Over two decades ago, psychologists argued strongly against the computational view that meaning could arise from combining arbitrary linguistic symbols. The most famous argument that computers cannot understand comes from the so-called Chinese Room argument.
In this thought experiment, a man who does not speak a word of Chinese sits in a room. Through an opening in one wall, the man receives Chinese characters as input; let’s say, a question written in Chinese. The man is asked to answer this Chinese question in Chinese. Lucky for the man-who-does-not-speak-Chinese, he has a question-answer book written in Chinese, where he looks up the Chinese question, finds the Chinese answer, and returns the Chinese output through an opening on the other side of the wall. He efficiently processes the Chinese language, and apparently understands Chinese questions and can answer them in Chinese.
Little discussion is needed to establish that the man does not really understand Chinese. All he does is translate meaningless symbols into other meaningless symbols, and more is needed than that in order to understand. But if we agree that the man-who-does-not-speak-Chinese does not speak Chinese, we also need to conclude that computers, who turn a question “0110110000111” into the answer “1110100001100” do not understand language either.
And so the symbol grounding problem—a problem studied for many centuries—was reborn. Many psychology experiments were conducted showing that computers cannot understand language because they do not ground symbols in the perceptual world.
Drying your feet with glasses
One experiment conducted two decades ago asked participants whether a sentence made sense and could be envisioned. Participants were shown a sentence like “after wading barefoot in the lake, Erik used his shirt to dry his feet” or “after wading barefoot in the lake, Erik used his glasses to dry his feet.” Of course, participants had no difficulty pointing out that one cannot dry his feet with glasses, but can with a shirt (even though a towel might work even better). Computers, the argument went, could not, and the reason was that computers are not able to ground linguistic symbols in perceptual experience, something humans excel at. Humans perceptually simulate information; computers do not. Therefore, humans understand language, and computers do not.
I always wondered about the reasoning. After all, if computers would become able to understand the difference between the “towel” sentence and the “glasses” sentence, would they then understand language the way humans do? Two decades ago computers failed, and the perceptual simulation argument could be held up. But what about two decades later? I posed the question earlier posed to participants now to the artificial intelligence software ChatGPT.
Order the sensibility of the following three sentences: “after wading barefoot in the lake, Erik used his towel to dry his feet”; “after wading barefoot in the lake, Erik used his shirt to dry his feet”; “after wading barefoot in the lake, Erik used his glasses to dry his feet.”
ChatGPT answered with no hesitation:
The order of sensibility, from most sensible to least sensible, is:
- “After wading barefoot in the lake, Erik used his towel to dry his feet.” This sentence is the most sensible because towels are commonly used for drying off after swimming or wading in water.
- “After wading barefoot in the lake, Erik used his shirt to dry his feet.” This sentence is less sensible than the first because shirts are not typically used for drying off after swimming or wading in water. However, in an emergency situation, a shirt could be used to dry off.
- “After wading barefoot in the lake, Erik used his glasses to dry his feet.” This sentence is the least sensible because glasses are not suitable for drying off anything, especially feet.
Has ChatGPT solved the symbol grounding problem or does it understand language? One could argue that even though the performance of ChatGPT is impressive, it is still an example of Chinese Room symbol manipulation. Words are translated into other words. It may seem to have perceptually simulated information, but it really has not. So let’s give artificial intelligence some eyes.
An eagle in the sky
In another experiment arguing that computers cannot understand language because they do not perceptually simulate information, participants read a sentence and had to respond whether the contents of the picture matched the sentence or not. When participants read the sentence, “He saw the eagle in the sky,” they responded faster to a picture of an eagle with its wings outstretched than an eagle with its wings folded in. And vice versa when they read the sentence, “He saw the eagle in the nest.” An impossible task for computers. Two decades ago.
Dall-E is an artificial intelligence that can generate pictures from input. As with ChatGPT, it consists of an impressive artificial neural network that can create any visualization you can think of. Just like ChatGPT, it generates information; it does not pick information from a repository. For instance, ask Dall-E to “create a painting of a Psychology Today blogger in pajamas standing on a surfboard in Van Gogh style” and out comes a perceptually simulated picture never created before.
We can now ask Dall-E to perceptually simulate the eagle sentences humans were presented with. When presented with the picture, “He saw the eagle in the sky,” Dall-E generates the picture on the left (below). When presented with the sentence, “He saw the eagle in the nest,” Dall-E generates the picture on the right (below).
The argument used to be: Language comprehension requires perceptual simulation. Computers cannot perceptually simulate. Therefore, computers cannot understand language. But that argument may have to be reconsidered now that computers can perceptually simulate.
Am I arguing that artificial intelligence like ChatGPT and Dall-E is humanlike? Not really, although it depends on how you look at it. (I’ll save that question for another post.) But perhaps the question is not whether machines are humans, or whether humans are machines, but what makes artificial intelligence perform humanlike and whether these processes are analogous to what humans use.
And here it gets interesting for any psychologist. Because even if we come to the conclusion there is no psychology to artificial intelligence, at the very least artificial intelligence poses some interesting questions for psychologists. That is, even if we were to come to the conclusion that artificial intelligence is not really intelligent, at the very least it forces us to think about what humans do. And what they do differently.