Discussing AI at a Journalism Tech Conference

I recently attended SRCCON, a lovely conference held in Minneapolis that focuses on the tooling needed to help modern news organizations innovate and also, sadly, stay alive. 

AI came up a lot. The notes and transcripts from the sessions are online if you’re interested in reading (or auto-summarizing) them. My takeaway is that there’s a lot of trepidation and curiosity around AI in media, but so far, not a lot of action. 

A few of the concrete examples of AI impacting newsrooms I came across: 

• Many publications are writing AI policies to add transparency for their readers, and offer guidance to their staff. There's a fear that AI will erode trust with readers, even if the organization isn’t using it to write articles. 
• A Reynolds Journalism Institute fellow built a tool that suggested SEO headlines for articles.
• An engineer at The Washington Post built a local semantic search tool to help journalists quickly search through a large number of documents. 
• One organization has experimented with using language model embeddings to analyze reader engagement with different lines of reporting. They previously had to rely on hand-added tags to do this kind of analysis.

Nobody that I spoke to was using AI to generate articles, and there was broad awareness of the shortcomings of generative AI (bias, hallucinations, copyright issues, security risks, etc). Even with those concerns, most attendees seemed eager to see if AI could save them time on the “production” work necessary to get an article in front of readers. 

Wouldn’t it be nice to just write an article and not have to tag it, write social copy, add captions and generate SEO headlines? These tasks are ripe for automation, and current AI technology can almost solve the problem. Problem is, most newsrooms don’t have engineers on call who can wire all the pieces together, set up a testing framework to gain confidence in the system, and then integrate it into a CMS. 

Several years ago, I worked on a project to speed up the production of recipes for The New York Times by “parsing” the recipe text from their print CMS using a machine learning model. I had to train my own text prediction model, (something called a CRF that you will never need to know about,) because OpenAI didn’t exist at the time. 

Back then, it took me one week to train the model and three months of part-time work to integrate the model into our CMS. Recently, I decided to reproduce that system with ChatGPT, and got it working in 10 minutes with a three sentence prompt (copied below).

LLMs and OpenAI’s API have made it possible for anyone to build a recipe parsing system in minutes, but you’ll still need some engineering help to integrate them into complex, legacy CMS systems. I hope organizations put resources behind the latter work because there are huge efficiency and quality-of-life improvements to be had. But I understand why these aren’t point-and-click solutions yet. 

—--

Prompt

Identify the original input, ingredient name, ingredient quantity, ingredient unit and any additional text from the following strings. Give me the output in JSON. The JSON keys should be "input", "quantity", "unit", "name" and "comment". Include all the fields even when they're empty.

"1/2 to 3/4 cup sugar"

"2 tablespoons unflavored gelatin, dissolved in 1/2 cup water"

"1 garlic clove, minced (optional)"

Response

{

"input": "1/2 to 3/4 cup sugar",

"quantity": "1/2 to 3/4",

"unit": "cup",

"name": "sugar",

"comment": ""

}, {

"input": "2 tablespoons unflavored gelatin, dissolved in 1/2 cup water",

 "quantity": "2",

 "unit": "tablespoons",

 "name": "unflavored gelatin",

 "comment": "dissolved in 1/2 cup water"

},{

"input": "1 garlic clove, minced (optional)",

"quantity": "1",

"unit": "clove",

"name": "garlic",

"comment": "minced (optional)"

}