Sunday, June 29, 2008

The Toilet Test for Machine Intelligence

A machine could be considered intelligent if it makes a person uncomfortable to use the toilet in its presence.


Anonymous said...

You need to make your robots smarter than my cat, then. He watches me shower.


Tyler Streeter said...

Ha ha... your cat's a pervert.

Seriously, though, I was thinking that the binary nature of most intelligence metrics (including the toilet test) often makes them less useful because they end up calling too many things intelligent. It's probably better to avoid the sharp distinction of intelligent vs. non-intelligent and just use a continuous range.

Here's a different form of the toilet test:

"A machine's intelligence is proportional to the degree to which it makes a person uncomfortable using the toilet in its presence."

So I would guess that a machine that is about as intelligent as your cat should make you about as uncomfortable.

Anonymous said...

Wouldn't you say that the toilet test has more to do with visual cues rather than intelligence, though?

Think of a super-intelligent, sighted, interactive, non-humanoid machine (think HAL, R2D2, etc.) versus a mannequin that is only capable of following you with its eyes, but has no other function.

If you had to choose, which would you be more comfortable with having installed across from your toilet?


Tyler Streeter said...

I think would be more comfortable with the mannequin. It would still be pretty creepy, of course, but more of a shallow kind of creepiness. Better than knowing that a sentient machine is *choosing* to watch you.

My idea was that this test would be based upon a person's prior subjective assessment of a machine's intelligence, which would take place over a period of time (maybe a few weeks) before performing the test. (It could be a machine that you know personally, like some kind of household worker bot.) In that sense this is reminiscent of Turing's original test, where you sort of "get to know" the machine for a while before making a final decision.

If you had never met the machine before the test, then I think you're right... the decision would be based more upon visual cues.

Anonymous said...

As funny as this is to contemplate, I think it may be a fallacy to suggest that just a generic intelligence level causes us to be uncomfortable with someone else in the loo. I would argue that it's far more a function of context and social cues/situation than anything.

For example, someone may be horrified by a relative or friend having witnessed them on the john, but a perfect stranger (maybe a janitor/washroom attendant) might somehow be less appalling.

Stepping farther, you could also consider someone who is completely disgusted by an animal watching during the deed, yet they have no problems with the far more intelligent partner being present.

I think that if there was some sort of super-humanly intelligent bidet, we would still feel more comfortable with it than our pets. The bidet would have a prescribed place in society's bathroom paradigm, while the pet feels out-of-place at that moment.

It's not really contradictory, it's just something you have to control for.

Tyler Streeter said...


I think you're right that there is no generic intelligence level that would "pass the test." For one thing, the threshold would be different for each person (so this definition of intelligence is subjective, but then... so is the Turing test to some extent). But, as you mentioned, even for the same person the threshold is different for each context. I would bet, though, that a person's threshold wouldn't vary too much around the mean for different situations.

About the super-intelligent bidet: that's sort of a different situation than what I was getting at. I'm imagining a general purpose household robot that would interact with you outside the restroom, too. You're probably right, though, that a person would be comfortable with their sentient bidet as long as it was confined to that context.

One thing I have realized, regarding ambiguous ideas like intelligence, is that it's impossible to create definitions that make sense in all situations. Instead, I think we just have to invent definitions that are useful for our particular purposes. Binary definitions of intelligence (like the Turing test) are useful in a certain way because they create a sharp dividing line between categories. I seriously doubt that this toilet test definition is useful in most situations, but it might make people think about the issue in a new way - it's informality is its virtue.