Artificial intelligence

By G.F. | SEATTLE

WORD games reward people with a vast vocabulary and the ability to sort a jumble of letters into the most unusual concatenations. Cheating is hard because consulting a dictionary or an anagram-solving website would be rather conspicuous to the other players. Online, such obvious skullduggery would go undetected. Or would it?

Take the recently launched Letterpress. The game, to which Babbage has become addicted, requires players to create words from letters displayed on a five-by-five grid. The grid is generated for each match according to rules that ensure a sufficient supply of vowels and consonants. (Letterpress is free, but 99¢ will buy you the ability to play multiple games at once.) Playing a word claims the letter tiles that comprise it, whether ones that another player currently has totted up toward his score or unclaimed tiles, with a few nuances to allow locking up letters. A word may be played just once, which also excludes any truncation of the words: sailors eliminates sailor and sail, but still allows sailing, sails and mainsail. The game proceeds until the last tile is claimed. This may happen after a handful of rounds or take nearly 100 turns, as Babbage recently discovered.

Making sense of ZVBTYRVCBPE (letters left in one recent game) can be trying. This is where abundant "cheater" apps come in. Not long after Letterpress came Letterpress Solver and a panoply of other programs purposely designed to game Letterpress. They joined the ranks of generalist cheater apps that work for a range of lexical challenges. All such programs make use of a simple expedient. Apple's mobile operating system, iOS, allows for screen capture. In order to cut corners, you grab the challenging screen in Letterpress, Scrabble or another similar game, launch the helper app and select the appropriate image from the photo library. The insidious software performs optical character recognition (OCR) to extract available letters and to analyse the grid.

Letterpress' creator, Loren Brichter (formerly of Twitter), has no clue why people cheat. The game does not track statisitics. If you win, you do not get a prize or a badge, he says. You certainly do not get money. Crucially, an opponent who had stuck to simple words and then sprouts "sesquipedalian" will inevitably stoke suspicion that he availed himself of an algorithmic aid. "Using it, you're a dick," Mr Brichter says. But building one is something else entirely. That, he thinks, "is awesome".

The simplest apps simply provide lists of potential answers. The best employ elaborate forms of artificial intelligence to provide the optimal solution. Although only a finite number of words may be formed in any given grid, with that number reduced each turn by the words played (along with truncations), an artificial-intelligence routine can look at how a word played in one round affects possible paths to victory in subsequent rounds, based on the moves available to the opponent. A single Letterpress game has fewer possible paths to walk down than chess, say, but there are nonetheless enough of them to make computing them all impractical. So, as with chess algorithms, clever techniques speed up the analysis by identifying and ditching the least promising solutions. And as with chess software, there are innumberable ways in which algorithms can be honed.

Apple has approved dozens of such programs in its online marketplace. Several developers have actually contacted Mr Brichter to ask for permission of sorts to roll out their cheater software. He obliges, asking only for a disclaimer to be included. A colleague even suggested a tournament where Letterpress AIs would battle it out among themselves, in the time-honoured tradition of contests pitting chess algorithms against one another. There, at least, no cheating would be possible.