Rise of the robot journalist

This article was made by humans. It was written by a human, designed by a human, edited by a human and proofread by a human. It features interviews with two humans, which were transcribed by two further humans. In terms of accumulated human hours, it took the best part of a working week.

The following article, published by Associated Press on 11th November, was not made by humans:

Aerie Pharmaceuticals Inc. (AERI) on Tuesday reported a loss of $13.1 million in its third quarter. The Research Triangle Park, North Carolina-based company said it had a loss of 54 cents per share. Losses, adjusted for stock option expense, came to 44 cents per share.

Although this took less than a second to produce it’s everything a report on Aerie Pharmaceuticals’ share price should be – accurate, detailed and mercifully to-the-point. Its author is called Wordsmith. He or she – Wordsmith has yet to be anthropomorphised – is a natural-language generation platform developed by North Carolina tech company Automated Insights, and what it lacks in character it makes up for in efficiency. In 2013 it generated 300 million news reports for Automated Insights’ various clients, more than all the human journalists in the world managed between them – and at a vanishingly small proportion of the cost.

When Associated Press (AP) announced in June that it was hiring Wordsmith to write quarterly-earnings stories such as the one above, it was presented as a win-win situation. The technology would “free journalists do to more journalism”, said AP managing editor Lou Ferrara on the company’s blog. He added that finance reporters were relieved to be spared from tedious work, but admitted that a “healthy dose of scepticism” remained among his staff. Doesn’t automation always lead to job losses?

James Kotecki, manager of media and public relations at Automated Insights, insists that his technology isn’t currently putting journalists out of work. “We’re producing articles that never would have existed in the first place,” he says. “AP was doing 300 corporate-earnings stories per quarter; they’re now doing about 4,440. So 4,100 of these stories would not exist without Wordsmith.”

Robots are still rubbish at the human bits of journalism – empathy, sympathy, warmth, understanding and finding great quotes to present in large, fetching type”

The best way to think of Wordsmith, Kotecki says, is as an assistant reporter. It handles the legwork, the boring facts and figures, to which human reporters can add complex analysis and colour. “The computer handles the who, what, where and when,” he explains, paraphrasing a journalism professor at the University of North Carolina. “And humans are freed up to ask why and how.”

The LA Times uses Quakebot to automatically generate articles about earthquakes above a certain magnitude. Photo: AP/Press Association Images/Damian Dovarganes

It’s an approach that works for the LA Times, whose Quakebot software picks up data from the US Geological Survey when there’s a quake above a certain magnitude and whacks it into a pre-written article template. The bot publishes the story online, satisfying readers’ demands for speedy breaking news coverage, and then journalists can head out and produce additional content on the quake’s impact on people’s lives. Reporters are needed because robots are still rubbish at the human bits of journalism – empathy, sympathy, warmth, understanding and finding great quotes to present in large, fetching type.

Like Quakebot, Wordsmith grabs standardised data, the kind that can be inputted into a spreadsheet, and applies algorithms to it that generate journalistic language. Constructing impressive sentences isn’t a huge challenge for a big bot with big data – you could always feed it the complete works of Shakespeare – but it would still be incapable of independent creativity. However, as data becomes more varied and complex, it can begin to produce writing with human-like insight.

“We’re going to be able to do a lot more in the future,” says Kotecki. “Our algorithms will get more sophisticated and more people will collect data sets [for the purpose of] creating stories. What’s exciting is the prospect of qualitative data being turned into what feels like quantitative data, so we can turn out stories with more nuance.” One way of capturing the human experience is to detect and measure human sensory experience – touch, sight, smell, sound, taste – which can be done with sensor-driven data. Sound recognition software is advanced. Shazam has been identifying music for more than a decade. Image recognition software is getting increasingly sophisticated and companies that produce it are attracting the big digital suitors. Google paid an undisclosed sum for Jetpac, an app that analyses public location-tagged Instagram images for key indicators of human experience or culture such as smiles, cocktails or moustaches to automatically generate lists of the most happy, drunk or hip towns. The software is a work in progress: Jetpac used to assert that Turkey was full of hipsters until tweaks were made to adjust for areas where facial hair isn’t necessarily a symbol of a counter-culture.

Google recently claimed to have made its own breakthrough in image-recognition software. The project uses Jetpac-style image recognition and Wordsmith-style natural-language generation side by side so that it can scan a photograph and accurately tell you, for example, that it’s “a group of young people playing a game of frisbee”. If Google were to do to private spaces what they’ve done to public spaces with Street View, they could feasibly produce instant reports on restaurants, bars or beaches. For example, if you were a person of a certain disposition going out in Soho you could automate reports for bars on Wardour Street and go to the one with the most well-dressed people aged under 35 drinking martinis.

