The world is currently facing an artificial intelligence (AI) bubble: the proliferation of generative pre-trained transformers, a type of large language model (LLM). Just two months after its initial launch, ChatGPT, the most prominent LLM platform, has amassed about 100 million users, and every 14 days, generates texts equal in mass to all printed texts in human history.
Already, several media outlets and creative industry employers – including advertising agencies, photographers and film, TV and radio producers – have begun adopting some form of generative AI tools. The newsroom is not immune to this trend. AI in journalism is not an entirely new concept though; for example, earlier usage included the automation of data-driven news reports, text analytics and NLP (natural language processing) for linguistic analysis. However, the emergence of generative AI has occurred at an unprecedented and rapidly evolving scale. Now, computers can connect normative judgements with facts, a crucial aspect of journalism that was previously exclusive to humans.
Yet, an AI panic follows the generative AI buzz; especially within the newsroom context. There is fear that AI will replace journalists in performing generic tasks such as content creation, news editing, audio transcriptions, and voice-overs faster – and cheaper – than humans.
However, the news value chain goes beyond news production, and while an AI model can perform traditionally human tasks at a faster scale, its role in newsrooms must be seen as augmentative, not substitutional. For instance, while ChatGPT can narrate past events and write news insights given the right prompts, it cannot perform tasks that involve human judgements, such as crafting breaking news or integrating human sources into news stories. Besides, generative AI models are already inherently limited by the cut-off date of their training data, and as such cannot capture current events.
How can Local Newsrooms in Africa Utilise Generative AI to their Advantage?
For this piece, we asked ChatGPT what impact generative AI might have on African newsrooms, and it generated the following response:
The African news ecosystem already faces challenges that threaten the sustainability of its traditional media, such as the asymmetry between news content and the changing nature of consumer demand. On the one hand, increased urbanisation means news supply has become more city-focused – leaving many behind; whereas hyper-locality of data is redefining how content is being demanded (consumers now prefer content geared toward local problems). Yet, increasing platformisation of news has led to the decline of local news sources as well as a shrinking revenue base for traditional media in Africa.
This creates a crucial gap: lack of representative news distribution. For example, despite numerous country-level and international data on health, education, climate change, poverty, and several other development issues, only a limited amount of this data is produced as insights, analysis, or news that is tailored to the communities with the highest stake in such issues.
Often, the solution to this is more digitalisation by way of improving efficiency in the newsroom. But more digitalisation does not necessarily eliminate the prevailing challenges. With generative AI, the more important question to ask is: how might local newsrooms leverage the potential of AI to personalise and customise news distribution while increasing revenues? At the surface level, there are several long-term benefits to deploying AI in journalism, including reduced marginal costs of news production (e.g., publishing costs) and better economies of scale and scope (news media can vertically integrate new products using the same technology, cost, and inputs).
By focusing on the three main anchors of the traditional newsroom – news production, news distribution and revenue generation – we show how generative AI can be leveraged to improve each hook, thereby strengthening the local news ecosystem in Africa.
News production
The use of AI in news production is perhaps the most explored area of AI in journalism to date, from newsrooms such as Bloomberg building automated systems that analyse financial reports and produce news predictions about the financial market, to Xinhua News Agency developing the world’s first AI-powered news anchor. Others such as AP use AI tools to automate transcription of audio sources. There are also many tools such as Otter.ai used for transcriptions, ChatGPT that helps journalists write faster given the right prompts, and Hypotenuse AI which generates news headlines and analyses them for audience appeal. Recently, Buzzfeed announced that it would begin using ChatGPT to write personalised stories for its readers. Beyond writing news headlines, another common area is the use of AI to synthetically create media outputs such as voices, imagery and videos.
In Africa, a key approach to news production is investigative journalism and many media outlets such as BBC Africa that engage in such journalism now use AI platforms to gather and generate content through models that classify information trends and use flagging alerts. But, studies show that this adoption has been slow and limited across African newsrooms. One key reason for this is inadequate capacity and awareness on the part of journalists. More so, the current AI innovation system, including research on its use, appears to be disproportionately focused on Western countries and China.
