Forbes has been publishing hundreds of articles a day for more than five years. As a Data Scientist, when I hear about that volume of data, the first thing that jumps to mind is: what can we learn from it? Are there best practices we can glean? Can our volume of work make future content better? The answer is yes.
We started by looking at what colleagues are doing with their own systems. We really enjoyed reading about BuzzFeed’s in-house headline A/B testing system, about the New York Times’ clever image cropping application, and also about the Washington Posts’ feature-rich CMS-for-hire Arc Publishing, among others.
What we noticed is that each publishing platform generally focused on two things:
- Improving the workflow of writing a story (from author → editor → publication)
- Providing authors, editors, and producers with tools to create better stories
“That’s fantastic”, we thought. These areas need work and as far as we can tell our colleagues are making great progress.
The question for us is: how far can we realistically go and create something that adds value to our writers and business?
In his upcoming book Superminds, MIT professor Thomas Malone suggests a way forward. He writes that there is an evolutionary process between tools and “collaboration machines”: first they evolve into assistants and from there into peers.
Prof. Malone writes that the key difference between tools and assistants is essentially two things: initiative and ability to collaborate.
Creating a machine that has initiative isn’t too difficult: we can create a computer program that constantly searches for patterns and sends notifications to users when it identifies something that is worth their attention. We have to balance that with great user experience so we don’t annoy users, but that is something we can tackle. What about inputting the ability to collaborate?
Learning to collaborate
We think that writers will only collaborate with a publishing AI if they both find it enjoyable and useful. So we came up with three principles to achieve that goal:
- Learn from authors at all times: we are balancing both the strengths of all authors and the specifics of each author.
- Don’t interrupt: our AI system is being designed to learn how authors interact with it, optimizing for the sweet-spot between usefulness and subtleness.
- Be personable: all the suggestions that our assistant generates “sound” and “feel” as if they are coming from a single entity – maybe even give it a name.