How Insider and Business Insider will use artificial intelligence

Insider executive editor Julie Gerstein sent out the following to the staff on Friday:

I am sending an update on the AI pilot group’s progress and offering some newsroom-wide guidance as we explore the possibilities of working with AI tools.

Over the next six weeks, the pilot group will test a series of prompts and hacks. Each of our four mini-groups will tackle one of the following:

1. Summary Bullet Prompts
WHAT: Develop standardized prompts for generating summary bullets.If our experimentation is successful, these prompts will be integrated into Muse and used to generate suggested summary bullets for stories for the whole newsroom

2. Translation services
WHAT: Experiment with quick and accurate AI translation services for 1) translated video captions 2) Interview/source translations 3) online webpage translations

3. Editorial Prompt Creation
WHAT: Determine the best prompts for the following use cases, including research, story editing and improvement, and pitch generation.
  
4. Productivity Hacks
WHAT: Determine the best prompts for taking and synthesizing meeting notes, writing emails, and synthesizing feedback.

Our aim is to come up with a prompt library that we’ll be able to share with the entire newsroom, so stay tuned.

We welcome everyone in the newsroom to experiment with AI whether you’re in the pilot group or not. If you’d like to get access to our paid account for GPT-4, which is the more advanced version of the free ChatGPT, please reach out to Robin Ngai for access. 

If you’re experimenting on your own, please be aware that AI has some pitfalls.

It shouldn’t be used for:

• Directly quoting facts from it or using it to fact-check other sources

• It’s much better at creating effective language around information you have already sourced and know to be true.

• Facts relayed by AI can be effective leads but should be independently confirmed

• Providing up-to-date information. Most of ChatGPT’s training data runs off models that were last updated several years ago.

• Data analysis

• It starts to produce inaccurate results even on small data sets.

• It will often mimic data analysis but provide inaccurate results.

• In one case, it told a pilot group member it was “roleplaying” at providing data analysis but was actually unable to “directly analyze a real-world dataset.”

The pilot group will continue to provide updates on its progress. 

 

Please don’t hesitate to reach out with any questions. Thanks!