News distribution
Value is better derived from AI when it fuses news production and distribution. However, the current AI-based distribution of news seems to come mostly from non-media organisations (e.g., platforms’ use of personal cookies to suggest news stories on users’ web browsers). News organisations can leverage AI for news distribution by using web-based data and content analytics to target the audience that would be mostly interested in a piece.
One way to imagine this is using AI to transform web-based data into short reports and news alerts tailored for local communities represented in the data. Other models could include using crowd-sourced primary data, pooled into a central platform to curate news that is tailored to underserved groups or blocs, based on specific characteristics of the data sources.
For-web based data, the AI model could gather, verify, and create content powered by geospatial and web camera data – sourced from open databases, satellite imagery, and sensors, and distributing it to target thousands of unique communities based on geography, interests, and orientation. Newsrooms can also exploit natural language processing to aggregate thousands of unstructured news texts online into stories and visualisations that are specific to local groups and communities.
Automating news insight:
Source: Dataphyte
There are two ways to achieve this. In the short-term, such models could be tightly tied to news production (sourcing), by using an alert system to notify journalists of events and data that align with their region and areas of interests. Based on this system, journalists are able to “follow the news” and “tailor” their writing to the communities they serve, reducing the gap between the tons of data produced daily and those actually utilised to form reliable insights and analysis. A closely related example here is AppliedXL, which combines machine learning with investigative journalism to map out data patterns before they get reported in the news.
In the long term, newsrooms could build applications that write several news stories and reports (production) based on algorithms that predict links between locations, events, and other web-based data. A relevant example here is Nubia AI, an Africa-focused model, which produces community-tailored development news reports according to segmentations in data sources for such news. Through this, hundreds of data-driven reports tailored for underserved communities can be produced in a timely fashion. It can also become possible to generate stories for individual users based on algorithms learning audience segmentation (language, age, location, income level, gender) and consumer preferences (industry, products, etc.). Of course, while this enables great personalisation of news content, it also carries potential risks such as biased reporting, lack of empathy and creativity in news reports, further filter bubbles and polarisation, loss of privacy, among others (further discussed in the section on ethical challenges below).
Revenue
The end goal of AI in journalism is improving efficiency by freeing up more time for journalists to perform other tasks – not replacing them. But this process could also boost revenue and optimise sustainability for traditional news outlets by reducing their over-reliance on ad-based revenue sources. This is one area where current adoption in Africa is limited but potential is numerous. Future strategies could, for example, leverage AI to create a streamlined sales funnel by identifying news consumption patterns across demographic groups and tailoring products accordingly.
African newsrooms can also use large and detailed consumers’ data, such as geolocation, combined with personalised messages and other predictive marketing techniques to improve their marketing strategies. Lastly, news media managers can more effectively engage in dynamic pricing by employing algorithms that predict the optimal price for products and the right time to open or close a pricing deal, using the vast amount of consumer preferences and demographic data available. Needless to mention, this must be undertaken responsibly, as there are significant ethical issues with personalised marketing techniques, such as reinforcing filter bubbles and the possibility of consumers’ data being sold or re-used for purposes other than those they agreed to (see more in the ‘Ethical challenges’ section).
Indeed, the potentials of AI in local journalism are numerous – from increased publication of personalised and locally grounded stories, to cost reduction, and expanding the financial funnel of newsrooms. Yet, this comes at no easy cost; there are obvious ethical challenges that stymie these prospects.
What are the Ethical Challenges?
The first point of worry is that the majority of newsroom AI adopters are the big players in developed economies, such as Associated Press (AP), Bloomberg, BBC, Thomson Reuters, among others, and there is little to no adoption by media outlets in Africa, or even by smaller newsrooms generally. A recent study by AP on AI in U.S. newsrooms found a wide gap between large newsrooms and smaller/local news organisations, and that only a few local newsrooms utilise AI for news production. As with other forms of innovations, new technologies require time to learn, and as such generative AI replicates the “winner-take-all” effect where technologically advanced big names in the industry can more easily enhance their market power by leveraging already existing tacit knowledge to optimise productivity with the new technologies.
Yet, the most transformative effect comes from using AI to close the divide in efficiency and productivity between large and small newsrooms. A key question here is whether policymakers could step in to provide both financial and technical leverage for smaller newsrooms, such as the financial support and subsidies given to SpaceX and Tesla in their early days, helping both ventures to revolutionise the space and electric vehicle industry, respectively.
