What happens if you take a sophisticated, artificially intelligent software programme, feed it 2 million clickbait news articles, and then ask it to write its own?
You get something a lot like Click-o-Tron.
It’s the brainchild of Lars Eidnes, a software developer from Norway, and is an experiment in recurring neural networks, a type of learning AI technology.
It looks like your standard viral news site, but its headlines read like BuzzFeed or Upworthy through a black mirror: “A Tour Of The Future Of Hot Dogs In The United States.” “Rihanna Shows Off A Black Fan At London Concert.” “How To Get Your Kids To See The Light.”
And as well they might: A recurring neural network (RNN) learns by analysing a data set and trying to derive lessons from it. To create Click-o-Tron, Eidnes took 2 million articles from sources including BuzzFeed, Gawker, Huffington Post, and Upworthy, plugged them into a pre-existing RNN (with some alterations). And then he let it run.
The results are spewed out like clockwork onto Click-o-Tron, a new article once every twenty minutes. Photos are taken from Wikimedia Commons, and are supposed to — but don’t always — correspond with the headline.
“Modern writers have become very good at squeezing out the maximum clickability out of every headline. But this sort of writing seems formulaic and unoriginal,” Eidnes wrote in a blog post explaining the tech behind Click-o-Tron. “What if we could automate the writing of these, thus freeing up clickbait writers to do useful work?”
He continues: “If this sort of writing truly is formulaic and unoriginal, we should be able to produce it automatically. Using Recurrent Neural Networks, we can try to pull this off.”
Some of the headlines are, clearly, garbage. Take “82nd & Light: Designer G 21 Jobs,” or “Boy – Eating Company Has Become The Family And School.” But others are almost indistinguishable from reality:
- “Keeping Out The American Dream.”
- “Mary J. Williams On Coming Out As A Woman.”
- “How To Get T-Pain” (my personal favourite).
The RNN isn’t just throwing words taken from the headlines together at random. It actually looks for patterns, and tries to figure out what works. For example, Eidnes writes, it completes “Kim Kardashian Says …” with “She Looks Fake,” “She’s A Hero,” and “She’s Married With A Baby In New Mexico.”
“Barack Obama Says …” meanwhile, completes with “It’s Wrong To Talk About Iraq,” “He’s like ‘A Single Mother’ And ‘Over The Top,'” and “He Is Wrong.”
Click-o-Tron “gives us an an infinite source of useless journalism, available at no cost,” Eidnes concludes on his blog. “If I remember correctly from economics class, this should drive the market value of useless journalism down to zero, forcing other producers of useless journalism to produce something else.”
Eidnes told Business Insider that the site was an experiment in the technology, and he didn’t expect it to work so well as it did. “I made it from a technological interest,” he said. I want to see if I tried it, how would it work out, so this would be a fun experiment. The results were a lot more hilarious than I thought they would be — I was laughing all the way through this.”
Click-o-Tron isn’t as outlandish as it first sounds. Robots already exist to automatically do journalism. Bots exist that automatically aggregate financial data and post articles; an AP bot publishes 3,000 such pieces every quarter, The Verge reported in January 2015. Similar software also exists for reporting on sports.
If it contains easily quantifiable data (scores, revenues, and so on), then a robot can theoretically learn its formula. Capturing the more nebulous vagaries of human interaction is more difficult, however.
So how soon until the Click-o-Tron produces an article capable of tricking the reader into thinking it has a human author? “As soon as someone Googles something and they end up on the site I made, I think they will be fooled pretty quickly … but I think that’s more about [how] humans are easily fooled.”