AI anthropomorphization
I often encounter arguments criticizing the tendency for humans to anthropomorphize AI such as LLMs. I think this criticism is hasty.
Humans have always anthropomorphized
Clifford Nass argued in his 1993 paper “Anthropomorphism, Agency, and Ethopoeia: Computers as Social Actors” that even when computer interfaces provide only minimal ‘social cues,’ users tend to anthropomorphize computers, and that through this it is possible to induce users to engage in social interaction with computers. (CASA paradigm)
The tendency itself for people to anthropomorphize computers had already been known for some time, but Nass’s research was significant in two respects.
- First, before Nass, it had been assumed that fairly sophisticated and complex social cues would be necessary to induce anthropomorphism, but Nass showed that even minimal cues could induce anthropomorphism.
- Second, before Nass, it had been thought that anthropomorphic tendencies appeared because of users’ ignorance or psychological error, but Nass showed that even users familiar with computers (that is, people who know perfectly well that computers do not have a ‘self’) anthropomorphize them.
This claim seems to extend to LLMs without much strain. Anthropomorphization of LLMs is not a new tendency, and this phenomenon occurs neither because LLMs are so excellent (Nass’s first claim) nor because people are stupid (Nass’s second claim).
The advantages of anthropomorphization
In the 1971 paper Intentional Systems, philosopher Daniel Dennett argues that in order to understand and predict the various phenomena that occur in the world, humans can adopt one of three stances.
- physical stance: a way of predicting behavior on the basis of physical laws. For example, the reason that ball is rolling is that I kicked it and applied physical force.
- design stance: a way of predicting behavior on the basis of functional design. For example, the reason the TV turned on when I pressed the remote control button is that it was designed that way.
- intentional stance: a way of predicting behavior on the basis of beliefs and desires. For example, the reason Younghee is going over there is that she wants to see the exhibition (desire) and thinks the exhibition hall is over there (belief).
Dennett says this.
I wish to examine the concept of a system whose behavior can be (at least sometimes) explained and predicted by relying on ascriptions to the system of beliefs and desires (and hopes, fears, intentions, hunches, …). I will call such systems Intentional systems. … The best chess-playing computers these days are practically inaccessible to prediction from either the design stance or the physical stance; they have become too complex for even their own designers to view from the design stance.
I wish to examine the concept of a system whose behavior can be (at least sometimes) explained and predicted by relying on ascriptions to the system of beliefs and desires (and hopes, fears, intentions, hunches, …). I will call such systems Intentional systems. … The best chess-playing computers these days are practically inaccessible to prediction from either the design stance or the physical stance; they have become too complex for even their own designers to view from the design stance.
For some kinds of systems, their behavior can be explained and predicted more easily by taking the intentional stance (often rendered in Korean as “지향적 입장”), rather than the physical stance or the design stance.
That is true of AI.
Is AI essentially different from humans?
As one of the grounds for saying AI should not be anthropomorphized, I also often see the claim that “LLMs are nothing more than gigantic matrix operations and are essentially different from the human mind.”
But the claim that, because the way some artifact is physically implemented is quite different from the human brain, that artifact therefore “has no consciousness” or is “essentially different from humans” can basically hold only if one does not accept functionalism (the view that regardless of the mode of physical implementation, if the function performed is the same, then it is the same thing). But as with many topics in philosophy of mind, it is not easy to firmly prove or dismiss functionalism.
About whether AI’s computation is essentially the same as or different from human consciousness, and in what respects and to what degree they are the same, we still know almost nothing.
Good anthropomorphization and bad anthropomorphization
“Right/wrong” and “beneficial/useless” are separate axes, and I think claims like “all anthropomorphization is inappropriate” or “the reason people anthropomorphize is that humans are stupid” are not only wrong but also not useful.
Anthropomorphization is a natural tendency, and sometimes it is beneficial. Humans anthropomorphize sometimes knowingly and sometimes without realizing it. I think it is both valid and useful to distinguish between good anthropomorphization and bad anthropomorphization.
I’ve said a few things about the advantages of anthropomorphization, so now let’s think about bad anthropomorphization.
- The providers of AI products excessively anthropomorphizing them for the purpose of making money, or designing or promoting products in ways that induce emotional identification (related research: Computers as bad social actors)
- Worshipping them(e/acc), or fearing them(longtermism)