How Propublica uses AI responsibly in its reporting – Propublica

In February, my colleague Kane Shwankke saw a post on a social media network Blussky about a database released by Sen. Ted Kruz asked the National Science Foundation to show more than 3,400 “Vok” grants that “promoted diversity, equity, and inclusion (DEI) or Advanced New-Marxist Class Propaga.”

Given that Schwencke is our senior editor for data and news apps, he downloaded the data, pked around and saw some grants, which looked far away from Cruz, a Texas Republican, called “The Radical Left’s fuck nonsense”. The grants included Schwencke thought that the University of Florida, the University of Florida, was a “very quiet project” on the development of advanced mirror coatings for gravitational wave detectors at her Alma Mater.

However, the grant details have mentioned that the project “promotes education and diversity, provides research opportunities for students at various education levels and carries forward the participation of women and reduces minorities.”

Schwencke wondered that it would be interesting to run data through AI large language model – one of those powering chatgits – to understand the type of grants that make a list of cruise, as well as why they were marked. He felt that there was an accountability story to tell.

In that article, Agnell Philip and Lisa Song found that “the dragnet of Cruise had swept of several examples of scientific projects funded by the National Science Foundation, which only accepted social inequalities or was completely unrelated to social or economic subjects cited by their committee.”

Among them: grant of $ 470,000 to study mint plants and how they spread to continents. As the best Philip and Song can explain, the project was marked due to the two specific words used in its application in NSF: “Variety,” referring to the biodiversity of plants, and “women”, where the application mentioned how the project would support a young woman scientist on the research team.

Another one involves developing a device that can treat severe bleeding. It included the word “victim” – as in a gun tablet – and “trauma”.

Neither Cruise’s office nor a spokesperson responded to our requests for the article for the article on the Senate Committee on Commerce, Science and Transport.

The story was a great example of how artificial intelligence could help reporters analyze large versions of data and try to identify patterns.

First, we told the AI ​​model whether to identify a investigative journalist through each of these grants to identify whether they include the topics that are looking for “Vokens”. And significantly, we ensured the model to ensure whether it was not sure. (AI models are known for hallucinations, and we wanted to protect it.)

For AI and readers, for the New News Room, which is eager how it works in practice, here is a fraction of the real signal we used by us:

Of course, members of our employees reviewed and confirmed every detail before publishing their story, and we called all the nominated people and agencies seeking comment, which remains a tax in the world of AI as well.

Philip, one of the journalists who wrote the above query and the story, is excited about the possible new techniques, but also moving with caution, as is our entire news room.

“The technique promises a ton in the lead generation and indicates us in the right direction,” he told me. “But in my experience, it still requires a lot of human supervision and vet. If used correctly, both of them can actually speed up the process of understanding large sets of information, and if you are creative with your signals and read the output severely, it can help highlight the things you would not have thought of. ,

It was just the latest attempt to use AI to help us to use AI to help and help our work better, while its use was also responsible, even in the ways that help our human journalists.

In 2023, in partnership with a local reporting network partner, Salt Lake Tribune, we used AI to help highlight the pattern of sexual misconduct among the disciplined mental health professionals by the licensing agency of Utah. The investigation trusted a large collection of disciplinary reports, including a wide range of potential violations.

To narrow in cases we were interested, we inspired AI to review documents and identify those who were related to sexual misconduct. To help the bot do his work, we gave it examples of confirmed cases of sexual misconduct, which we were already familiar to and specific keywords. Each result was reviewed by two correspondents, who used licensing records to confirm that it was classified correctly.

In addition, during our reporting on 2022 school shooting in Uvalde, Texas, Propublica and Texas Tribune, he received a contingent of unpublished raw materials collected during the state investigation. It consisted of hundreds of hours of audio and video recording, which was difficult to throw through. The footage was not organized or clearly label, and some of it were incredibly graphic and harassed to see journalists.

We used the self-hosted open-source AI software to safely classify the material, which enabled reporters to match the respective files and recreate the day-to-day events, which shows the laborious detail how the lack of law enforcement preparations contributed to the delay in facing the shooter.

We are well aware that AI does not repeat the much time-intensity done by us. Our journalists write at the top of our stories, our newspapers, our headlines and long stories. We also know that there is a lot about AI that needs to be examined, including companies marketing their products, how they train them and risk risk by them.

But for us, AI also has the ability to use as one of several reporting tools that enable us to creatively examine data and pursue stories that help you understand the forces that shape our world.