never have enough funds to give everyone the service they need.
The harsh reality is that some deserving people lose out.
The harsher alternative is to be coldly rational by deciding who and who not to treat. It would be most cost-effective to exterminate terminally ill or even chronically ill patients or sickly children. Those funds would be better spent on patients who might be returned to health — and to productive, tax-paying lives.
This is, of course, an approach too repugnant for civilized societies to contemplate. But decision-making AIs such as AlphaZero don’t use compassionate human logic because it gets in the way. (The ‘Zero’ in that program’s name indicates it needs no human input.)
The same sort of computer mind that can conjure up new chess moves might easily decide that the most efficient way to streamline the health service would be to get rid of the vulnerable and needy.
How we keep control of deep learning machines that will soon be employed in every area of our lives is a challenge that may well prove insurmountable. Already top IT experts warn that deep-learning algorithms can run riotously out of control because we don’t know what they’re teaching themselves.
And the programs can develop distinctly worrying ideas. A system developed in America for probation services to predict the risk of parole-seekers reoffending was recently discovered to have quickly become unfairly racially biased.
DeepMind certainly acknowledges the potential for problems. In October it launched a research team to investigate the ethics of AI decision-making. The team has eight full-time staff at the moment, but DeepMind wants to have around 25 in a year’s time.
But, one wonders, are 25 human minds enough to take on the super-intelligent, constantly learning and strategizing powers of a monstrously developed AI?
The genie is out of the bottle. In building a machine that may revolutionize healthcare, we have created a system that can out-think us in a trice. It’s a marvel of human ingenuity. But we must somehow ensure that we stay in charge — or it may be checkmate for humanity.