Since Scrabble adopted an official lexicon in 1978, one thing has been constant: People have never stopped arguing about what is or isn't a word. Players have defended the game by noting that its letter strings--from AA (a kind of Hawaiian lava) to ZZZ (an interjection for sleep)--could be found in a bunch of standard North American dictionaries, books that have been used through the years to compile and revise Scrabble's tournament word list. But after an update last month introduced dozens of suspect words, riling up the community of competitive players, that's becoming harder to do. The linguistic tumult began in September, when the organization that maintains the word list used in club and tournament Scrabble, NASPA Games, published a draft of its update. The NASPA list includes all of the words in the Official Scrabble Players Dictionary, the go-to source for living-room and app players in North America, plus a lot more.

The 2023 World Scrabble Championship, held last month in Las Vegas, was an instant classic. The best-of-seven finals went the distance, with tense games, obscure words, strategic genius, and a Scrabble-record audience of 900 watching on Twitch. The winner was David Eldar, 33, of Melbourne, Australia, who defeated Harshan Lamabadusuriya, 44, a pediatrician who lives in Southmoor, England, to capture the $10,000 first prize. To reach the finals, they topped a field of 134 players from 29 countries--from Poland to Pakistan, Singapore to Sierra Leone--in a four-day, 32-game tournament. Game 6 of the finals was hailed by Scrabble experts as one of most exciting high-stakes games ever. I discussed it online with the two competitors. Our conversation has been edited and condensed for clarity. Note: The event used the international-English Scrabble dictionary, which includes substantially more words than the lexicon governing competitive play in North America. To avoid confusion, words acceptable only in the international word list are marked with a #. Stefan Fatsis: David, you trail three games to two and are up first in Game 6.

In the mid-1970s, top players in an emerging tournament Scrabble scene persuaded the game's corporate owner to adopt a universal lexicon for competition. Players manually scraped five standard college dictionaries, recording every unique two- through eight-letter word (plus inflections) that met the game's rules. When the Official Scrabble Players Dictionary was published, in 1978, players rejoiced. "You can retire the boxing gloves and put up your swords," the Scrabble Players Newspaper wrote. "You now have an arbiter to settle all arguments."

In recent weeks, a web-based word puzzle called Wordle has become a popular daily distraction. Suddenly, millions of people are focused on their vocabulary of five-letter words, and are newly aware of concepts like letter frequency and letter position as they strategize about the best opening words and faster solutions. For these people, Wordle is captivating. Previous research can help us understand how our brains respond to word games, and why we love them. Wordle is a single-player puzzle that combines elements of several games, including Scrabble and Battleship.

Games are often designed to shape player behavior in a desired way; however, it can be unclear how design decisions affect the space of behaviors in a game. Designers usually explore this space through human playtesting, which can be time-consuming and of limited effectiveness in exhausting the space of possible behaviors. In this paper, we propose the use of automated planning agents to simulate humans of varying skill levels to generate game playthroughs. Metrics can then be gathered from these playthroughs to evaluate the current game design and identify its potential flaws. We demonstrate this technique in two games: the popular word game Scrabble and a collectible card game of our own design named Cardonomicon. Using these case studies, we show how using simulated agents to model humans of varying skill levels allows us to extract metrics to describe game balance (in the case of Scrabble) and highlight potential design flaws (in the case of Cardonomicon).

The current state-of-the-art Scrabble agents are not learning-based but depend on truncated Monte Carlo simulations and the quality of such agents is contingent upon the time available for running the simulations. This thesis takes steps towards building a learning-based Scrabble agent using self-play. Specifically, we try to find a better function approximation for the static evaluation function used in Scrabble which determines the move goodness at a given board configuration. In this work, we experimented with evolutionary algorithms and Bayesian Optimization to learn the weights for an approximate feature-based evaluation function. However, these optimization methods were not quite effective, which lead us to explore the given problem from an Imitation Learning point of view. We also tried to imitate the ranking of moves produced by the Quackle simulation agent using supervised learning with a neural network function approximator which takes the raw representation of the Scrabble board as the input instead of using only a fixed number of handcrafted features.

The game board was classic CES misdirection--a way to set the IVS booth apart from the convention's 4,000 other exhibitors and to draw journalists like me into learning about a piece of factory machinery.

Look, there's plenty of boring industrial robots on the floor of the Las Vegas convention center for CES. But I've got to hand it to the Industrial Technology Research Institute (ITRI) -- it really knows how to make a demo fun and interactive. The company combined a number of its technologies into a robot that is able to sit and play Scrabble against a human opponent ... and win. This isn't some "AI" playing you in a digital world, to be clear. The humanoid machine, powered by what ITRI calls its Intelligent Vision System for Companion Robots, plays a physical (if slightly modified) game of Scrabble.

The BBC's Dave Lee challenged a robot to a game of Scrabble at CES 2018.

The current favored lookahead strategy is traditionally called "Monte Carlo sampling"; it has been used in several areas of endeavour including game playing for many years; I believe the first computer program to use this technique in Scrabble was Richard Hooker and Tony Guilfoyle's Quetzal; it has also been used more famously in Brian Sheppard's Maven, where he calls it "simulation". Very roughly, Monte Carlo sampling in Scrabble consists of picking a random rack from the bag such as your opponent might have, and then finding his best move using that rack. You do this for each of your own moves that you are considering playing. Typically it might do this 20,000 times, to ensure a statistically significant sample. Although in classic AI, this would be done as a Minimax, with a branching ratio of 1 at each ply (possibly with some pruning such as alpha-beta cut-off to keep the exponential explosion under control); in the Scrabble programs listed here, what they call a multi-ply search is actually done as follows: Take two candidate moves A and B. For each candidate, generate a 4-ply sequence of moves, A1-A2-A3-A4 and B1-B2-B3-B4.