We find ourselves in a peculiar historical moment. The brightest (and most expensive) minds in technology are locked in a frantic race to build Artificial General Intelligence—a synthetic mind to match, or surpass, our own.
This quest, dominating headlines, financial markets, and possibly creating the biggest speculative bubble in human history, is predicated on a singular, almost obsessive focus on intelligence.
Yet, in the rush toward simulating what we see as smart, we’re neglecting to consider its supposed opposite: stupidity. Because if we’re teaching a machine to be intelligent, the least it should know is when it’s being stupid, right?
This is a harder problem on many levels than one might assume. First of all, humans generally prefer to write and think about intelligence—perhaps because when writing and thinking about stupidity, one would feel compelled to prove one is not stupid themselves.
The term “stupid” is often used as a blunt instrument, a casual and broad insult implying a lack of intelligence. But this is a lazy definition. The real phenomenon however is far more complex and interesting. Intelligent people, after all, do remarkably stupid things with alarming frequency.
We recognise stupid when we see it. We also (sometimes) recognize it in ourselves. And yet, for behaviour that is so abundantly present in society, it’s damned hard to define accurately—or measure.
In the late 1930’s, the Austrian writer Robert Musil held a lecture titled “Über die Dummheit” on stupidity in which he tried to define what stupidity might be. According to him, it was not akin to dumbness, but rather something far more dangerous; an exhibition of a person lacking the correct conceptual framework to deal with a situation, instead falling back on crude, unchallenged generalisations as a solution.
Stupidity in that regard is indeed less of an individual failing than a societal one—of which Musil was acutely aware; he died a few years later exiled in Switzerland, after his writings were banned by the Nazis— because most of our concepts, generalisations and mental tools are taught to us by the societies we’re raised in.
This is the reason we can look back and wonder how people were stupid enough to put lead in paint, radium in beauty products, leeches on the ill, or, in the case of this essay, convince ourselves that ChatGPT is smart.
I find it quite telling that a lecture on stupidity given in a time of mass delusion and authoritarian populism contains so many lessons for the time we’re in—one of mass delusion and authoritarian populism. There are many parallels between the utopia created around the potential of AI, by means of fear of being made obsolete (by smart computers), and the utopia sketched by authoritarian demagogues, by means of fear of being made obsolete (by scary foreigners).
The most obvious one is stupidity, and we could go down that route, but I hold your intelligence in high enough esteem that I trust you can venture down that path yourself. Yet, if we don’t know—or don’t want to admit that we know—what stupidity is, how will we teach a machine to recognise when it’s being stupid?
Subscribe to AGI Ethics News for monthly deep dives into AI’s ethical challenges.
Perhaps more empirical research can help—it’s easier to code statistical patterns than philosophical, culturally-bound concepts after all.
In 2015, a group of researchers published a paper titled “What is stupid?: People’s conception of unintelligent behavior,” in which they attempted to use the human ability of recognising stupidity to categorise its different forms. It is one of very few documented experiments attempting such a feat.
They identified three general groups. ‘Confident Ignorance’, where an individual tragically overestimates their own knowledge or skill; ‘Lack of Practicality’, a failure of common sense where a technically correct action is applied in a contextually absurd manner; and ‘Lack of Control’, where absentmindedness or impulsivity causes someone to fail to apply the knowledge they do possess.
All three of these forms of stupidity have been widely documented – and mocked – when exhibited by all the state of the art language models. Lol! The AI (or actually, the VC-backed corporation running the model) confidently assumes the flawed data is a true representation of merit, and it lacks the practical common sense to question a nonsensical pattern! Stupid machine. We have nothing to fear from it.
But looping back to Musil, this is exactly the type of stupidity we should fear the most—on two levels. First, as he said, “stupidity lulls mistrust; it ‘disarms’”, making us less aware of the risks of trusting a stupid machine. And second, by relying on data created by humans to build its conceptual frameworks, we’re building in the cultural and conceptual stupidity of the society it exists in.
The quasi-religious fervour around the potential of AI is not helping. Dietrich Bonhoeffer, a German theologian and anti-Nazi dissident who was hanged for participating in a plot to kill Hitler, wrote in a letter that “Upon closer observation, it becomes apparent that every strong upsurge of power in the public sphere, be it of a political or of a religious nature, infects a large part of humankind with stupidity.”
In this case, our biggest stupidity might be building these systems without a robust, generalisable understanding of the stupidity we want them to avoid. Because before we succeed in creating a machine that can think, we must first truly understand all the ways in which we fail to. And to understand this requirement completely, we must, according to Musil, apply what he sees as the antithesis of stupidity: humility.
Unfortunately, this is a behaviour quite apparently alien to those running the companies developing this technology. Or those championing its omnipresent rollout.
A final thought on Musil’s lecture from philosopher Sacha Golob, who wrote a great piece on it for Psyche, simply because it’s relevant on so many levels:
“If history is anything to go by, a few hundred years from now, our ancestors will find at least one part of contemporary morality almost unintelligible – ‘How could decent people ever have believed that?’ If they are not to condemn us as evil, they might well have to conclude that we were stupid.”
And quite probably, we are.
Ethical considerations:
- How much transparency and public scrutiny should there be around the development and deployment of powerful AGI systems, and how are risks versus benefits balanced in disclosure?
- How might AGIs be designed to experience or manifest humility, recognizing their place within a larger universe and thus avoiding arrogance or unchecked agency?
- How should institutions ensure ethical treatment of AGIs, particularly if they advance toward higher cognitive or sentient capacities?
Stay informed and ahead of the curve on all things AGI ethics — subscribe to AGI Ethics News and receive in-depth analysis and exclusive insights directly to your inbox every month!
Alejandro Tauber is the publisher of Brussels-based EU affairs website EUobserver. He occasionally writes about technology, science and society.