Arguably the main challenge with videogame AI is to make NPCs - or Non Player Characters - move in a convincing manner through a particular locale. Once that is achieved, the next ruse is to make each NPC fallible in some way or other. When you hide in a room as Solid Snake, the PlayStation is fully aware of where you are hiding. It is, after all, reading your joypad responses and displaying your position on the TV. Were it to pursue each player mistake to the extent of its knowledge, Metal Gear Solid would be frustrating, short-lived experience.

In attempting to re-create human (or animal) behaviour, programmers introduce innumerable complications to any given design brief. For an example of the unpredictable nature of people in a game-based context, think of football. Or, more specifically, videogame representations of it. Football is a game with clearly defined rules. Yet, one match differs entirely from the other, no matter the eventual score.

Looking at Jon Ritman's Match Day - an early, celebrated slice of 8bit soccer life - it's easy to notice that CPU-controlled players 'home in' on the human player in possession of the ball. Their appreciation of the tactical aspects of the sport are restricted to getting the ball, and hoofing it upfield. On paper, that doesn't seem to be unduly artificial behaviour. Fans of the much-maligned Wimbledon will understand that it is.

Watch a game of ISS '98 on the N64, and the subtleties of Konami's game are immediately apparent. The ball moves in far more directions than Match Day's simple angles, thus increasing the demands on AI routines. Again, the programmer's code is perfectly capable of scoring at any given point. Aware of your goalkeeper's ability to dive, and the power of the shot a player can make, Konami's AI routines are effectively hamstrung. This is, potentially, the most interesting (and, in the future, important) area of intelligence programming - making a computer less able than their human opponents. There is, after all, no fun in losing.

Cyberlife's Creatures, or its sequel, are very different in their approach to artificial intelligence. Primarily, they differ because they attribute a larger number of 'registers' to each individual character. Let's look at the behaviour of a typical Norn. Is it happy? That can be one register - a yes, a no, or somewhere in between. Is it hungry? If not, when will it become hungry? By adding similar attributes, the 'character' becomes complete. Desires and emotions can be simulated via simple (or complex) maths, with a low total engendering a smiling Norn, a high figure leading to more adverse expression.

Little Computer People - an old 8bit favourite, being very much like a basic Tamagotchi in a limited domestic environment - may appear vastly divorced from Creatures, due to the visual gulf between the two. In truth, they both preach a common gospel. Artificial intelligence is an exercise in convincing a player that on-screen life actually 'lives'. No matter the relative complexity, games seek to fool a player that their actions are acknowledged - and considered - by the console or computer in question.

On a curious closing note, it's worth examining the early Sim City games. When you create a city, you also introduce an elementary consciousness into the world. The 'sims' - as Maxis call them - are an androgynous collective that react to your every decision. Build a hospital, and they are pleased. Pay no regard to pollution, however, and they become less satisfied with their lot. This is indicated, in a low-tech, unaffected manner, by text-based opinion poll screens. Turn your computer off, and you effectively 'kill' a basic, yet living, thinking organism. Now that is frightening.