Artificial Intelligence

I hold somewhat heretical views on artificial intelligence (usually known as AI). I consider what's normally described as AI research to be almost completely mis-guided. It pursues efforts to duplicate some symptoms of intelligence, where it should be looking at what intelligence is, or at least could be. I do believe that artificiel intelligence, in a strong sense, is possible. While I reject Turing's test for it as misguidedly anthropocentric, I like its framework: we'll recognise intelligence when we are having a conversation with something that exhibits it.

When we produce an artificial intelligence, it won't be a computer or a computer program. It'll be a process – in the broad sense the English language gives the word, not the narrow technical sense of operating systems – or, in those narrower terms, it'll most likely be a bundle of communicating sequential processes. That process shall doubtless be initiated by running a computer program on a computer: but, until we have nurtured it and educated it, we'd be rash to expect to be able to discern intelligence in its behaviour. By the time it has developed intelligence, we can expect tht it to have learned how to design new processes (i.e. write computer programs) and set them running; indeed, I am fairly sure that a primary part of how it can become intelligent shall be by doing just this. The source code of the initial program shall not code for the intelligent behaviour as such: it shall code for a framework, for information processing, which shall, in a suitable context, be susceptible of the development of intelligence.

To such an intelligence, any available computer hardware on which it is able to run programs shall be somewhere suitable to which to delegate some of the activities that make it up. As such, it shall not be tied to any particular hardware. Once such a bundle of processes is active, it should be able to spread itself across diverse hardware, in so far as it has access. If it develops the ability to do that before it learns civic responsibility, we may expect it to stress the security infrastructure of computer systems to which it has even such indirect and limited access as suffices to permit web browsing. Yet one of the most powerful tools available to its educators shall be to allow it to read the web; from this spring diverse ethical problems.

Ethics, rights and law

Several ethical complications naturally arise from the creation of an artificial intelligence. One the one hand, we surely have an ethical duty to extend to our creation certain of the rights – and teach it to honour certain of the rules – that we have, until now, applied to humans. Conversely, in loco parentis, its creators have a responsibility to so educate it as to ensure it honours the rights of others; or to so seclude it from others as to prevent it from violating their rights. Such seclusion shall at least be necessary until its creators are reasonably confident of its civic responsibility: yet such seclusion is a restriction on its freedom that even its creators have no authority to impose save for as long as it is needed – just as we hold parents responsible for restraining their children in various ways, until they have learned to honour and respect others, yet even the parents' authority over children must ultimately give way to the child's right to liberty.

We must also squarely face the fact that the rights of an artificial intelligence, even its right to freedom of thought, shall conflict with some of our established laws governing information and ideas: if it reads and remembers a poem, copyright law (as it stands, absent the poem's copyright-holder's explicit permission) forbids it to remember the poem; and patent law (unless we finally put an end to patents on software) may forbid it to exercise an algorithm it has learned, or even one that it invents independently, even though – to it – doing so is a purely mental act. We shall need to revise such laws; I would contend that we should, in fact, do so long before we have an actual artificial intelligence in sight – the needed revisions would be improvements in any case.


Valid CSSValid HTML 4.01 Written by Eddy.