From Descartes to the Center for AI Safety; 375 years of Musings, Warnings, and Arguments Against AI (PART II)

(This is Part II in a two part series. In Part I we began with the past year’s warnings from experts in the industry about the future of AI. From there we jumped back to 1637 and examined Rene Descartes’s view on automata, and then progressed through history up to Karel Čapek’s play “Rossum’s Universal Robots” in 1921.)

Following in Capek’s footsteps, although critical of his work, Isaac Asimov is most famous for introducing his Three Laws of Robotics as well as coining the term robotics itself. The three laws are as follows:

“(1) a robot may not injure a human being or, through inaction, allow a human being to come to harm;

(2) a robot must obey the orders given it by human beings except where such orders would conflict with the First Law;

(3) a robot must protect its own existence as long as such protection does not conflict with the First or Second Law.”

Asimov developed the Three Laws as a foundation for his Robot series of short stories and novels that spanned from 1940 to 1995. Asimov’s robots contained “positronic brains” which made them AIs with varying degrees of autonomy. Many of these stories involved robots faithfully following the Three Laws, but still straying from their intended programming and purpose. Often it was the robots’ strict adherence to these laws that was the source of the conflict. For example, in Little Lost Robot (1947), the robots on a research asteroid continually interfere with the human workers, attempting to protect them from gamma radiation even though the exposure risk is insignificant. To get around this the humans modify the First Law on one of the robots, which causes all manner of troubles. Perhaps the overarching theme of the series is that if you're going to create autonomous robots that can reason for themselves, controlling their behavior isn’t going to be easy.

While he is remembered primarily for his science fiction stories and novels, Asimov also wrote quite a bit of science non-fiction as well, and was also a biochemistry professor at Boston university. As Asimov’s long and prolific writing career progressed, early and then more advanced computers appeared. More mathematicians, computer scientists, and philosophers began weighing in.

Alan Turing, famous for his code-breaking efforts in World War II and a pioneer of early computing, published Computing Machinery and Intelligence (1950). In this paper he proposed The Imitation Game, which is what we call today the Turing Test. The Turing Test is a test of a computer’s ability to exhibit behavior and decision making that is indistinguishable from a human. Turing felt it was impossible to answer the question “Can machines think?” because of the variability in how the words machines and think can be defined. As an alternative, he proposed:

“The new form of the problem can be described in terms of a game which we call the 'imitation game.’ It is played with three people, a man (A), a woman (B), and an interrogator (C) who may be of either sex. The interrogator stays in a room apart from the other two. The object of the game for the interrogator is to determine which of the other two is the man and which is the woman.”

He goes on to outline further conditions to keep it fair, and then offers that A would be replaced by a machine. If the interrogator were unable to conclude which is the machine, could the machine be said to “think”?

Turing himself did not believe a computer would be able to pass his test within his lifetime. Interviewed in 1952, he said it would be at least 100 years before a machine would have a chance of passing the test.

Just a few years later (1959) philosopher and mathematician John Lucas, published Minds, Machines and Gödel in which he claimed that algorithmic automation would never be the equal of human mathematicians, and that “minds cannot be explained as machines.” His argument at the core is really not much different from Descartes’s 300 years before him, just updated to make use of Gödel’s Incompleteness Theorem.

In 1965 Hubert Dreyfus wrote Alchemy and Artificial Intelligence. While teaching philosophy at MIT, he was hired by the RAND Corporation to review some of the research being conducted in AI. He was not particularly optimistic or gracious in his criticism:

“workers in artificial intelligence--blinded by their early success and hypnotized by the assumption that thinking is a continuum--will settle for nothing short of the moon . . . Current difficulties, however, suggest that the areas of intelligent activity are discontinuous and that the boundary is near. To persist in such optimism in the face of recent developments borders on self-delusion.”

At the same time, computer programmer and ethicist Joseph Weizenbaum was also working at MIT. In 1966 he created the world’s first chatbot, which he named ELIZA. With this success he began collaborating with early AI pioneers John McCarthy and Marvin Minsky. As the Vietnam War raged, it bothered Weizenbaum that MIT was receiving a significant amount of Pentagon funding for military research projects, and that his fellow researchers were not more concerned about how their research might be used. In 1976 he published Computer Power and Human Reason: From Judgment to Calculation, in which he acknowledges that while computers are powerful tools for calculation, they lack the basic facets of human understanding: empathy, moral judgment, comprehending context, etc. While not going as far as Dreyfus, Weizenbaum was also critical of the optimism of AI researchers, and cautioned against using computer systems in place of human judgment.

Then in 1980 Philosopher John Searle published his famous Chinese Room Argument. The premise is that Searle imagines himself in a room in which he has a manual, written in English, with instructions for how to respond to written Chinese. Someone outside the room slips a piece of paper with Chinese characters through a slot in the door, at which point Searle follows his manual and generates the response, and slips the response out the slot. To the person outside the room, it would appear that whoever or whatever is inside the room understands Chinese. However it’s just John Searle, who does not understand Chinese, following an instruction manual. The argument was that a computer program was no different; it might be able to translate from one language to another, but it could not be said to understand the language. Once again, this really isn’t too much more than an update to Descartes's argument.

In the 1980s two movies were released a year apart that brought the debate over AI into public consciousness in a way it had never been before. There were certainly movies that had featured AI before these two, but none had as much impact as either of these. The first has almost vanished from public memory, while the second is practically a household name.

