AI in Journalism 2024: AI as a Tool for Efficiency, Research and Improvements

In this next part of my series on AI in journalism 2024 – where I previously wrote about my own journey into the field – I’ll now explore some of the practical ways media organizations are using AI today. I will not pretend to have a full picture – of course I can’t. Instead, this is a way for me to structure some of my observations about what’s been happening in newsrooms in 2024. To aid me in this, I’ve used Notebook LM, Google’s AI tool for chatting with documents.

As newsrooms struggle to meet the demands of a changing audience, they need to work smarter, redefine their business models and meet the needs of this audience in new and more relevant ways.

Many newsrooms are trying to do more with less, applying AI to streamline production.
While AI can be a valuable tool for efficiency, I believe it’s important to build a broader strategy – one that considers every step of the journalistic process with the audience’s need at heart. Ideally, freeing up time by letting AI handle routine tasks could open the door to more and deeper investigative journalism.

Automating repetitive tasks

Many of these tools have been around for a while. For instance, the book Automating The News (2019) predates much of the current hype around large language models (LLMs) and generative AI. Yet with the rise of LLMs, tools for transcription, translation, and data analysis have greatly improved, leading to widespread adoption in newsrooms. In a survey conducted by JournalismAI in 2023, 85 percent of the newsrooms had experimented with generative AI. Today I’d say were most likely getting close to 100.

1. Transcription tools. This might be the most widespread use of AI in newsrooms today, and that is because it’s such a no-brainer. Instead of spending hours transcribing your interviews, you get an AI-powered tool to do it for you, and double-check the parts you want to use.
At Dagens Arbete, we used both the built-in transcription feature in Microsoft Word and Good Tape, but there are many solutions out there.
YouTube also provides automatic transcripts, which can be useful when locating specific quotes in a published video. There are tools to extract the whole transcripts that can be useful.

2. Translation. AI-driven translation lets you interview sources who speak languages you don’t know, through realtime voice-to-voice tools. It also offers quick document translation, helping you grasp key points in foreign-language materials.

3. Suggestions for headlines. Stuck on a headline? An LLM can generate multiple suggestions, serving as the tireless colleague who keeps brainstorming even after ten tries. You might not use the suggestions verbatim, but they can spark new ideas, and I’ve had at least part of suggestions make it into the final headline.

4. Create content for social media. Generating posts for social media can be time-consuming. Tools that generate short, platform-optimized texts – complete with hashtags, smileys or images – can be a welcome addition. When I tried to integrate something like this into WordPress in March, results were unimpressive, but I believe the technology has improved considerably since then.

5. Search engine headlines and texts. SEO is changing rapidly now that search engines are turning more and more towards presenting AI-generated summaries instead of just linking to sources, but it will still be relevant for some time to make SEO friendly headlines and texts. There are plenty of AI-driven tools that can assist you.

6. Summarize articles. Aftonbladet introduced three-bulletpoint summaries and saw an increase in readership for articles that included them. A small change, perhaps, but it shows how succinct summaries can boost engagement. I built a similar tool into my wordpress toolbox, getting pretty good results.

7. Getting quality comments. At the Graham Media Group in Detroit, the strategy has been to have a reporter pose a question in the comments section of their articles. This has led to higher engagement and better quality discussions. Now they have an AI tool to generate those questions.
It is not only about engagement for its own sake, as a social good or the possibility for getting tips for follow-ups, but also about getting to know the audience better. In order to comment you have to log in.
“We … know that those users who do comment are actually our highly valuable users,” Michael Newman, Graham Media Group’s director of transformation, said to NiemanLab in 2023. An active comments section means more users will register and return to the site. That in turn will make them more valuable to advertisers.

A similar project is in place at German public broadcaster Bayerischer Rundfunk (BR) where an AI-powered tool checks for user comments ”that directly address questions or nudges the editorial team” and sends them to editors. These are used in a new editorial format called Dein Argument, writes the The EBU News Report 2024.

AI for research and finding news

The EBU News Report 2024 predicts that journalism will be mainly about research in the future, less about production. ”The human factor in production will change and partially disappear.” Yet AI tools can also be used for research.

1. Exploring large documents. LLMs can give you a head start in finding what’s interesting in a newly released report or press release, guiding you to the passages that you might want to dig deeper into. They can also summarize key findings. These are helpful tools when you are in a hurry, but you still have to do the actual reporting and double-check what you get from the tools.

