By now, we’ve all seen the headlines: 

“(Magazine) accused of publishing articles written by AI.” 

 “(Tech site) pauses publishing AI-written stories after disclosure controversy.” 

 “(Newspaper chain) to halt AI experiment after botched articles.”

At Hearst Newspapers (HNP), the last thing we wanted was to be another ugly headline or lose the trust of our readership. However, we knew we couldn’t sit still and ignore the opportunities for local journalism afforded by the rapid advances in generative AI. 

So we started tinkering. Carefully. 

How it started 

HNP’s DevHub — a team of a dozen journalist-developers and content strategists who collaborate with newsrooms to conceive, build, and publish interactive projects, news apps, internal tools, and storytelling templates — hired an AI+automation engineer who had started building a smart and simple audience optimisation tool, dubbed Producer-P, for one of our local newsrooms. 

The Slack-based tool reflects DevHub’s “innovate, carefully” strategy. It streamlines digital content production across several newsrooms, including the San Francisco Chronicle, Houston Chronicle, and San Antonio Express-News.

Producer-P leverages OpenAI’s GPT-4 and fine-tuned GPT-3.5-Turbo models to assist journalists in creating optimised headlines, SEO titles, URLs, related links, and notification summaries. By quickly analysing article content, Producer-P not only cuts down the time needed for these tasks but also brings a variety of perspectives to content creation. 

The system is designed with a focus on active user engagement. Rather than being directly embedded into HNP’s CMS, which could lead to an overreliance by editors on AI-generated content, it requires users to critically assess the content it generates in Slack and manually implement the bot’s suggestions. 

Ethical and responsible AI usage is a cornerstone of Producer-P’s implementation. All users undergo a training program covering technical usage and broader issues relating to large language models. This training focuses on three essential rules:

• Maintaining human oversight.
• Limiting text generation to summarising, aggregating or remixing reported, edited, and vetted information.
• Ensuring transparency with both readers and colleagues.

What we’ve learned

Since its launch, Producer-P has been embraced by HNP journalists, handling more than 1,000 requests a month across our network.

The tool’s influence goes beyond tangible metrics, however. More than 200 journalists have been trained in responsible AI use through the Producer-P rollout, shifting our newsroom culture toward seeing AI as a “copilot” — an enhancer of journalistic skills, not a replacement.

Journalists across the group report a deeper understanding of AI’s role and limitations, enriching their reporting capabilities and opening up our company for future innovations.

While Producer-P is available to all our newsrooms, big and small, our hyperlocal newsrooms see the most benefit. With fewer audience-focused journalists on staff, our community newsrooms are more likely to see Producer-P as akin to having an audience producer or SEO editor on their team, while larger newsrooms already have journalists working in those areas. 

We’ve also learned that an AI tool trained on our content can be accurate, despite our initial concerns. A testament to Producer-P’s reliability is its unblemished record of accuracy, with no incidents of factual errors. 

The tool’s contribution extends to reshaping our approach to AI in local journalism. Following on the success of Producer-P, we are developing more AI-powered internal products focused on translation, newsletter summarisation and transcription, as well as public-facing experiences that repackage our reported, edited, and vetted content. 

Producer-P’s use also underscores the balance between technological advancement and the value of human judgment. By reinforcing this balance, Producer-P boosts efficiency and fortifies our commitment to responsible, innovative journalism.

. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .