New York Times: To succeed with data in the newsroom, put people first

It is 2014 and the New York Times realized that digital innovation had to be prioritized to keep up with its competitors – one response was to set up a data science team. A group of specialists that would become of significant importance to​ ​the entire organization.  

The data science team is supporting both newsroom and business, and according to Chris Wiggins, Chief Data Scientist at the New York Times, it has been quite the cultural transformation. Many journalists struggle with churning numbers and analyzing data. And it is a lot to ask a busy news writer to access and understand an app that quite literally speaks in code.  

To get them onboard, Chris Wiggins and his team had to create both meaning and context for their colleagues in other departments. The team needed to figure out how to think about the interoperability of their work. Making sure it was both tangible and valuable for their peers to look into what could seem like a black box. A mindset that together with support from the CEO, led his team of two to grow to a team of twenty-two.  

NYT works with a people-first approach. This was a consequence of realizing that the structure of their teams had a high impact on their ability to deliver results.  

“How do we organize the teams? We have learned how important it is to structure our teams well in terms of their communication structures. We produce the very design of the organization within these structures. So, we worked very hard on this, making sure that our data scientists and software engineers are sharing desks, sitting side by side, sharing repositories, and sharing snacks. We want them to talk together and learn from each other,” Chris Wiggins said at CUE Days 2022.  

Besides mixing teams of software engineers and data scientists, the Data Science team also works in close collaboration with their friends in Product, Marketing, and the newsroom. Having the newsroom co-creating the recommendation algorithm has been a big success for the team, showing exactly how well input from other departments can make a difference.

The New York Times’ data department works with data in three different ways, built on the idea of Gartner’s Analytics Ascendency Model. By doing so they make sure everyone in the company can follow the need for analysis, meanwhile, the employees get the tools and support necessary to incorporate data into their everyday workflow. No matter if they are working in advertising or as an investigative journalist.  

It’s quite simple. Depending on the problem you either work with a “what happened”, “what will happen” or “how can we make it happen”-approach.  

Also called descriptive, predictive, or prescriptive modeling. Or in machine learning communities: unsupervised, supervised and reinforcement learning. 

And guess what – it aligns quite well with the journalistic way of thinking.  

When you work with descriptive modeling, you want to make sense of a large and complex dataset. In short: We want to know what happened by creating an overview everyone can understand. 

“A good example of this,” Chris Wiggins says, “is our ​R​eaderscope. Years ago, during our global expansion, we realized we wanted a more sophisticated, real-time insight into what is happening across our website. Who is reading what and where?” 

And while it started as a curiosity about the readers’ behavior on their website, the insight they gained led them to be able to range stories based on behavior and thereby engage with the reader for a longer time online. This became a great tool for advertising, making the New York Times able to monetize the insights created, while also creating an overview of best-performing stories for the journalists to evaluate on.

At the New York Times, working with data does not necessarily have to be about ads, subscriptions, or monetization.  

A great example of predictive modeling is the Takata project. Together with a journalist, Chris Wiggins and his team created an AI setup to figure out what incident reports involved a certain type of airbag. The AI setup suggested which incident reports should be further investigated, saving the journalist a lot of time. It’s a great example of how to empower journalists to find stories lurking in the thousands of databases maintained by governments, etc. The project resulted in a big recall from the airbag company, as a fault in the airbag was discovered.  

Project Feels is another example of a predictive project at The New York Times. The AI setup predicts what emotion a story might invoke in ​readers​. It turned into what the advertising department calls ​Perspective Targeting​​. ​Perspective Targeting​ at The New York Times now makes it possible for ​marketers​ to advertise their products next to stories that may ​set the right tone​​ ​with the​ir​ ads.  

But no matter if the predictive modeling analysis is for investigation or monetization, the outputs are always non-actionable. They simply identify the need to make a decision and they are almost always based on predetermined scenarios with finite options.  

Prescriptive modeling analyses the problem and asks, how we can make something happen. A good example could be: Where on social media do we post our content to gain the most engagement? 

The Audience Development editors asked the question and as a result, the data science team built a Python-based application. The challenge was, it would have been a disruption of the editors’ process to open Python to get to the insights.  

“We realized that we could do the machine learning and we can do the prediction on how much engagement a story would get as a function of what social media is chosen. But we cannot get our editors to open up Python or a new web app. We needed to meet the editors where they work," said Chris Wiggins to the audience at CUE Days.  

The solution was to build something understandable into Slack, as it had become quite important for the editors in their workflows.  

“So, we decided to meet them there by building a mirrored text interface in a project we called Blossom. Making it easy for the editor to figure out if the story would be better suited posted on a subaccount or promoted on for example Facebook”  

This a great example of how to meet editors in their already established environment.