2. Exploring many documents. Tools like Google’s Notebook LM lets you chat with multiple documents at once. This can for instance be useful for small newsrooms needing to sift through transcripts of local government meetings. Reporters at Michigan Radio WUOM-FM do this with City Council and Subcommittee meetings in over 40 cities. In a draft to the Partnership on AI’s AI Procurement and Use Guidebook for Newsrooms, reporter Dustin Dwyer says: “It allows a reporter like me to get a lot more coverage to the audience than I could just doing it by myself”. The transcripts are used as a starting point and are reviewed for accuracy.
Dare we hope for more coverage of local politics with the help of AI?

3. Research agents. Agentic AI is on the rise, and research agents can help you gather information on a topic. Already before the latest reasoning models were released, AI agents let you review the research they made and thus form an opinion whether it is valid. With reasoning models, they will become even more powerful.
I have been trying out CrewAI‘s tools, and will continue to do so. My first takeaway is that it matters a lot which model you choose. I got what looked like convincing results, but when I clicked the links to the sources that I had asked for, they were broken. I’ll try it out with a different model next.

4. Track social media. AI can be used to find breaking news in social media feeds. For instance Reuters have a News Tracer that does this.

5. AI tools for facial recognition. In a session on how to integrate AI into creative storytelling at the JournalismAI festival, visual investigative journalist Peter Andringa of the Financial Times talked about how they use AI for visual investigation and storytelling.

– When we think about what AI means for our storytelling, oftentimes the components of these stories are databases, satellite data, social media resources, public records. All of these are big, messy data sources, Andringa said.

He examplified this with an investigation into missing children in Ukraine, who were abducted in contested territories in eastern Ukraine and put up for adoption in Russia. They used a facial recognition model to compare photos of children in a missing children database from Ukraine with those of children up for adoption in Russia.

– Given the size and the messiness of the data, we were concerned that traditional either hand- or computational methods of joining these two databases wasn’t going to be enough.

– Comparing these side by side would be 182 million combinations, far more than anyone could possibly look at on their own. So this is where we thought AI comes in. This is a needle in a haystack scenario. We need to prioritize which of these matches are most important for us to dig further into.

The AI model returned about 1000 possible matches. After human oversight, which could rule out for instance matches where the metadata of the two photos didn’t add up, they ended up with 46 stong cases, that were then investigated by reporters on the ground, finding family members and checking with autorities. In the end 26 cases were confirmed.

– Some of these matches were cases where the names had been changed, the birth dates had been changed.

Without an algorithmic approach, we wouldn’t have been able to join these listings from these databases to identify these matches. But it’s important to see that AI is one step of this multi-step process. It would not have worked without that initial step, but the AI was not nearly sufficient to reach the conclusion that we did at the end of the story, Peter Andringa said at the JournalismAI Festival session, which can be viewed below.

 

AI for article improvements and accessibility

LLMs can provide you with proofreading and feedback during the writing process. Some media companies train models on their own content, to get suggestions for edits that align with their style of writing. Another such example is Finish Helsingin Sanomat, who use a tool called Hennibot. ”Hennibot is an automated editorial assistant used daily by journalists at Sanoma. It helps us enrich our content and understand the audience”, said Managing editor Esa Mäkinen.

Radio for deaf people? At Sveriges Radio, the Swedish Public Broadcast Radio station, they are trying out an AI tool for texting their radio news.

AI for writing articles

The 2024 State of the Media Report shows that while 53 percent of the journalists who responded haven’t used generative AI at all, almost one in five of those who have use generative AI to create outlines or early drafts of articles.

Simpler articles based on data, a type of robotic journalism, has been in production for years. With the addition of AI, these can become more varied.
In areas like sport results, weather, fuel prices and traffic conditions, AI-driven stories are regularly published. At News Corp Australia a team of four staff use AI technology to generate thousands of local stories each week.
Washington Post’s Heliograph is another early example, and in Sweden Newsworthy is doing a great job with combining data and localization for similar data-driven stories.

As pointed out in the EBU News Report 2024, ”A significant benefit for any type of knowledge work could be that by using generative AI, people can play to their strengths and fill the gaps on what they are not as good at or not directly skilled to do.”

Some publications are taking their AI use one step further. At Express.de, they publish AI-generated articles under the name Klara Indernach. ”She” is presented with an AI-generated byline image, and a statement about the material being AI-generated and then checked and edited by humans.

***

Those were some use-cases that I’ve come across. Please let me know in the comments if there are better examples.

Next up I will look at AI for working with data and new formats.