A TEST FOR MACHINE CONSCIOUSNESS
A photo or other kind of visual image could be used to test for machine consciousness. Various tests for determining such a thing have been proposed over the years, with famous examples being the ability to play strategy games well enough to compete at championship level, or an ability to converse in natural language. Prior to there being machines or software that were capable of performing such feats, both examples were thought to be uniquely human attributes. However, it is now generally acknowledged that neither chatbots nor strategy game-playing programs are conscious or even intelligent in anything other than a narrow sense. The question is: Why not?
Imagine there is a dark room inside which there has been placed a person and a machine consisting of a light sensor, speech synthesizer and loudspeaker. Whenever a light is turned on or off, both machine and person say “light” or “dark”. Although both person and machine register photons striking light-sensitive parts like retinas or photodiodes, only the person can be said to be conscious of the fact it is light (or dark). The reason why this is so has to do with how ‘information’ is classically defined, ie, as ‘the reduction of uncertainty that occurs when one among many possible states is chosen’. The machine enters one of two possible states, and so for it a state corresponds to one bit of information. But, when the person registers the light, not only is it ‘not dark’, the light is also ‘not green’, ‘not blue’, ‘not purple’. There are no elephants in the room; the room is not triangular in shape. Clearly, the person can rule out countless possibilities, whereas the machine can only rule out one.
Differentiating between many possible states is not all there is to consciousness. After all, a one mega pixel camera has a sensor chip that can record 2^1000,000 states, but that does not mean to say it is any closer to being conscious than a single photodiode. A major reason why not is because the camera’s sensor chip consists of many individual and independent photodiodes. This is very different to a brain, whose neurons (according to Henry Markham- more on him later) “are not islands. They need a group of neurons around them that turns out to be approximately the size of a column”. The neocortex is essentially composed of millions of these columns and it is incorrect to think of the brain as one organ; it is an intricate and intertwined collection of hundreds of specialised regions.
The fact that the repertoire of states available to a person cannot be divided has lead Christof Koch and Giulio Tononi to propose ‘Integrated Information Theory’ or “the availability of a large repertoire of states belonging to a single, integrated system” as a means of testing for consciousness. Since those internal states must be highly informative about the world if they are to be useful, the extent to which a candidate ‘conscious machine’ is indeed conscious could be determined in the following way: Show it a picture and ask it for a concise description. This would entail not just labelling objects in the picture, but also understanding the causal relationships between those objects in order to ascertain the gist of the image. Why does THE HUMAN bend over close to THE ENGINE of THE CAR? Because he is a mechanic trying to fix the car. On the other hand, if the AI failed to notice that the car is too small for an adult to sit in it, is made of yellow plastic and the ‘mechanic’ is a child, one might suspect that it has been explicitly programmed to conclude that the combination of ‘human’, ‘car’, ‘spanners’ equals ‘professional mechanic’ in which case it would fail the IIT Test for consciousness.
Today, the amount of visual and audio footage being uploaded to the Web makes it ever more necessary to crack the problem of designing software that can perform the kinds of pattern-recognition that humans do so well. Just think of how useful a search engine that could actually understand audio and video footage would be. It could watch an online video at super-high speed and find the particular segment that you want to watch. It could help automatically edit home movies. It could scan through YouTube and remove copyrighted material.
On what might be a darker note, security cameras are becoming increasingly prevalent in towns and cities, but unless somebody is watching the monitors those cameras are not really spying on us. You can bet that security firms would be very interested in software able to watch CCTV footage 24 hours a day. If I were asked to write a science fiction story detailing how we ended up in a ‘Big Brother’ society with omnipresent survaillence making privacy impossible, it would probably be based on people gradually giving up their privacy in favour of ever-more effective search engines.
