Connected through the London School of Economics JournalismAI Collab, representatives from eight news organisations around the world — Reuters, AFP, Nice Matin, Nikkei, Schibsted, Reach PLC, La Nacion, and Deutsche Welle — decided to explore how AI could help us serve diverse audiences. 

Together, we created the AIJO project. 

The name represents our ambition to pair AI with JOurnalism, but also is symbolic; AIJO (Aijō, 愛情) means “love” in Japanese. If used in ways that align with and serve our publishing missions, we believe there is great potential in using AI technologies to create more inclusive — and even loving — journalism. 

Making the case for diversity

Failure to represent and report on different people, points of view, and lived realities can feed the injustices of our current world. With global flashpoint moments in 2020, news organisations across the world began asking how well we serve communities big and small, powerful, and marginalised. 

If the moral argument isn’t enough, money talks: If we as publishers under-represent or misrepresent parts of our population, we will struggle to find new audiences and subscribers beyond our established base. 

Focusing on gender bias

 We believed that bias in the newsroom influences what level of diversity is achieved in both its organisation and output, so we focused on exploring how we can leverage AI to understand, identify, and mitigate bias. 

Coming from different cultures, our newsrooms (and societies) struggle with different kinds of bias. What is an issue in Sweden is not necessarily one in Japan, and vice versa. 

But there are shared challenges. 

We all reported how men and women are given different opportunities in our societies, albeit in varying ways. We decided to assess how gender bias is manifested in our shared activity of producing and publishing content. 

Gender representation in news imagery

We believed that computer vision, one of the many concepts behind the AI term, could help us identify newsroom gender biases by uncovering who we depict. Who is visually represented in our news stories?

By focusing on images, we could aggregate news media content from all over the world to see how we are doing on a global level. Using an open-source AI model, we analysed 28,051 images and measured the percentage of females represented in the image, then aggregated the data to get the average ratio of all images. 

The AI model determined that, across our eight news organisations, the average ratio of females represented in images was 22.9%. Among the participating publishers, this ratio ranged from 16.6% to 35.7%. 

We reviewed 10% of the total images with a human eye. In this sample, the share of females increased to 27.2%. The female ratio among publishers in the human-reviewed samples ranged from 19.4% to 36.2%. An important note here is that while we did not review the full dataset manually, the actual share of females is likely somewhere in between the AI assessed ratio and the sampled 10%. 

Analysing our texts

The AIJO-organisations publishing in English also conducted a natural language processing (NLP) analysis. Partnering with the discourse lab at Simon Fraser University, the Gender Gap Tracker noted the proportion of men and women mentioned and quoted. 

AFP, Nikkei, Deutsche Welle, and Reach ran the tracker on one week of their content, leading to an analysis of 1,430 articles.

On average, women represented 21% of the total people mentioned in these articles, compared to 73.3% men. Around 5.6% of the names mentioned could not be identified as either gender by the AI tool. As for quoted sources, 21.92% were women.

Events in the world that week and our own editorial priorities naturally impacted these results; the news reflects what’s happening in society.

But that’s not the full story.

The way we talk about the world can be subject to bias, including gender bias. One way of understanding if and how that manifests itself is to examine quotes. Are we giving men and women the same amount of airtime?

The Gender Gap Tracker found the average quote length was 103 characters for men and 97 for women. The difference is subtle, but when added to the rest of our findings, it illustrates a worrying tendency of dominant male representation globally. 

Looking ahead

With racist facial recognition, rabbit holes of radicalisation, and questionable cases of automated journalism, critical reports on AI have added up in recent years. There are indeed significant risks associated with employing AI in a news media context. But there is also immense opportunity to do good and use these new technologies in ways that align with and serve our publishing missions. 

AI alone will not be the answer to the media’s diversity crisis. We need committed people to work toward a much-needed cultural change. We believe that AI can help us — humans — understand where action is needed. 

Visit www.aijoproject.com for more information about the project and our experiments. Banner graphic by Aurore Malval.

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