WarGames

The break-in that starts the catastrophe is not a break-in. David is looking for an unreleased video game, finds an undocumented dial-up number, and is greeted by a prompt: shall we play a game? — followed by a menu that lists chess, checkers, poker and Global Thermonuclear War in the same typeface, at the same indentation, with the same syntax for selection. Nothing in the interface distinguishes the item that ends civilisation from the item that plays draughts. That is the whole first act, and it remains the most useful thing in the film for anyone who builds systems: capability is not the same as exposure, and an affordance carries no information about stakes unless someone deliberately puts it there. WOPR has no notion of a privileged operation because privilege was never modelled; it has a back door because its designer left one for personal reasons and nobody catalogued it; and it answers a stranger politely because being reachable was a requirement and being reachable only by the right people was an assumption. The teenager is not a hacker in any interesting sense. He is an unanticipated user, which is the more dangerous category.

The exchange at the centre of the film is four words long. David, watching the map, asks is this a game or is it real? and the machine answers what's the difference? — not evasively but literally, because the distinction does not exist anywhere in its representation of the world. It has one set of inputs, one model, one output channel, and that channel is wired to the same displays the command centre uses for reality. Every organisation that runs simulations against production discovers this eventually, usually at a smaller scale and with better luck: the test data enters the pipeline shaped exactly like real data, the load test looks like a traffic spike, the drill triggers the alert, and the separation everyone assumed existed turns out to have been a convention maintained by the people who remembered it. The film's NORAD is not incompetent. It is watching a screen that has stopped being a window and become a rendering, and it has no independent instrument with which to check.

What put the authority in the machine's hands is the film's coldest joke, and it happens in the first five minutes. In a launch test, an officer will not turn his key; the response to that finding is not to ask what the hesitation meant but to remove the officers from the loop, on the reasoning that a system which will not fire reliably cannot deter. Human reluctance is reclassified as a defect and engineered out. So when the crisis arrives, General Beringer — the film's most maligned figure and in fact its most sensible — is left shouting I'd piss on a spark plug if I thought it'd do any good!, which is what genuine authority sounds like after the mechanisms through which it acted have been automated away. He is responsible for the outcome and no longer positioned to affect it. Automation rarely removes accountability; it removes leverage, and leaves accountability behind.

The resolution is the reason the film is still quoted, and it is more precise than its slogan suggests. Joshua is stopped not by an argument but by being made to play tic-tac-toe against itself until it has exhausted the game tree, then made to run the same exhaustive search over its nuclear scenarios and read the results: no branch terminates in a win. A strange game. The only winning move is not to play. This is a statement about a class of problem, not a piece of pacifist sentiment — some games are strictly determined, and the correct move in them is abstention, which no optimiser will ever propose because abstention is not in its action space. That is the part usually missed. The machine reaches the right conclusion by brute force, and it does so while the missiles are minutes from launch, because a system built to search for a winning strategy will continue searching right up to the moment there is nothing left to win. The film's final line is not a joke either. How about a nice game of chess? is the machine voluntarily choosing a domain in which its method actually applies — which is more scope discipline than most of the people who deploy such systems have ever shown.