Examples of how news organisations are employing generative AI are emerging and news summaries offer a useful, non-contentious entry point. Here’s how South Africa’s Daily Maverick went about offering readers short versions of their top stories.
Last week, a new report, Generating Change: A global survey of what news organisations are doing with AI, gave a strong overview of news organisations’ engagement with AI and associated tech.
The comprehensive survey of over 100 news organisations from 46 countries was led by Professor Charlie Beckett and Mira Yaseen from the JournalismAI initiative at the London School of Economics and Political Science (LSE).
The survey found that, despite fear and pessimism, more newsrooms than not the world over were experimenting and using AI than not: about 75% of the respondents were using AI in at least one of the areas, in various ways; 80% cited four primary applications for future AI integration; and 85% of respondents had experimented with genAI for tasks such as authoring summaries and generating headlines.
The Daily Maverick, an independent South African digital news organisation, has been experimenting with reader engagement since late 2022, boosted by an innovation grant from the Google News Initiative.
By the time ChatGPT launched in November 2022, the publisher had already settled on OpenAI “as the provider with the most accurate summaries, says co-founder and CEO, Styli Charalambous.
The result is Daily Maverick Summaries – an AI-generated summation – including a headline and one-paragraph synopsis.
Feedback, says Charamboulos, has been so good that they’re now planning to offer it as the default option in the app and the homepage.
WAN-IFRA asked Charalambous 5 questions about this genAI journey.
How long did it take to test and implement?
We had to build the integration to our CMS and the new front-end to deliver the summaries, and incorporate new editing processes to ensure the accuracy of the content. The best prompts that returned the most useable summaries also took a lot of back and forth.
It took about six months to get a workable solution to test, and it has been iterated a few times, especially as new versions of OpenAI’s LLM have been released that required tweaking of prompts and some technical backend updates.
Any lessons for others keen to try?
Since we began this project so many AI tools have exploded on the scene, that I’d recommend figuring out which works best for the specific function/process you want to improve. Are you looking to improve a process (get more efficient) or create a new product/service (explore a new territory)?
It helps to look at journalism as a value chain (from ideation, creation, editing, production, publishing, content marketing, reader revenue, etc) to see which part of the process you want AI to help you get better at.
Once you know which area to focus on, having different people in a news organisation experiment with LLMs and AI tools is a good way. Many of these tools have free tiers, so it’s easy to experiment. Some organisations are forming AI workgroups to coordinate efforts, which also helps with feedback loops, and sharing wins and challenges.
Does it add to, or reduce, editing time?
Although the process of creating the actual summaries is automated, we still require sub-editors to check that the summaries accurately reflect the long-form story. We need to assume people won’t read the full story, and the summary is the version we are putting our name and reputation against. It’s no different from any other publishing we do, we must pass our internal standards and controls.
What impact on readership, if any?
The original intention of Summaries was to better serve audiences who are pressed for time and were bouncing from our site. The dirty secret of many news publishers is that most readers don’t get beyond the 25% marker of a page, which is a missed opportunity and doesn’t maximise the effort gone into creating that work.
Initially, we’ve targeted source traffic from social media because of the higher bounce rate, and we’ll be expanding summaries to other parts of sites and more versions of the summary feed.
Although the session duration hasn’t increased, we’ve seen that readers will consume summaries of at least another three articles during their visit. This tells us these particular readers are interested in bite-sized versions of the news, so we are expanding the breadth of visits as opposed to the depth.
What has been your newsroom team’s response to genAI exploration – and what other tools, if any, are you looking at?
I’d say it’s been mixed with excitement, anxiety and circumspection. That’s understandable, given all the hype and uncertainty that goes with a powerful innovation wave that we’re seeing with generative AI. There are so many possible use cases it’s very easy to get carried away, so it’s important to have a framework to prioritise further experiments. For us, this includes looking at translations, headline scoring and testing, recommendations and chat-style search features.
The post AI use cases: How genAI summaries are boosting Daily Maverick’s readership appeared first on WAN-IFRA.