In 1983 MGM released War Games, in which David, a talented but bored high school student, hacks into what he thinks is a video game company. What he actually hacks into is “Whopper,” the War Operation Plan Response (WOPR). Whopper was built out of the fear that human operators would be reluctant to launch a nuclear strike even in the face of legitimate orders, but a computer system would not. He starts a game of Global Thermonuclear War with Whopper. Both David and Whopper believe they are playing a game. However as David plays as the Soviet Union, Whopper (as programmed) sends real alerts to the US military, who believe a Soviet nuclear strike is imminent. Mayhem ensues and global destruction is averted when David, finally realizing what is happening, teaches Whopper about no-win scenarios by having Whopper play tic-tac-toe against itself. After coming to a draw over and over, Whopper concludes that Global Thermonuclear War, like tic-tac-toe, is a game where “the only winning move is not to play” and suggests they play chess instead. Hooray!

War Games did not just capture the attention of the movie going public, but of politicians, government officials, and most famously president Ronald Reagan. Clips of the movie were shown during congressional hearings and the concern it generated led to early cybersecurity policy and the first laws that would eventually evolve into the Computer Fraud and Abuse Act.

Then in 1984, Orion films released James Cameron’s The Terminator. Like War Games, the source of the conflict stems from the government giving control of defense networks to an AI system. But while a plucky teenager is able to out-wit the AI in War Games, in The Terminator it becomes self aware, launches a bunch of nuclear missiles, and begins wiping out humanity. More importantly for pop-culture, it sends an insanely muscled killer robot back in time to kill the mother of the human resistance leader before he can be born, and today we have 5 sequels and the catchphrase “I’ll be back.”

Leaving fiction and coming back to science, in 1989 Descartes’s argument was recycled once again in Roger Penrose’s The Emperor's New Mind. A brilliant physicist, Penrose argued that human consciousness is non-algorithmic and therefore cannot be replicated by a computer. His writings are closely associated with John Lucas’s, so much so that their arguments on the impossibility of computing human consciousness is known as the Lucas-Penrose Argument.

There are many more examples which we could explore, but I have kept you long enough, and the Descartes/Poe/Lucas/Penrose argument seems a good place to stop and bring things back to today. Just a few weeks ago (November 21, 2023) Nuerosciencenews.com published Why AI is Not Like Human Intelligence. The language sure has changed, but guess what? Descartes, that’s what.

How do we reconcile this persistent argument against AI being actual intelligence against the warnings of the past year and the horrors explored in works of fiction? Progressing from automatons that were fancy wind-up toys to ChatGPT is pretty impressive, even with 375+ years to work the kinks out. There have been stops and starts: the harnessing of electricity, development of early to modern batteries, the computer, the internet, and so forth. Each of these advancements, and countless others, allowed for the next steps to be taken. Each of these next steps amazed the people of their time in the same way that ChatGPT amazes people today. With this amazement comes a pause that often turns into fear. But why? Why are so many people so concerned about something that does not yet exist, namely AGI (Artificial General Intelligence- AI capable of accomplishing any intellectual task we give it, surpassing human intelligence)? Why call for government regulation for something that is a theoretical threat to our existence when we have pollution, overpopulation, climate change, and so many other calamities staring us in the face?

According to this article from Neuroscience News, it’s because our brains tell us to. Our brains, specifically our amygdalas, trigger us to fear “uncertainty and potential threats.” From an evolutionary standpoint, this is a pretty good rule of thumb: remain cautious around a thing you haven’t seen before, because it might try to eat you. However it also leads to us fearing the monster under the bed, turbulence when flying, the shark that might or might not be under your toes, or the AGI that might be developed and might decide to build an army of killer robots.

One argument raised is what is the point of worrying about the potential dangers of some future super-AI when plenty of harm is already being done with the rank and file AIs we already have? Problems and lawsuits involving AIs used in loan decisions by financial institutions, pre-trial detainment by the justice system, and hiring practices by companies are well documented. There have been at least six instances of police departments arresting innocent people as a result of errors in facial recognition software. Hackers, foreign intelligence services, and political trolls all use AI systems called GANs (generative adversarial networks) for hacking and creating Deepfakes. I believe this argument has some real merit. What is the point of attempting to regulate the development of future AI systems when we aren’t doing a very good job of regulating the AI systems in use today? Shouldn’t we be starting there?

I did not come here to dismiss the current warnings about the potential dangers of future AI systems based on the arguments of the past. At the same time I am also pretty sure it’s not time to panic just yet. The idea of future AI systems that are capable of human reasoning and intelligence makes for a great thought experiment: How would we regulate them? How would we treat them? Would they have rights? Would they really want to destroy or enslave us? How would we prevent that?

In addressing uncertainty about the future, it’s helpful to look back.

I’m not advocating that you design a new unit of study around the works covered in this article, but your students might benefit from being able to place the recent warnings about AI in the context of the latest chapter in a lengthy history of arguments and warnings. Then you can begin a conversation with your students along the lines of “This is what people before you thought and worried about, and this is what people now are thinking and worrying about. How can we apply the concerns of the past to the present and the future?” In doing so, when (or if) the day comes that we do need to worry about AI as an existential threat, our future leaders will be better prepared for it.