AI Is Supposedly Getting Closer, Yet 11 Fallacies About It Are Everywhere | One More List a Day

If your picture of artificial intelligence comes from sci-fi movies and tech CEOs' speeches, you would do well to find some other sources.

"We've built a computer that works like the human brain"

Tech companies large and small are all talking about AI, and neural network computing is their favorite proof that they stand at the technological frontier. After marketing departments have simplified its principles again and again, it has turned into something that "works like the neural networks of the human brain."

But neural network computing, which already existed in 1980, has only the vaguest connection to the human brain. One big problem is that scientists still have no idea how the brain's network of countless neuronal synapses actually "computes." As Michael I. Jordan, an AI and machine learning expert at UC Berkeley, put it:

"We have no idea how the brain stores information and how it operates. What are the rules? What is the algorithm? So right now we can't say we're using theories of the brain to guide AI systems."

"It passed the Turing test, so it's intelligent"

You probably remember a headline from 2014: "Russian team develops AI robot that passes the Turing test for the first time." The computer, called Eugene Goostman, was described as having "posed as a 13-year-old boy and fooled more than 30% of the judges."

In 1950, Alan Turing, the father of computer science, first proposed his famous experiment on whether machines can think. A human tester, not told what is on the other side, converses in text with a computer. If the computer fools the tester into taking it for a real human, it is deemed to "think."

The Royal Society later made Turing's standard concrete: in a series of five-minute keyboard conversations, a computer that is mistaken for a human more than 30% of the time passes the Turing test.

The annual Loebner Prize competition uses this standard, and it produced bots that "passed the test" long ago. In 2011, for example, the program Cleverbot achieved a 59% pass rate.

Critics of Eugene Goostman argued that the so-called robot was built for one purpose: to pass the five-minute test. Its makers were also rather crafty. They cast it as a 13-year-old whose first language is not English, so that whenever it couldn't answer a question or got something wrong, the judges would put its incoherence down to its age. In other words, Goostman was neither the "first program to pass the test" nor an AI robot.

Incidentally, getting a computer to bluff its way through isn't hard. Here are 10 tricks, if you're curious.

"This robot is already nearly as smart as an XX-year-old child"

"Inside Baidu there is a project called 'Baidu Brain' that uses technology to simulate human thinking. It is now roughly as intelligent as a 2- or 3-year-old. In ten or twenty years, a brain like this may be smarter than the human brain." So said Baidu CEO Robin Li.

Plenty of others have assigned ages to AI programs. Researchers at the University of Illinois at Chicago gave a computer an IQ test and concluded that the machine's "intelligence" was close to that of a 4-year-old human child. Microsoft's marketing copy for Xiaoice describes the voice assistant as "a cute 17-year-old girl."

In 2012, Swedish scientists developed an AI program said to have an IQ of 150, and the press reported that it "beat 96% of humans."

But if these tech companies could drop the marketing nonsense and talk honestly about their AI, "this robot has the intelligence of an intellectually challenged 5-year-old" might be a better description.

For a start, intelligence testing is itself controversial. Intelligence is made up of three abilities: short-term memory, reasoning and language. The three interact; you may be a gifted reasoner with a poor memory. A test of 11 items cannot fully measure anyone's intelligence.

As Jeff Hawkins says, the human brain can learn almost anything. A child of two or three has already mastered at least one language and can pick out their parents at a glance in a very cluttered scene, whatever their posture or expression. AI is nowhere near that, let alone able to pin down something as nebulous as "intelligence."

"Algorithms are taking control of everything"

"With algorithms, we can make better choices than humans can." Every few years this claim comes round again. From Techmeme, founded in 2005 to aggregate news from tech sites by machine, to Toutiao, which talks up personalized recommendation in every interview, all have used algorithms to pull in news automatically from across the internet. But in the past few years, things have changed.

In 2008, Techmeme hired Megan McCarthy, a former tech reporter at Wired, and her job was to step in by hand on the news the computer picked. Techmeme founder Gabe Rivera said that content pushed purely by algorithm has problems, such as uneven quality and poor timing. The ideal news aggregator, he said, mixes "automatic plus human," which handles news more effectively and covers the weaknesses of both people and computers.

Today Techmeme keeps a nine-person editorial team, and Toutiao is hiring more and more editors to select news by hand on top of the algorithm.

Smart assistants have seen a similar reversal. Siri and its kind painted a future in which you say one sentence into a microphone and the AI takes care of everything, whether ordering food, finding movie tickets or checking a game score. But to keep their accuracy up, the assistant-style lifestyle services that have recently appeared in China and abroad have mostly switched to human operators. Even Facebook's newly launched Facebook M uses people to handle requests, while the machine learns how they do it.

For a good while yet, at least, algorithms and humans will keep helping each other.

"This robot has learned to reproduce itself"

Recently a laboratory at the University of Cambridge gave birth to a robot that the media hailed as able to "have babies." Its "mother" is a robotic arm, and its "babies" are small blue 3D-printed cubes, each with its own electric motor inside.

On the far-fetched idea of robots reproducing, John von Neumann, the father of the modern computer, pointed out that any self-reproducing system must have two basic abilities: it must be able to build certain components and assemble them into a next generation like itself, and it must be able to pass a description of itself on to that next generation. By that measure, the so-called "baby-making" robot does not meet the definition of "reproduction".

Even if an AI like "Skynet" that could spread like a virus appeared one day, its reach would be very limited, because a computer that could wipe out humanity needs a physical body. As Jeff Hawkins, an AI expert and author of On Intelligence, has said: "Can you imagine it [AI] building a factory on its own and then having babies nonstop?"

