A More General Critique of AI

There is, however, a critique of even the connectionist paradigm, set forth recently by Terry Winograd, the very same researcher who helped jumpstart the GOFAI revolution of the 1960's with his SHRDLU program. Winograd argues that the danger of the rationalist tradition is the assumption that reality has an objective existence independent of the observer, and that cognition consists of manipulating mental representations of this world. Instead, they argue, cognition is a dynamic interaction with the world, and our brain helps determine how the world is perceived, rather than simply grasping a subject world. Forester and Morrison give the example of a stick illuminated from one side with white light and from the other with red light, and the green shadow that results, even though there is no actual light shining on the stick in the range of the spectrum we normally call green. The internal pattern of our retinal states determines our perception, says Winograd, and not the other way around.

The problem with "building" a cognitive system is therefore that cognition becomes a series of provocations of the nervous system by the environment. Both the range of stimuli that can successfully provoke our nervous system and the range of possible effects of these provocations are determined by human evolution, in the case of innate cognitive properties, and in the case of learning, a smaller-scale evolutionary and historical process of interacting with the world and selecting certain stimuli and responses. That is, we have evolved to perceive green in certain circumstances, and there is no one-to-one mapping between "green in the world" and "green in our minds."

This raises quite a few major questions for AI. It indicates, for instance, that language might not be the decidable logical construct most of AI assumes it to be, but rather a historically constructed phenomenon tied closely into the nature of human experience. This would mean that for computers to develop human-like intelligence they would need to develop interactively with their environment, and while this is not outside the realm of possibility, it is not clear how this interaction would take place. Moreover, the question is raised as to whether computers can interact with their environment in a human-like way without having human-like bodies. For instance, if computers cannot see the world in a similar way that humans do, or touch objects in similar ways to humans, could they develop human-like intelligence? Or to take a more simple example, would developing a computer's intelligence to the level of a five-year old require leaving a computer turned on for five straight years, an unreasonable request of almost any computer system today? As our conception of the world around us changes, our conception of the project of artificial intelligence must change as well.

back