When robots write the articles
The automatic production of content, from the Associated Press to the Washington Post
Up until three years ago, a pool of financial journalists from the Associated Press, one of the world's most important and respected press agencies, would publish around 300 meticulously compiled articles (each between 150-300 words long) to report the quarterly financial statements of large American companies. Currently, the quarterly earnings reports available to AP’s more than 1,400 clients are twelvefold and concern about 4,400 companies. The services’ enhancement is not a consequence of a greater commitment by the editorial staff but of the partnership between AP and Automated Insights, a company specialized in tools to automatically generate content from a database.
WordSmith is a sophisticated set of configurable linguistic templates for producing content that has found applications in various fields, including journalism. While the mechanism can appear to be complex, the production flow is very linear. The key elements are the linguistic templates that are developed by a small group of specialized journalists (in AP’s case, journalists specialized in covering economic and financial news). The editorial staff is asked to create templates that can be adapted to various hypothetical scenarios: a profit or a loss, results that are in line with analysts’ expectations or that come as a surprise either positively or negatively…Coverage ‘in absentia’ however goes into great detail by taking into account percentage differences with respect to the same quarter of the year before or of previous years, recognizing elements of particular importance for each business sector such as main products, units sold, investments, and by including the markets’ reaction following publication of the results and eventual comments.
Once the template is prepared, the article’s composition can occur without the direct intervention of the editorial staff. WordSmith ‘writes’ its the final text on its own and does not require any type of review before publishing. The data flow AP provides Automated Insights with is more than enough.
AP continues to assign a journalist to cover only the biggest US companies, large lenders like Citibank or technology giants such as Apple or Google. And even in these cases, WordSmith usually contributes a paragraph or two with the quarterly figures.
This does not mean however that the work of the editorial staff in the profit season is reduced, indeed, the opposite. Freed from the duty of dealing with numbers, ranges and percentages, journalists can focus on added value information such as analysis, commentary, comparisons and forecasts.
In 2015, a NPR reporter challenged WordSmith to write a short article on a publicly-listed company’s earnings as soon as it published its financial statements. The program took two minutes to write the article, while the journalists took seven minutes.
In 2017, the executive editor of Quartz (a site owned by Atlantic Media) wrote a tweet in which he provided an example of how Tencent, the Chinese technology giant which had organized a conference at which he was taking part, had published a “not-half-bad” articled written by Dreamwriter. It is estimated that through that program Tencent publishes about 2,500 daily articles concerning finance, technology and sports.
The Washington Post is also at the forefront of this field and has developed its own artificial intelligence tool to automatically generate content. The product, which currently is only used internally, is called Heliograf and is considered to be the brightest star of the entire sector.
As in AP’s case, Heliograf does not rob journalists of their work but opens up new frontiers in the production of content by allowing them to cover hyper-local issues that would have been otherwise ignored, such as extremely small-scale sporting events (like high school football games) or the electoral results from the smallest and most remote parts of the US.