"The Terminator is here, and robots now kill"

In July this year, an assembly-line robot at a contractor's factory for Germany's Volkswagen came under a storm of criticism. While working, the robotic arm suddenly seized a worker and crushed him against a metal plate. He suffered severe chest injuries and died.

Next came headlines like "Robot kills factory worker," along with articles quoting celebrities of the tech world on the "AI threat," such as Stephen Hawking's "AI could lead to the extinction of humanity" and Elon Musk's "AI is more dangerous than nuclear weapons."

But Volkswagen's "Terminator" was really no different from a car accident. Volkswagen confirmed that the accident was caused by "human error." The "robot" that killed can be programmed to carry out movements on its own, but it is a world away from "artificial intelligence."

The first recorded death caused by a factory robot happened at a Ford plant in the US in 1979. Since then, such incidents have averaged fewer than one a year, a figure that is negligible next to car crashes.

"AI has to look like a human"

Hiroshi Ishiguro, a well-known AI expert in Japan, builds robots that look as close to human as possible. He has even gone so far as to make a copy of himself.

We often fantasize about a mechanical companion exactly like us, but don't expect him or her to do your housework, because you'll have been scared out of your wits long before that. According to the "uncanny valley" theory, if a robot could truly reach 100% likeness to humans, people would react warmly. But before that, in the "close but not quite" stage, which is where Ishiguro's robots sit, it feels like looking at a walking corpse.

Doesn't the round, chubby robot vacuum seem warm and cute by comparison?

"Robots will obey Asimov's Three Laws"

In 1942, the science fiction writer Isaac Asimov set out his famous "Three Laws of Robotics": a robot may not injure a human being or, through inaction, allow a human being to come to harm; a robot must obey orders given by human beings, except where such orders conflict with the First Law; and a robot must protect its own existence, as long as doing so does not conflict with the First or Second Law.

Science fiction fans treat these as the ground rules of the future. In fact, letting machines decide who lives and dies is no heresy. More than 30 countries already have "lethal autonomous weapons systems," and in parts of future battlefields, AI may decide by itself whether to shoot a moving object that appears on a screen.

From the military's point of view, letting AI decide life and death has many advantages: it carries out its mission without hesitation, and a killing machine feels no guilt. Handing such decisions entirely to AI also means drone operators no longer have to pull the trigger. Flying drones by remote control in a sealed building in Virginia by day, then sitting down to dinner with your family at night and seeing news of a mistaken strike, is not an easy job.

AI deciding life and death will no doubt stir up plenty of controversy. In May 2013, the moral, ethical and human rights questions it raises came before the UN Human Rights Council for the first time. But the debate is not one-sided, and Asimov's Three Laws are not necessarily the future we will face.

"My AI can do..., because it uses deep learning"

Since roughly the time Google acquired DeepMind, deep learning has become the magic word on the lips of every startup with any connection to AI. It seems that once you have "deep learning," which supposedly "learns like the brain," your phone will understand every word you say in a noisy room and translate it, tone and all, and will recognize the clothes the woman in front of you is wearing and find the same item on Taobao.

In startup pitches, "deep learning" has nearly become what all-powerful "socialism" was half a century ago.

Yann LeCun, a researcher with more than 30 years of AI experience and director of AI research at Facebook, explains deep learning this way: build learning machines and train them to learn everything with the same method.

For a training system, massive amounts of training are still the foundation, and many startups have failed to get this right.

And even with massive training, the old limits remain. When Google trained a machine heavily and used deep learning to get it to recognize cats, the machine could reliably spot a cat only when it faced the camera head-on.

"Comparing AI to the brain gives it an aura of magic. That kind of description is dangerous, and it leads to hype."

LeCun went on:

"Such descriptions create expectations among investment firms, the public, potential customers, startups and investors, making them believe we are at the cutting edge and building systems as powerful as the brain. In reality we are still very far from that goal."

"A few months of hard work, and the robot becomes self-aware"

AI in sci-fi movies always leaves a wildly optimistic impression: building a convincing robot is easy.

In Chappie, for example, the robot hero is rebuilt in a garage-like storeroom. From the moment it is born, it teaches itself rapidly and grows into an adult within days. In the Iron Man films, JARVIS, the British-accented AI butler, comes out of Tony Stark's flashy lab. In Transcendence, the scientist played by Johnny Depp easily uploads his consciousness at home and becomes an immortal virtual being.

Don't assume real scientists are that capable. Professor Stuart Russell of the University of California, author of Artificial Intelligence: A Modern Approach, says that one person building an AI robot single-handedly, shut away from the world, is simply unrealistic. Such work usually takes a team of scientists, and results come only through a long, slow process.

As for uploading consciousness, Russell bluntly calls it "nonsense," because we humans still can't say what "consciousness" even is.

"We'll copy your memories and use AI to let you live forever as a virtual person"

A startup working on data analytics and AI says it can build a virtual you that imitates everything you say and do and how you think. Your memories, it says, will be preserved forever, and you will become immortal.

Don't get carried away with that sweet dream just yet. All this company stores is data such as your social media and emails, and right now there is no way to transplant memories. The reason is simple: no one knows how the brain actually stores memories. In fact, even where in the brain memories are kept is an unsolved mystery, never mind preserving them.

Or let's get (clumsily) a little more philosophical: what counts as a "memory" in the first place? What you remember, or what you think you remember?

Cover image from Hellorf