This is interesting in light of all the worry about generative language models spreading misinformation (e.g. GPT-2 and company). With mainstream news trying to automate reporting on real facts, tools to detect auto-generated reporting won't be sufficient to combat automatic fake news. We'd have to have a way of telling the good bots from the bad, which is a whole can of worms.
The headline is kind of misleading. It makes it sound like a machine would be churning out these news, out of thin air, and that's not really the case. The article itself isn't helping much with all the hype.
The article isn't really about bots writing news. It's also far from OpenAI's GPT-2. It's about regular reporters writing massively localized stories with some fine tools. It's pretty cool stuff actually.
What radar is doing is definitely shaping the media landscape, but it is not news created by bots. It's news generated from templates, which are written by humans:
> One of six human journalists working at Radar will write a story “template” with wording for each of the various possible scenarios — for example, a boom, modest rise or sharp fall in violent crime. Then, at the click of a mouse, versions of the story are created for each of the UK’s 391 local authority areas, pulling in the statistics specific to that region.
They are basically doing what Reuters does, except Radar only provides a handful of stories a week instead of hundreds of stories. But the trick is that while stories by Reuters are static, the articles provided by Radar are localized for each region, based on the data used. (All the localizations are written into the template, from which the program picks the best fitting pieces based on the data per region.)
> The UK’s most prolific reporter churns out thousands of stories a month for hundreds of publications across the country — a superhuman effort, were it not for the fact the journalist is not fully human.
The numbers in the previous quote mean that Radar generates hundreds of different, localized versions from these dozen or so weekly article templates.
The cool thing here isn't that a model would create these articles, but that it's possible to create good quality local reporting, from data and a well written master template, by non-local reporters.
Basically, as news publishers are sunsetting their local papers, Radar has shown that it is still possible to serve local news to smaller regions.
And all this can be done by more or less regular reporters. There's some programming involved, but nothing complicated - AFAIK, the programming for the templates is similar to writing scripts in Excel.
> Gary Rogers, Radar’s editor-in-chief, estimates that half of its stories that appear in print are added to by local reporters. However, that drops to just 20 per cent for stories republished by online news sites.
Like mentioned in the quote, half of the time, for stories that are going to be published in physical papers, local journalists add their own reporting to the stories. Again, these stories don't magically enter the front page. They are selected to be placed there by the editors, and in many cases, local journalists add their knowledge to the articles. But still, the language in them is good enough to be used on prime estate.
In web and mobile, it appears, the rim is lower, and it goes more often faster to the site, without modifications. But then again, the rules in internet are different.
Plenty of the problems that people have had with previous template-based robot journalist writings, has been with repetitive textual structures. With Radar's model that is not as big a problem, as the master template is only used once per story. If a reader reads the story from one magazine, it's less likely that she will also read the same story in another paper. Compare this to sport robots where you see the same structure repeating in every story.
I've understood that Radar isn't the first in this place, but they have shown a business model for template based news that manages to avoid the biggest problems with repetitive structures.
In summary, Radar is doing something cool by automating local news, although the article isn't the most clearest on what the cool thing is.
There's plenty of ambiguity in stories about Radar, so my guess is this is partly due to their press release lingo / hype.
There are certain open questions with Radar style localized template based news.
Mainly it's about automatization, changing the way things are done. New skillsets are needed, while old ones are out of the door.
There's the existential problem this kind of centralized local news creation has on smaller desks, mentioned in the article, but then again the trend appears to be that owners are running the small, unprofitable ones down anyways, regardless of Radar.
There's also the question whether this style of localization, but not personalization, of news, is the best way to serve readers. Could these same local news be served in some other way?
IMO, the way Radar works, mainly happens the way it does, as there exists a need from small papers, for quality journalism that they don't have their own resources to conduct. Of course, nothing is stopping Radar of doing something different in the future.
Then there's also the hypotetical (?) threat on press diversity. Basically a multitude of superficially distinct voices are repeating the same liturgy. This zerg of articles on a specific topic, might appear to a news consumer, as many credible sources supporting a specific view or point.
There's similarity with Reuters et al in this problem, and clearly Reuters hasn't destroyed news. But with Reuters, the text is most of the time identical in all the places where it is published. In the case of localized templates, it might not be clear to the reader that all the different versions are instances of the same master template, with same bias from a single writer, if the template is well enough written.
