Can AI Write the Next Game of Thrones Book?

Game of Thrones is one of the most popular TV shows of all time. Its duels, dragons and dramatic deaths make it completely unpredictable, with fans repeatedly shocked at the passing of their favourite characters. It’s this that has given it a viewership of nearly 5 million in the UK alone. But is it as unpredictable as we think? A software engineer by the name of Zack Thoutt begs to differ…

A seven-part series

Like many great TV and film productions, Game of Thrones is based on a series of books. Written by George R.R. Martin, the series of novels is actually called A Song of Ice and Fire, with the TV series’ name taken from the first volume – A Game of Thrones.

In total, it will be made up of seven novels, but – as of yet – only five have been written. The first, A Game of Thrones, was released in 1996, with further volumes in 1998, 2000, 2005 and finally the fifth, A Dance with Dragons, in 2011. Currently, Martin is working on the sixth book, The Winds of Winter, and it’s proving a painful wait for most fans.

The wait is over… Kind of

Lucky for those who can’t wait, software engineer Zack Thoutt has tried to write the sixth book himself using artificial intelligence (AI). How? He used a recurrent neural network (RNN) to go through over 5,000 pages that make up the first five books. It’s then used the information from this to attempt the sixth of the series. All Thoutt had to do was enter a prime word to begin each chapter and tell the programme how many words to write.

The results? Well, let’s just say the wait isn’t over yet. Uploaded onto development platform GitHub, the first five chapters have a few surprises in store for readers. From new characters – such as Greenbeard – as well as some new grammatical structures – like “why did you proper?” – the AI has certainly provided a mixed bag.

Improving the results

The project was clearly a bit of fun for Thoutt, to see how well AI could replicate an author’s style. However, he thinks it may have fared better if it had much more content, with simpler language. Another problem he identified was its inability to remember character deaths – leading to deceased characters being unexpectedly brought back into the story.

“It's obviously not perfect,” says Thoutt. “It isn't building a long-term story and the grammar isn't perfect. But the network is able to learn the basics of the English language and structure of George R.R. Martin's style on its own.”

The speed of science

AI novel-writing is just one example of science and technology speeding things up. But it’s not really the most successful. A better example is LC-MS, which can improve the speed of drug screening while maintaining impressive accuracy, as discussed in the article ‘Pushing the Limits of Speed and Sensitivity in Drug Screening – an LC-MS solution’