We have emphasized that nothing replaces original reporting. And when we discuss AI, we focus on responsible use of AI tools in a newsroom to help analyze big datasets, speed up research efforts, and automate more routine tasks—with a core principle being that we ensure a human is always involved and nothing gets published without human oversight.

Other core principles are transparency (clearly label when AI is being used in a significant way), accountability (every journalist is responsible for every word they publish, and plagiarism is a fireable offense), and authenticity (don’t use AI to write or edit a story from scratch).

At the time of our first policy draft, in 2023, very few AI guidelines had been published. Wired was one of the first, and we were motivated by their work. Research by Hannes Cools and Nicholos Diakopoulos was also incredibly helpful, as their early roundup allowed us to see the different “flavors” of policies emerging across the industry. Over the past three years, we have iterated on the policies by maintaining a running list and analyzing what works for each publisher, what works for us, and what we might want to consider for our next revision.

We worked hard to get the right balance between encouraging enthusiasm for experimentation and maintaining journalistic integrity and protecting our intellectual property. It was very important for us to lead with excitement rather than fear, which is why our guidelines open with the line “Advances in artificial intelligence create incredible opportunities for our newsroom.” And only when you get five grafs down do you see the first “don’t do this” statement.

Before the policy, before the tools, organizations need to be able to clearly answer: Who are we serving, and what do they actually need from us now, and in the future? Policy is only as good as the thing it’s serving, and the starting point comes from answering those questions clearly.

We want our policies to be editorially driven and our tools also to be editorially driven. Our policy right now is that we do not create any text using generative AI but may use it to help create some data visualizations. We use suggesting capacities, like suggesting a stronger headline or alternative text for images, but that also should be edited by the reporters. But if the policy does change, we will be transparent about what is being generated, if anything.

My hope is that AI can help us scale into areas where there are already news deserts. There aren’t enough people covering city council and school board meetings already, including us. I would love it if AI could help us get transcriptions and summarize them so reporters can see if there’s anything important we should cover. Then the reporter can follow through, call people, and do their investigative reporting based on what they find from these transcript summaries. To me, this is where AI would be golden.

In developing a policy: don’t overthink, and don’t be too concrete. We like working with principles instead of rules. Rules, model recommendations, and dos and don’ts all quickly become outdated. Principles should be able to last longer. We did have to revisit ours, and probably will again soon.

Zetland should feel like a very human product. We’ve increasingly been doubling down on this vision. Figuring out how to use the robot and become more human is the big challenge.

My manager, chief growth officer Michael Tortorello, and I were the “draft committee” in charge of writing the first draft of the AI guidelines. We scheduled one-on-one meetings with department heads and managers to talk about specific areas in this policy statement that would affect their team’s work. We talked through the types of tools they were using and the use cases they imagined for experimenting with AI in the future. We took all of their feedback and concerns—and hopes and dreams—into consideration.

We also had an all-staff meeting with breakout sessions to talk about the policy and questions they had about it. We showed this draft to all of our staff, so they had time to read it and raise questions. I then hosted open office hours with our staff for them to bring concerns. That all went into the final revision work before we published.

We require everybody to take a refreshed AI 101 training every year so they are up to date with our policies, changes in the AI world and the tools at their disposal, how some of their peers are using AI in their journalism. We also have monthly town halls, regular demos, and emails to keep people engaged, and provide space for sharing and debate.

We built our transcription service, Good Tape, in 2022. We recruited and trained AI experts fairly early to work alongside our newsroom, our tech team, and designers. The service created a positive buzz around AI in-house very early. (We sold Good Tape last year, but Zetland is still using it.)

We formed a group of internal champions, and they shared their knowledge and also arranged informal sessions for the company. The most important thing has been to share best practices. The policy is not at all important. I doubt anyone ever looks at it. We’re not a company with a huge book of standard operating procedures. But the editorial ethics and culture are quite ingrained. What we do is to spend time talking about data privacy and what the ingrained rules of journalism mean in this new reality.

We started with training. We trained all our journalists in prompting, and then it was ChatGPT—the only thing that was around back then. Then we moved on to editorial tools, transcription tools, summarizations, SEO stuff, text-to-speech.

While it’s important to have a policy document, we’ve found it even more useful to discuss those guidelines in person and explain things like why using enterprise-grade tools is critical to protect our journalism and IP. “AI 101” training sessions have been a huge help in explaining the “why” behind our rules, rather than just handing out a list of requirements.

For a busy newsroom, we found that show-and-tell sessions are even more effective. We recently hosted a series of three “lunch and learns” where colleagues demonstrated how they use AI for reporting, fact-finding, editing, or “vibe coding” new tools. This summer, we’re diving into AI image detection and a hands-on vibe-coding workshop.

