Skip to main content
May 16, 2019 - Technology

AI learns to write headlines (but not this one)

Kaveh Waddell
Illustration of a typewriter typing on its own

Illustration: Rebecca Zisser/Axios

One of the most frustrating parts of journalism is writing headlines — they need to be pithy and smart, drawing in readers but not infuriating them with cheap clickbait.

What's new: Perhaps the simplest solution is to summarize an article as efficiently as possible. And because machines are getting increasingly good at that, AI headline writers can now nearly instantly generate titles that outshine even some human-made ones.

Details: Primer, an AI company, built a tool to do this, and spoke first with Axios about it.

  • To learn how actual news editors write headlines, Primer's system read more than 1 million news articles and the headlines they were paired with — but only those where the headline was made up entirely of words found in the story.
  • Once trained, it can read a new article and string together the best possible series of words to turn into a headline, according to Primer.
  • In what Primer director of science John Bohannon calls a "headline Turing Test," evaluators were asked to rate computer-generated headlines against the originals — without knowing which is which. In its final form, Primer tied or beat out humans more than half the time, the company said.

The big picture: Understanding and generating natural language is still one of the hardest problems in AI, but machines are making significant advances.

  • New research has supercharged the method that Primer used for its headline generator, packing it with untold levels of data and computing power. The result: shockingly realistic, but still fact-free, prose.
  • Publishers are experimenting with condensing an entire scientific field into a book. They're not very good yet.
  • In the future, better summarization can help create a connected web of academic research that can make it much easier to understand the state of the art in a field, and encourage scientists to re-test colleagues' results.

I asked Bohannon to headline some of our recent stories. The results:

What's next: Summarizing huge troves of documents can help fish out useful information from a bottomless sea. Eventually, says Bohannon, a machine that has a good idea of what you care about can read through millions of documents and send you summaries of only the most relevant information.

Go deeper