Макс Тегмарк - Life 3.0 - Being Human in the Age of Artificial Intelligence

When Air Inter Flight 148 crashed into the Vosges Mountains near Strasbourg in France on January 20, 1992, killing 87 people, the cause wasn’t lack of machine-human communication, but a confusing user interface. The pilots entered “33” on a keypad because they wanted to descend at an angle of 3.3 degrees, but the autopilot interpreted this as 3,300 feet per minute because it was in a different mode—and the display screen was too small to show the mode and allow the pilots to realize their mistake.

AI for Energy

Information technology has done wonders for power generation and distribution, with sophisticated algorithms balancing production and consumption across the world’s electrical grids, and sophisticated control systems keeping power plants operating safely and efficiently. Future AI progress is likely to make the “smart grid” even smarter, to optimally adapt to changing supply and demand even down to the level of individual rooftop solar panels and home-battery systems. But on Thursday, August 14, 2003, it was lights-out for about 55 million people in the United States and Canada, many of whom remained powerless for days. Here, too, the primary cause was determined to be failed machine-human communications: a software bug prevented the alarm system in an Ohio control room from alerting operators to the need to redistribute power before a minor problem (overloaded transmission lines hitting unpruned foliage) cascaded out of control.25

The partial nuclear meltdown in a reactor on Three Mile Island in Pennsylvania on March 28, 1979, led to about a billion dollars in cleanup cost and a major backlash against nuclear power. The final accident report identified multiple contributing factors, including confusion caused by a poor user interface.26 In particular, the warning light that the operators thought indicated whether a safety-critical valve was open or closed merely indicated whether a signal had been sent to close the valve—so the operators didn’t realize that the valve had gotten stuck open.

These energy and transportation accidents teach us that as we put AI in charge of ever more physical systems, we need to put serious research efforts into not only making the machines work well on their own, but also into making machines collaborate effectively with their human controllers. As AI gets smarter, this will involve not merely building good user interfaces for information sharing, but also figuring out how to optimally allocate tasks within human-computer teams—for example, identifying situations where control should be transferred, and for applying human judgment efficiently to the highest-value decisions rather than distracting human controllers with a flood of unimportant information.

AI for Healthcare

AI has huge potential for improving healthcare. Digitization of medical records has already enabled doctors and patients to make faster and better decisions, and to get instant help from experts around the world in diagnosing digital images. Indeed, the best experts for performing such diagnosis may soon be AI systems, given the rapid progress in computer vision and deep learning. For example, a 2015 Dutch study showed that computer diagnosis of prostate cancer using magnetic resonance imaging (MRI) was as good as that of human radiologists,27 and a 2016 Stanford study showed that AI could diagnose lung cancer using microscope images even better than human pathologists.28 If machine learning can help reveal relationships between genes, diseases and treatment responses, it could revolutionize personalized medicine, make farm animals healthier and enable more resilient crops. Moreover, robots have the potential to become more accurate and reliable surgeons than humans, even without using advanced AI. A wide variety of robotic surgeries have been successfully performed in recent years, often allowing precision, miniaturization and smaller incisions that lead to decreased blood loss, less pain and shorter healing time.

Alas, there have been painful lessons about the importance of robust software also in the healthcare industry. For example, the Canadian-built Therac-25 radiation therapy machine was designed to treat cancer patients in two different modes: either with a low-power beam of electrons or with a high-power beam of megavolt X-rays that was kept on target by a special shield. Unfortunately, unverified buggy software occasionally caused technicians to deliver the megavolt beam when they thought they were administering the low-power beam, and without the shield, which ended up claiming the lives of several patients.29 Many more patients died from radiation overdoses at the National Oncologic Institute in Panama, where radiotherapy equipment using radioactive cobalt-60 was programmed to excessive exposure times in 2000 and 2001 because of a confusing user interface that hadn’t been properly validated.30 According to a recent report,31 robotic surgery accidents were linked to 144 deaths and 1,391 injuries in the United States between 2000 and 2013, with common problems including not only hardware issues such as electrical arcing and burnt or broken pieces of instruments falling into the patient, but also software problems such as uncontrolled movements and spontaneous powering-off.

The good news is that the rest of almost two million robotic surgeries covered by the report went smoothly, and robots appear to be making surgery more rather than less safe. According to a U.S. government study, bad hospital care contributes to over 100,000 deaths per year in the United States alone,32 so the moral imperative for developing better AI for medicine is arguably even stronger than that for self-driving cars.

AI for Communication

The communication industry is arguably the one where computers have had the greatest impact of all so far. After the introduction of computerized telephone switchboards in the fifties, the internet in the sixties, and the World Wide Web in 1989, billions of people now go online to communicate, shop, read news, watch movies or play games, accustomed to having the world’s information just a click away—and often for free. The emerging internet of things promises improved efficiency, accuracy, convenience and economic benefit from bringing online everything from lamps, thermostats and freezers to biochip transponders on farm animals.

These spectacular successes in connecting the world have brought computer scientists a fourth challenge: they need to improve not only verification, validation and control, but also security against malicious software (“malware”) and hacks. Whereas the aforementioned problems all resulted from unintentional mistakes, security is directed at deliberate malfeasance . The first malware to draw significant media attention was the so-called Morris worm, unleashed on November 2, 1988, which exploited bugs in the UNIX operating system. It was allegedly a misguided attempt to count how many computers were online, and although it infected and crashed about 10% of the 60,000 computers that made up the internet back then, this didn’t stop its creator, Robert Morris, from eventually getting a tenured professorship in computer science at MIT.

Other malware exploits vulnerabilities not in software but in people. On May 5, 2000, as if to celebrate my birthday, people got emails with the subject line “ILOVEYOU” from acquaintances and colleagues, and those Microsoft Windows users who clicked on the attachment “LOVE-LETTER-FOR-YOU.txt.vbs” unwittingly launched a script that damaged their computer and re-sent the email to everyone in their address book. Created by two young programmers in the Philippines, this worm infected about 10% of the internet, just as the Morris worm had done, but because the internet was a lot bigger by then, it became one of the greatest infections of all time, afflicting over 50 million computers and causing over $5 billion in damages. As you’re probably painfully aware, the internet remains infested with countless kinds of infectious malware, which security experts classify into worms, Trojans, viruses and other intimidating-sounding categories, and the damage they cause ranges from displaying harmless prank messages to deleting your files, stealing your personal information, spying on you and hijacking your computer to send out spam.