Story’s End: the artificial author

AIs will transform traditional story-writing and storytelling.

MindTheYarn is about the stories we tell, and the people we become by telling them. I’ll begin it with a story about The End of storytelling.

Okay, maybe not the end. But a quiet, worrying evolution: artificial authors.

Work, shaken and stirred

The Word Economic Forum recently released it’s 2018 Future of Jobs report, in which it concludes that more than half of all current workplace tasks will be performed by machines by 2025–just six years from now.

THe WEF estimates that in those six years, automation will kill off 75 million jobs worldwide, while creating 133 million jobs. The new jobs will be in areas like data analytics, programming, and advanced materials. A positive outlook, in one way. A cynical view is that humans could become attendants to the machines we create.

Human janitor cleans up around robot white-collar workers.
Image credit: Nathan Daniels

Isn’t this just the same old?  New technologies and automation have been taking over drudge work since before the Romans. There has always been fear and hysteria, and we’ve always moved past it. There’s one interesting difference this time: the machines are not just faster and stronger, they’re smarter.  They’re taking over the brain-work.  Surgeons, scientists, programmers, musicians, and yes, even novelists, are in the cross-hairs.

AI, insurgent

Machines are doing jobs that were done only by people five or ten years ago, like answering telephone support calls, flying passenger planes unaided, and writing news articles. Often, the people receiving these services don’t even know they’re being served by a machine.

This will have have profound social consequences, of a scale similar to those of the industrial revolution and the taming of fire.  Profound doesn’t necessarily mean bad, but the changes will be disruptive, and it won’t feel good to be on the wrong side of the disruption.

The Grace study is one of the more serious forecasts of how soon AIs will be able to perform as well or better than the best humans at similar tasks. The authors took a ‘wisdom of crowds’ approach, but with domain experts instead of punters. The predictions are less histrionic than Elon Musk’s.

I’ve pulled out one chart from the Grace study, highlighting an item relevant to storytelling.  The survey asked “When will an AI be able to write a novel good enough to make it to the New York Times best-seller list?”  The answer from the crowd was: 33 years after 2016, give or take. In other words, by around year 2049, AIs will be just as entertaining as Dan Brown or Danielle Steele.

Here’s a snippet that shows achievements predicted in the near term.  In eight years (the median guess), machines will write high school essays that get top grades and pass plagiarism tests.  Think of the job losses: all those smart, cash-strapped kids who survive by spitting out essays for the rich, lazy ones. And pop music composer?  With nine years left on the clock, it’s as bad a career choice as truck driver.

Rise of the robo-journalist

About eighteen years ago, I visited a Reuters office in New Jersey. My contact there proudly showed off a PC sitting in an otherwise unoccupied work cubicle.  This PC, he explained, was writing those little ‘market color’ articles that thousands of traders and other business people would see zip by on their newswire channels.

Market up 200 points on pending Fed announcement.
Market off highs on presidential polls.
Market drops to new low on Vatican crisis.

The simple program running on that PC made entirely spurious connections between recent news stories and recent market price movements. It filled in a template (“Market [up|down|whatever] on [recent event]”, and injected the tidbit into the feed. Rinse, repeat.

Thousands of people wasted precious seconds reading the stuff. I think most, at some level, understood that the little ‘stories’ were nonsense, but they seemed to fill a need. They’d been worth paying a hack to write before the program existed, and you still see them to this day.

Fast forward to 2019.

You know those community newspapers you can pick up free in the mall?  Cheap paper, full of ads and stories about the local team, last week’s crimes, and Mrs. Tucker’s restaurant going bust. The quality isn’t great, but you don’t expect it to be.  They’re comforting. They give you a warm sense of place, of community, a feeling that the whole world hasn’t yet been assimilated into the Borg.

Think again.  That local mall paper was probably written by some tireless, humming machines in a mile-long data center in Texas.  It’s a collection of contributions from several automated services, consolidated and arranged for maximum ad-revenue extraction by another program.  Here’s a human-written story about Patch, one of the companies that does this.

For example, most sports stories–the ones about college football, about minor league games–are machine-generated.  Associated Press has a service for that. The AI does quite a good job. It receives the scores and team compositions as soon as they’re recorded, does some research on the players, their quirks, previous games, and troubles, taps its artificial neural net to get the sports vernacular right and inject the right amount of verve into its writing style, then spits out a story.

Here’s a sample, written by a different AI, quoted in a Deadspin article:

… Kenny O’Brien gave the Cavaliers fits on the mound. Virginia managed just three hits off of the Colonials’ pitcher, who allowed no earned runs, walked two and struck out one during his four innings of work.

Twenty-seven Colonials came to the plate and the Virginia pitcher vanquished them all, pitching perfect game. He struck out 10 batters while recording his momentous feat. W. Roberts got Ryan Thomas to ground out for the final out of the game.

Tom Gately couldn’t get it done on the rubber for George Washington, taking a loss. …

Could you do better?

Google ‘robo-journalism’ and you’ll find more interesting examples.  One of the most intriguing is the GPT2 fake news generator, developed by OpenAI, a brain shop founded by Elon Musk.  A journalist starts an article, then the AI completes it, even adopting the journalist’s distinctive writing ‘voice’.  The GPT2 neural net was trained on a data set consisting of ten million articles. It writes plausible prose, and sounds like it knows what it’s saying. Only it makes it all up.

Want to spread disinformation about Hillary Clinton, or get people to vote for Trump? Give GPT2 a few leading lines with the right bias, and it will generate a million fake news stories, all different, all convincing.

OpenAI is funded by Peter Thiel, a Trump supporter whom some suspect of lending his influence over Facebook and his own data-mining companies to get Trump elected the first time around.

The Great Automatic Grammatizator arrives

Detail of the original cover of The Great Automated Grammatizator
Detail of the original cover

In Roald Dahl’s prophetic story, The Great Automatic Grammatizator, a frustrated author uses the ‘almost mathematical’ rules of grammar as the basis for a machine that composes stories. The writing algorithm’s controls are funny, whimsical, and weirdly plausible. It has switches for style and humor, and foot pedals to control passion.

The machine is successful, and soon its designer is buying up the names (brands) of all important human authors, until only one remains–a sad fellow with nine hungry children and no income.

Far-fetched whimsy?  Maybe not.

2018 saw the publication of the first novel billed as machine-written. The article contains some short samples. They’re weird.  We can’t be sure it’s the first, either. The same article would give the distinction of ‘first’ to The Policeman’s Beard is Half Constructed, written by a program in the ’80s.

Others are working on humbler writing assistants, like the one that helped write Mr. Penumbra’s 24-hour bookstore.

Perhaps the most brazenly commercial is Philip M. Parker’s patented system, that has apparently written 800,000 non-fiction books. 100,000 of them are for sale on Amazon, for prices ranging from $2.99 to $995.00. Topics include a global survey of Lambskin Condoms, a bibliography on Charlotte-Marie-Tooth-Disease, and a dictionary of Cebuano. Exciting stuff.

Detail of book entitled 'The 2020-2025 World Outlook for Lambskin Condoms'

And no, you can’t look inside before you buy. It’s the ‘long tail’, author edition: spread a wide net.

But these are early days. According to the Grace study, we shouldn’t expect that NYT best-seller until 2049. I can hardly wait.