How smart are the robots getting?

Franz Broseph seemed like any other Diplomacy player to Claes de Graaff. The handle was a joke — Austrian Emperor Franz Joseph I reborn as an online bro — but that was the kind of humor that people who play Diplomacy tend to enjoy. The game is a classic, beloved by the likes of John F. Kennedy and Henry Kissinger, combining military strategy with political intrigue as it re-creates the First World War: Players negotiate with allies, enemies and everyone in between as they plan how their armies will move across 20th-century Europe.

When Franz Broseph joined a 20-player online tournament at the end of August, he wooed other players, lying to them and ultimately betraying them. He finished in first place.

De Graaff, a chemist living in the Netherlands, finished fifth. He had spent nearly 10 years playing Diplomacy, both online and at face-to-face tournaments across the globe. He did not realize until it was revealed several weeks later that he had lost to a machine. Franz Broseph was a bot.

“I was flabbergasted,” de Graaff, 36, said. “It seemed so genuine — so lifelike. It could read my texts and converse with me and make plans that were mutually beneficial — that would allow both of us to get ahead. It also lied to me and betrayed me, like top players frequently do.”

Built by a team of artificial intelligence researchers from the tech giant Meta, the Massachusetts Institute of Technology and other prominent universities, Franz Broseph is among the new wave of online chatbots that are rapidly moving machines into new territory.

When you chat with these bots, it can feel like chatting with another person. It can feel, in other words, like machines have passed a test that was supposed to prove their intelligence.

For more than 70 years, computer scientists have struggled to build technology that could pass the Turing test: the technological inflection point where we humans are no longer sure whether we are chatting with a machine or a person. The test is named for Alan Turing, the famed British mathematician, philosopher and wartime code breaker who proposed the test back in 1950. He believed it could show the world when machines had finally reached true intelligence.

The Turing test is a subjective measure. It depends on whether the people asking the questions feel convinced that they are talking to another person when in fact they are talking to a device.

But whoever is asking the questions, machines will soon leave this test in the rearview mirror.

Bots like Franz Broseph have already passed the test in particular situations, like negotiating Diplomacy moves or calling a restaurant for dinner reservations. ChatGPT, a bot released in November by OpenAI, a San Francisco lab, leaves people feeling as if they were chatting with another person, not a bot. The lab said more than a million people had used it. Because ChatGPT can write just about anything, including term papers, universities are worried it will make a mockery of class work. When some people talk to these bots, they even describe them as sentient or conscious, believing that machines have somehow developed an awareness of the world around them.

Privately, OpenAI has built a system, GPT-4, that is even more powerful than ChatGPT. It may even generate images as well as words.

And yet these bots are not sentient. They are not conscious. They are not intelligent — at least not in the way that humans are intelligent. Even people building the technology acknowledge this point.

“These systems can do a lot of useful things,” said Ilya Sutskever, chief scientist at OpenAI and one of the most important AI researchers of the past decade, referring to the new wave of chatbots. “On the other hand, they are not there yet. People think they can do things they cannot.”

As the latest technologies emerge from research labs, it is now obvious — if it was not obvious before — that scientists must rethink and reshape how they track the progress of artificial intelligence. The Turing test is not up to the task.

Time and time again, AI technologies have surpassed supposedly insurmountable tests, including mastery of chess (1997), “Jeopardy!” (2011), Go (2016) and poker (2019). Now it is surpassing another, and again this does not necessarily mean what we thought it would.

We — the public — need a new framework for understanding what AI can do, what it cannot, what it will do in the future and how it will change our lives, for better or for worse.

The imitation game

In 1950, Alan Turing published a paper called “Computing Machinery and Intelligence.” Fifteen years after his ideas helped spawn the world’s first computers, he proposed a way of determining whether these new machines could think. At the time, the scientific world was struggling to understand what a computer was. Was it a digital brain? Or was it something else? Turing offered a way of answering this question.

He called it the “imitation game.”

It involved two lengthy conversations — one with a machine and another with a human being. Both conversations would be conducted via text chat, so that the person on the other end would not immediately know which one he or she was talking to. If the person could not tell the difference between the two as the conversations progressed, then you could rightly say the machine could think.

“The question and answer method seems to be suitable for introducing almost any one of the fields of human endeavour that we wish to include,” Turing wrote. The test could include everything from poetry to mathematics, he explained, laying out a hypothetical conversation:

Q: Please write me a sonnet on the subject of the Forth Bridge

A: Count me out on this one. I never could write poetry.

Q: Add 34957 to 70764

A: (Pause about 30 seconds and then give as answer) 105621.

Q: Do you play chess?

A: Yes.

Q: I have K at my K1, and no other pieces. You have only K at K6 and R at R1. It is your move. What do you play?

A: (After a pause of 15 seconds) R-R8 mate.

When Turing proposed the test, computers could not chat. Scientists communicated with these room-size machines by feeding mathematical and textual instructions into vacuum tubes via typewriters, magnetic tape and punched cards. But as the years passed and researchers created a new field they called artificial intelligence — a concerted effort to build machines that could think at the level of a human — many held up the imitation game as the ultimate goal.

“People were not building systems for fluent dialogue. That was just too hard,” said Stuart Shieber, a Harvard computer scientist who specializes in computational linguistics, including the Turing test. “But it was an aspiration.”

By the mid-1960s, machines could chat in small ways. And even then, they fooled people into believing they were more intelligent than they really were.

In 2014, after nearly 60 years of AI research, three researchers in St. Petersburg, Russia, built a bot, called Eugene Goostman, that imitated a 13-year-old Ukrainian who had learned English as a second language. But claims from its creators — and from the news media — that it had passed the Turing test were greatly exaggerated.

When asked, “Which is bigger, a shoe box or Mount Everest?,” this bot said: “I can’t make a choice right now.” When asked, “How many legs does a camel have?,” it replied: “Something between 2 and 4. Maybe, three? :-)))”

Then, about three years later, researchers at places like Google and OpenAI began building a new kind of artificial intelligence.

Write me a sonnet

On a recent morning, I asked ChatGPT the same questions that Turing had laid out in his 1950 paper. It instantly generated a poem about the Forth Bridge:

Its red paint gleams in the morning sun

A sight to behold, for all to see

Its majesty and grandeur never done

Then it correctly added 34,957 and 70,764. It did not need 30 seconds to do so. When I laid out the end of a chess game as Turing did, it responded with typically clear, concise, confident prose. It seemed to understand the situation.

But it did not. It mistook the end of the game for the beginning. “I would move my rook to R2,” it said. “It is generally a good idea to try to develop your pieces (move them out from their starting positions) as quickly as possible in chess.”

ChatGPT is what researchers call a neural network, a mathematical system loosely modeled on the network of neurons in the brain. This is the same technology that translates between English and Spanish on services like Google Translate and identifies pedestrians as self-driving cars weave through city streets.

A neural network learns skills by analyzing data. By pinpointing patterns in thousands of photos of stop signs, for example, it can learn to recognize a stop sign.

Five years ago, Google, OpenAI and other AI labs started designing neural networks that analyzed enormous amounts of digital text, including books, news stories, Wikipedia articles and online chat logs. Researchers call them “large language models.” Pinpointing billions of distinct patterns in the way people connect words, letters and symbols, these systems learned to generate their own text.