AI & News: Moving Beyond “if” to “how”
Today, understanding artificial intelligence is an essential skill for a media leader. This is not simply because you need to determine whether to allow scraping of your web site, whether to sue for copyright, or if you should do a deal with a player like Open AI. It’s also because you need to figure out which aspects of AI you will use in the service of impactful journalism and audience engagement. AI will reshape the media landscape, and the organizations who use it creatively and impactfully will thrive.
Reporting tools-
A news organization is only as good as its stories, and AI offers opportunities for new reporting discoveries and alert systems. AI isn’t appropriate for reporting — especially in the physical world where new knowledge and information is found that is not already on the Internet, where AI lives and operates. But here are some great examples of AI tools that help human journalists write deeper, faster, more interesting stories.
First, to discuss an area near and dear to my heart — investigations. AI and machine learning models can be used to find things on the Internet that would be unspottable by even the most diligent reporter. We found this to be true at The Journal in investigations we did about algorithms, like this one, which examined how TikTok’s feed works, or this one on Google. In addition, reporters are increasingly getting large data files. Using AI tools to query them is a no brainer. In every case I’ve seen, there’s plenty of human-only reporting to do on top of this sort of AI investigative digging. AI is just part of the puzzle solving.
A different type of AI reporting tool is a story alert system. This can apply to all sorts of topics but at the Wall Street Journal, we used it for business reporting. We built a stock movement alert tool. When our models noticed stocks of companies moving in ways that typically indicate news, our system pinged the relevant beat reporter in Slack so he or she could hit the phones and see what’s going on. It’s a great way to break news.
There will be many other AI-assisted reporting tools that get created. News organizations should have in-house experts who understand the level of care good journalists put into reporting and who understand AI enough to ideate, build and evaluate new tools.
Content creation–
This is perhaps the trickiest area, and it’s the area most people have focused on in recent months. Flubbed attempts by CNET and others have had journalists shaking their heads at the idea that AI can create content. People tend to react to new technologies in general by thinking of the most obvious use cases, and especially ones that cut costs, but those obvious uses are often the ones where new tech falls down at first. In general, more success may be found by thinking more creatively about how to use new technologies, like AI.
Content creation is one of the areas where creative uses of AI can have the most impact on improving what we provide to audiences. In my consulting work, I advise news companies to focus on using AI to create new content that we would not have otherwise had. Rather than focusing on how to get AI to do all the things human reporters do, why not make new types of content that will engage readers, but that are beyond what your human teams can create?
Like personalized content. No news organization can have its best reporters manually create content that is tailored to every reader. Yet there are occasions that readers want personalized content. A great example of this working at the WSJ was the member hedcuts we introduced in 2019. The Journal is famous for these hedcuts, which are speckled drawings of people’s faces, and The Journal runs them alongside many news stories. But it’s usually only famous people who have had a WSJ hedcut. We liked the idea that any subscriber could have one, but of course, the WSJ human artists could not create that many hedcuts. The solution? AI. We used a GANs model to create a tool that converted uploaded headshot photos into personalized drawings. It was a big hit with our members. Read more about the AI methods behind the hecuts here. (GANs are now superseded by diffusion models (Midjourney, Stable Diffusion) — I will be doing a more technical post on some of these technologies next week.)
There are many types of written stories that could be created that are individualized to the audience. One example idea is an experiment called the Flexicle we did at The Journal where articles expand out with personalized information. You can read more on that here.
As AI gets more powerful, the opportunities to use it to amplify and extend the work of reporting teams will only grow.
User Tools-
For all the controversy in the industry about whether news companies will do deals with OpenAI and other tech giants, there already has been a deal between a technology company and a news outlet that worked great.
In 2020, The Wall Street Journal partnered with AWS’s machine learning lab to create a user-facing tool that would answer the public’s questions about political candidates. The effort grew out of a tool that our teams had created for the reporters in the DC bureau to fact check politicians using the Factiva news database. Our consumer research showed that the public was highly interested in fact-checking candidates as well, and so we created a tool where people could ask a question, and get an answer. The models behind it were created in partnership with AWS’s machine learning lab. They represented an early deployment of technology that we now see in language models like Claude and ChatGPT. You can read more on the technology behind it here.
Called Talk2020, this tool is an example of something we couldn’t offer readers without AI. The models behind it were quite advanced and the user engagement from the audience was fantastic. It gave our marketing team something unique to advertise during the presidential debates. It led to more engagement with our stories, which we promoted alongside it. You can read more about that here.
Analytics Tools
Newsrooms are increasingly paying attention to audience data to inform their coverage. That’s good in some cases, but the data in some places is blunt and misleading. At The Journal, we used neural networks to create a topic model that allowed us to make more sophisticated recommendations about what the data showed. The idea was that an AI model could help us learn a lot more about what our stories were about, beyond the manual tags editors had put on them. There are many more avenues data scientists can explore in news data especially in the area of audience engagement. You can read more here.
Only the Beginning
It is unlikely that media companies will be the place of deep technological innovation with AI, in part due to cost constraints. (Media companies are unlikely to spend $150 million on GPUs to train large language models from scratch.) But news companies can benefit by hiring people who know how to use these technologies and can think creatively about how they improve news experiences. Next week I will share a technical article on some of the AI open source systems that I think are most important for news product and engineering teams to be examining.
We are still in the early days of what the news industry will do with AI, but let me leave you with one big thought. AI will change user behavior too. The public will begin acting differently online and come to have different expectations online. News organizations need to look to AI not only as a way to improve what they are doing today but also as a way to adapt along with the changing public.