Building our AI guidelines was a top priority for our newsroom AI task force immediately after we formed the group. The task force includes both AI enthusiasts and skeptics from across our photo, audio, graphics, audience, data journalism, newswires, coverage, and standards teams. Our AI guidelines specify what we’re comfortable with now, but that’s a moving target. We made sure our guidelines are treated as a living document that will continue to evolve. Each version has a “last updated” date, and we aim to review the policies every six months or after a big new technological shift. Up next: drafting our vibe-coding guidelines. We’ve been talking a lot about how we can safely and ethically use AI to build functional, specialized apps for each team or individual.

Our AI Advisory Board is a cross-functional group of leaders from our editorial, research, and product divisions, such as breaking news, television, data journalism, Bloomberg Economics, and Bloomberg Intelligence. Our role is to be advocates and champions for technology while making sure we’re adhering to our editorial guidelines. Members are located in the US, UK, Singapore, and Japan.

There was a lot of interest in forming our committee, which has representation from the business team, editors, and reporters, particularly after an incident at the Wisconsin State Journal in which a journalist was fired after one of her stories included an AI-hallucinated business owner, business, and quote about a local development. That kind of error would be caught at Wisconsin Watch, but reporters were still concerned about what our policy is. We’re now in the process of developing our policy. We’ve also been conscientious about equal committee representation from both the union and management.

The ABC is trialing more AI tools that can assist our journalism—always with editorial oversight. Over time we envision this work will power up our investigative journalism, help with routine production tasks such as transcribing and metadata, and deliver our content to the audience in new and impactful ways. This will allow our teams to focus on the core work of doing great journalism.

We had a documentary series examining the crimes of Rolf Harris, an Australian entertainer who became a convicted pedophile. He is now deceased. The producers were exploring the idea of using AI to re-create his voice to read some of his letters. The internal conversation and ethical framework that the ABC applied when considering the proposal were: What is our level of comfort? How do we do this ethically? How do we tell the audience this? Is there any way that the audience could be misled or confused by what they’re seeing, and how would that affect trust?

In the end the producers decided against using AI in this way. They found a more creative and authentic storytelling approach using real recordings of Harris. They took a very careful approach to consider how they could best tell this story with impact and care.

The AI world moves incredibly fast, so all of us publishers need to keep reviewing our policies. For Reuters, that means having a senior editor responsible for reviewing our guidance regularly and an AI Governance Committee, chaired by our editor in chief, which meets monthly to look at the tools and solutions people are using and building, testing them against our standards, feeding back before approval, and always maintaining the ability to hit the kill switch if something goes awry.

Our standards editor Carlos Cunha created Axiomizer, with the support of the standards and copy editor teams, to help reporters deliver cleaner copy to their editors. It started out as an OpenAI GPT. It offers suggestions to sharpen sentences, tidy up grammar, and ensure optimal usage of axioms.

What it can’t do—what we want it to do—is fact-checking. We keep running tests through it, and it won’t find everything, or sometimes it will find something that is correct and flag it as incorrect. We’re getting much closer to a good fact-checking tool.

The field is moving faster than any single organization can keep up with alone, and sharing, which journalism has historically been bad at, is now a competitive necessity. The organizations that try to figure this out in isolation are leaving hard-won learning on the table. That’s why NPA invests so heavily in creating spaces for practitioners to think out loud together—our programs, our Slack, the NPA Summit. The peer exchange that happens when people doing this work in very different contexts compare notes is irreplaceable.

You can’t just set some policies and forget. You need to be in line with audience expectations, making sure that they actually work in practice. We can set these things up and say, “Right, this is what it is.” But then maybe we need to revise and refine that a little bit more. It’s foolish to think that you set forward policies and then not review them. We have to be in constant review.

One of the really interesting changes we are seeing is that some of our journalists are becoming smart product builders and some of our engineers are spotting journalism problems to solve.

Our chatbot has responded to seven- or eight million questions. And we haven’t received a single email about that. So I guess people are just using it. Maybe they don’t love it. They don’t hate it. They just use it because it works.

Most people are not super interested in exactly what tools you’re using for various tasks. I think the only thing they want to know is: Is it a human or a journalist who makes the bigger decisions? Who chooses what you are reporting, for example, and who’s responsible for it? People are concerned about that.

As a nonprofit newsroom in Minnesota covering immigrants and communities of color in the state, we have a very dedicated demographic that we serve. There are a lot of disparities in AI output when it comes to communities of color, and that concerns us. The context really matters. At Sahan, we’re pretty transparent about our work. In the introduction to our “Sahan Sunday” weekly roundup newsletter, I wrote about using some AI tools in both our reporting and business side and about why this was a necessary step to take. I wrote about challenges and concerns that I had in developing relevant projects and the decisions that I made before publishing projects that involve AI.

We always ask, “What do you think? Give us your feedback.” We have pretty positive reviews: “I’m glad that you used this. This helped me save a lot of time and made your news more accessible.” But in cases where we just mentioned AI without explaining how and why we’re doing it—what exactly we are talking about when we use AI in that specific context—that’s when readers tend to be more reactive or negative about it. And that’s why our AI policy is so important.