The challenge for African countries is more acute: weaker levels of industrialisation and poor digital access inhibit mass adoption of AI by small and big newsrooms equally. In Africa, most AI solutions are focused on the financial services, agriculture, and healthcare sectors, with very few use cases in the newsroom.
The second ethical concern has to do with hallucination tendencies of LLMs. Even with constant human prompting, when generative AI models do not understand the questions they are asked, they tend to hallucinate (i.e. produce factually inaccurate content, or content that does not match existing inquiries or the database they are trained on). This becomes even more problematic in that the models could occasionally jumble up existing facts and misinterpret them due to inadequate training on those phenomena. The problem is that LLMs tend to be more useful in analytical pieces that draw on generic publicly available data, and less useful for breaking news, which requires current contextual understanding that has not been incorporated into existing models. As such, news consumers must constantly fact-check content from AI-powered news platforms. Innovators must also prioritise transparency, including defaulting to White Box models, acknowledging risks and inadequacies, and offering non-technical explanations about how their models work.
We must also be conscious of other social and ethical challenges associated with deploying AI in complex human phenomena like news production and consumption. For instance, many image-generation models lack diversity in their datasets and have been spotted reinforcing pre-existing real-life gender and racial biases. For example, one image classification model identified Black and Latino women as “homemakers” more than white men; some others were less likely to display women when prompted with search strings such as “doctors” or “gastroenterologists”. In fact, a popular image generation AI, Stable Diffusion, was recently found reinforcing such harmful and stereotypical depiction when it showed “African workers” in a poverty-stricken context, compared to “European workers” whom it portrayed in a more sophisticated and industrialised environment. If unresolved, using these models in newsrooms replays the poorly constructed narratives about Africa often seen in Western media. Perhaps it is time African innovators started considering Afro-centric image-generation models (a useful reference point here is Chidiebere Ibe’s use of black medical drawings to address the lack of diversity in medical illustrations).
Concerns about data privacy also threaten progress with generative AI models. Recently, the Italian Data Protection Authority placed a temporary ban on ChatGPT’s use of Italians’ data – essentially limiting its operations. Yet, there is a delicate balance between protecting individual data privacy and enabling data-sharing practices that could engender further innovative models; since LLMs are generally built on large troves of data which on one hand benefit from continuous access to human data for training new models/algorithms, but on the other require frontier AI companies to share their codes and data with peer companies.
In the newsroom, the bigger question is how to strike a balance between AI generation of content and human moderation. In this context, editorial boards must pre-define moderation terms, content generation limits, and stringent verification procedures, such as provenance documentation for AI generated media and other synthetically generated content.
Where do we go from here?
The path to AI innovation in Africa requires enormous co-creation, prioritisation, and financial investment. A peer learning community is required for pioneers and newcomers in the news industry to learn, share and collaborate across the spectrum of integrating AI into media operations. Open-source technology communities such as Open AI should prioritise local news media alongside other critical industries. Generative AI innovation is capital intensive, and as such there is a need to expand the conversation to philanthropy and foundations to provide the necessary catalytic funding for training, research, and business development.
There is also the need to prioritise the ethical issues discussed above, while engendering innovation. Safeguarding social and ethical values while using AI in the news ecosystem requires the input of multiple stakeholders – from AI companies to journalists, news consumers, and the government. But more than anyone else, newsrooms must have clearly spelt-out corporate digital responsibilities that guide their adoption and use of AI.
Relatedly, AI governance policy in Africa also needs to move as swiftly as innovation does. A report by ALT Advisory shows that as of September 2022, only about four African countries have a complete national AI strategy. As a response, African governments need to be proactive about developing both an enabling environment (investment in research and development, education, and skills) and a unique LLM governance approach that is sensitive to local and cultural news context.
Ayantola Alayande is a Research Consultant in Civic Technologies and Public Policy at Dataphyte.
Joshua Olufemi is the CEO of Dataphyte and Co-founder of Nubia AI —an AI-powered news platform that auto-creates development news and data insights on Africa’s local communities.
