by Push Singh, MIT Media Lab

October 16, 1999

We are beginning work on a project to build a commonsense database by trying to get people on the internet to type it in.  We are still sorting out the many details, but I am anticipating a number of “brain-in-the-vat” criticisms complaining that such a system could never be intelligent, and I’d like to try to head them off immediately.

Critic: Your system will not be intelligent because ...

First, let us try to understand what people mean when their complaint starts out this way.  Let us parse this as

“Your system differs from a person because ...”
"Your system will be poor at solving the same problems that people can because ...”

Critic: Your system will never really understand what it means be an actual living person.

In much the same way that anthropologists make theories of other cultures and ethologists make theories of other species, machines can make theories of human beings.  It is possible to study and articulate the behavior of things different from yourself.  We may never really understand them in the same way that they understand themselves, but at the same time, there are ways we might understand them better than they understand themselves!  It is not uncommon for people to sometimes have, in some ways, better models of their friends than their friends have of themselves.

Critic: Your system will not ever be intelligent because it is not grounded in sensory perception.

I have three responses:

1. Once more, what is “intelligent”?  If you mean the ability to solve problems and answer questions, there are some questions a "brain in a box" will be able to answer and some it will not.  I agree that without the kind of rich learning about the sensory world that comes from living it, the system will likely be poorer than a person at answering many questions about the spatial and physical world, but is that weakness enough to render it unintelligent?  Alien, yes, but unintelligent no.
   
   
2. In general, the concept of “sensory grounding” needs to be broken down into separate visual, spatial, auditory representations.  To say that these cannot be even partially articulated is to ignore the power of words and literature.  I agree that a huge amount of what people know is spatial and physical, and difficult to articulate fully, but this doesn't imply it can't be partially articulated, and that that partial articulation isn't useful.  Nonetheless, there is nothing stopping us from gathering pictures, movies, and sounds in addition to statements of commonsense.
   
   
3. Many people think that an idea can only mean something if it has sensory grounding.  I believe that this is true in a perverse sense.  Something only comes to mean something if it has many meanings.  If you give something a sensory meaning, this helps by giving it an additional meaning.  Of course the sensory meaning is in some sense more true or reliable than some arbitrary statement we might make.  To the extent that we need many reliable meanings to make reliable inferences, sensory grounding is a good thing.  But senses can deceive us, and sometimes a statement in an encyclopedia, verified by hundreds of people, is more reliable than what our own eyes tell us.

This criticism points to a genuine problem.  Contextualizing the knowledge is important to connect pieces of knowledge that are used or learned together.  But at the same time, much of what children presumably must do when they learn is to separate accidentally associated items into separate mental bins.  In a sense, this project will start from the other direction from real children, because the knowledge is provided as isolated, separate items.  The problem then becomes re-associating these items into useful contexts, and we need to think of good ways to make it easier to do this.