Thankfully, such technology could also have more substantial uses. Might a Quakebot aided by image recognition and sensors bring that people-on-the-ground element to an automated earthquake report? Could an automated report on a demonstration in Trafalgar Square give you an accurate account of the number of protesters, who’s in the crowd, what they’re chanting and what’s on their placards? Computer probably says yes. You wouldn’t necessarily need access to install cameras. Increasingly, news organisations are experimenting with commercial drones, which can be flown over hard-to-reach places to gather information. For war reporting, for instance, a drone with a camera linked to image recognition and natural-language generation software could conceivably produce real-time reports from parts of the world journalists can’t get to.

Tiny sensors to be placed under NFL players’ shoulder pads provide instant data that Wordsmith could use to conjure up live blog commentary. Photo: AP Photo/Houston Chronicle, Brett Coomer

For the time being, Wordsmith isn’t quite so ambitious. Kotecki is excited by an innovation in North America’s National Football League (NFL) which will see tiny sensors placed under players’ shoulder pads. Not only will this provide instant data that Wordsmith could use to conjure up live blog commentary, it also increases the potential for a story with human elements. If sensors can measure tackles, tumbles and touchdowns, they can provide data to generate the vocabulary of bold decisions, nasty bruises and hard-fought triumphs. Auto-reporting might give you an idea of what it feels like to collide at full speed with a seven-foot-tall, 25-stone football player, but it’s probably not going to be able to do a ‘Fever Pitch’ and describe what it feels like when, after a lifetime of waiting, your team finally wins the league.

Automated Insights points to its Yahoo fantasy football coverage to back up its claims that Wordsmith can write with the “tone, personality and variability of a human writer”. The company produces thousands of report cards algorithmically tailored for individual competitors. “Yahoo wanted us to incorporate a lot of trash talk in this iteration of our software,” says Kotecki, “so we taught Wordsmith lots of snarky phrases.” In a report compiled for the Texas Misfits, in competition with teams such as Minnie Mouse and Multiple Scorgasms, Wordsmith wrote: “With an average of 9.1 years of NFL experience, Texas Misfits has the greatest chance of losing a player to osteoporosis.”

Wordsmith, it seems, is indeed capable of being snarky. But its jokes have been dropped in a spreadsheet by data scientists and writers. It still needs humans employed to write zingers. What would save publishers even more money is software that can write its own jokes.

Simon Colton is a professor of computational creativity at the University of London and the coordinator of the What-If Machine project, an EU initiative to create software capable of creativity independent of human help. The What-If Machine takes big data – in this case a bank of documents totalling 100 million words – and seeks commonalities in their language so it can learn how words are connected together. It then generates ‘what if’ premises for stories. “What if,” one of its better attempts went, “there was a banker whose regulator ran away and she suddenly became as excessive as traffic?” They’re not there yet.

“To me these projects [such as Wordsmith] are straightforward data visualisations,” says Colton. “Instead of pie charts and graphs they do words, and they miss various aspects needed for a good read. I tell people not to worry about their creative jobs being threatened by automation because of what I call ‘the humanity gap’. Even if software was good at wit, humour and writing style, it would not be human. If you want human insight, you’re not going to get it from a computer any time soon.”

If publishers could automate short clickbait stories, will they spend the money they save on more in-depth, personal, first-person stories? No, they won’t”

“If you just want the facts and a half-decent read, a piece written by software might be OK,” Colton says. “But if software writes a piece about the joys and pitfalls of pregnancy, you’re not going to connect with it. A piece of software might eventually write in the style of Will Self or Julie Burchill, but it wouldn’t have their personalities.”

Perhaps the key to surviving in journalism in the age of automation is having a personality a robot could never succeed in – or want to succeed in – emulating. But while Julie Burchill can relax, regular reporters remain vulnerable to the harsh realities of big business. Colton is a realist: “If [publishers] could automate short clickbait stories, will they spend the money they save on more in-depth, personal, first-person stories? No, they won’t.”

So, those pesky bots are stealing our jobs. Not exactly, says Colton. “The problem is not the technology. The problem is that businesses always want to rake in money.”

Don’t worry, though, we’re not that kind of business. There’s no possibility of Delayed Gratification ever being written by robots in [LANG GEN ERROR 4.2] written by robots in [LANG GEN ERROR 4.2] written by robots in [LANG GEN ERROR 4.2] written by [FATAL SYSTEM ERROR].

Posted on Wednesday, May 6th, 2015 - Categories: From the archive, Long reads.