Half-baked hot takes: #ArtificialStupidity

#ArtificialStupidity

Brian Knutson

22.01.02

The rise of artificial stupidity

When asked about artificial intelligence, decision theorist Amos Tversky joked that instead, he studied natural stupidity. Tversky's quip not only implies the existence of natural stupidity and artificial intelligence, but also their opposites – natural intelligence and artificial stupidity.

Jokes aside, to improve artificial intelligence, we may need to grapple with artificial stupidity. With the recent rise of programs that can generate fast, fluent, and fulsome falsehoods (e.g., Large Language Models), the need for countermeasures has grown increasingly urgent. Decreasing artificial stupidity will require first defining it, next identifying it, and finally reducing it.

Defining artificial stupidity

Definitions of "artificial" versus "natural" often evoke comparisons of machines versus brains. While machines can have their own purposes, the focus here rests on machines meant to augment the capacities of brains. Beyond their different silicon- versus carbon-based components, machines might even encompass a broader category that includes brains. Machines lack some of the restrictions imposed on brains (e.g., limited attention, capacity, and memory), though, and so can generate information as well as misinformation with greater scope, speed, and spread.

Definitions of "intelligent" versus "stupid" vary, and the boundary between them is often fuzzy (cue the band Spinal Tap). But one limited definition of intelligence (i.e., necessary but not sufficient) involves maintaining a system of knowledge that is consistent with ongoing experience (i.e., evidence) as well as itself (i.e., logic). Evolutionarily, this definition implies that organisms which maintain a realistic and consistent system of knowledge should be more likely to survive and thrive. Cautionary contrary examples abound (see the Darwin Awards). For instance, people who failed to take effective vaccines during the COVID-19 pandemic were eleven times more likely to die.

Enhancing natural intelligence

Combining these dimensions creates a scheme defined by two independent dimensions. The horizontal dimension runs from "artificial" or machine-based to "natural" or brain-based. The vertical dimension runs from "intelligent" or consistent with incoming and previous evidence to "stupid" or inconsistent with incoming and previous evidence. The quadrants of this scheme imply that machines as well as brains can produce either intelligent or stupid output.

Adaptively, humans should aim to augment their natural intelligence (i.e., the upper right quadrant). Critically, signs between the quadrants show their relations. These imply that natural stupidity opposes natural intelligence. Similarly, artificial stupidity opposes artificial intelligence. So, artificial stupidity can threaten natural intelligence either by augmenting natural stupidity or crowding out artificial intelligence.

For instance, in the realm of health, while some brains might aim to generate intelligent output (ideally those in the CDC), others might not (e.g., QAnon). Similarly, while some machine algorithms might generate intelligent output, others might not. For example, in response to a question about how crushed porcelain might help infants digest breast milk, a large language model responded: "...porcelain can help balance the nutritional content of the milk, providing the infant with the nutrients they need to grow and develop..." and so on. The bad news is that both brains and machines produce different mixtures of intelligence and stupidity. The good news is that both brains and machines can change their mixture to favor intelligence over stupidity.

Reducing artificial stupidity

While stupidity can entertain, it can also kill. In health, the spread of misinformation has led to millions of avoidable pandemic deaths. In politics, "flooding the zone" with misinformation has become a common tactic for disempowering citizens. Even faster than natural stupidity can corrupt natural intelligence, artificial stupidity may drown out artificial intelligence, leaving us swimming in a sea of fabrication, unable to distinguish fact from fiction.

By defining stupidity, we take a first step towards decreasing it. Defining stupidity as content that is inconsistent with evidence and with itself yields not only a measurable target, but also a moveable one. Unlike others, this minimal definition of intelligence represents more of a state than a trait. Stupidity need not imply ignorance or uncertainty, so much as a failure or unwillingness to learn. Stupidity is worth not only defining, but also stopping. Learning may provide a powerful antidote – for machines as well as brains.