Will AI replace our news anchors?

On 19 July, a Dhaka-based private television channel made headlines by introducing 'Aparajita,' the first AI presenter in Bangladesh. 

The emergence of AI presenters across the world has sparked intense discussions in both mainstream media and social media. While some worry about the potential impact on news anchor jobs, there's also a crucial question: Does Bangladesh possess the technology to create a fluent Bengali-speaking AI avatar?

Inside AI technology

"Building effective artificial intelligence technology necessitates vast amounts of data. Approximately 10,000 hours of audio data are needed to achieve a minimum sensible outcome," said Farig Sadeque, an Assistant Professor of Computer Science and Engineering at BRAC University. 

Sadeque holds a PhD from the University of Arizona with research experience in computational linguistics, applied natural language processing and machine learning. 

"For a language-based AI, extensive audio recordings are required to enable sensible speech, encompassing various regional accents and idioms. At the same time, it is important to have various noises around, so that noisy speech can also be identified. However, having data alone is insufficient; a complex AI model is also necessary to process and utilise the data effectively." 

The recent rise of AI is largely due to the Transformer Model – a neural network that learns context and meaning by analysing sequential data, such as sentences. High-performance computers are vital for designing and processing these advanced models. 

Unfortunately, such powerful AI models are still relatively scarce in Bangladesh, requiring significant investment. Moreover, the lack of skilled machine learning engineers in the country further hinders the development of complex AI products.

Creating generative AI, like ChatGPT, presents more challenges than virtual AI avatars, which primarily involve converting written words into audio. Meanwhile, generative AI requires the analysis of recorded audio and video data to achieve natural facial expressions and gestures.

In the case of the likes of Aparajita or AI avatars, a facial expression is added to the audio. News anchors don't need to make too many faces when reading the news. And so an AI avatar also does not require much video input due to the absence of facial expressions such as anger, sadness or laughter. 

"As a result, in the same way as audio, the video data is analysed to see how the shape and position of the lips, the facial muscles change during the utterance of a sound," said Sadeque. The more data is analysed, the more fluent and natural the face of the on-screen AI avatar will appear.

How AI is being used in newsrooms

Generative AI tools, including ChatGPT, are widely adopted in newsrooms worldwide for summarising long reports, offering topic research guidance, spell-checking, writing full articles, generating topic ideas and translation. 

However, the use of AI presenters has surged recently, with various countries, including Bangladesh, adopting them. Notably, some news channels rely on Software as a Service (SaaS) companies for their AI presenters, raising questions about the extent of in-house AI research and development.

According to a survey by the World Association of News Publishers published last May, more than half of newsrooms around the world use generative AI tools like ChatGPT. 

AI started as a news presenter in 2018 with China's Xinhua Agency. According to The Guardian, the AI presenter was modelled after Xinhua Agency presenter Qiu Hao by processing facial and voice data using machine learning.

By 2021, more AI presenters appeared in four newsrooms in China and South Korea in a similar fashion. South Korea-based company DeepBrain AI was involved in all of this.

Joe Murphy, Business Development Manager at DeepBrain AI, explained the entire process of making this AI presenter in an interview with VentureBeat, a technology-focused media outlet. It is known that the original presenter was given a special script of 500 to 1000 sentences to read out, which contains almost all the phonemes of the language. 

It was captured how the facial expression changes while pronouncing a specific sound, where the presenter pauses or takes a break. This data is then used to create an AI video avatar, capable of pronouncing any words like the original presenter. According to Murphy, it took about a day to do a video shoot and about three to four weeks of machine learning time on the computers to generate the first AI model.

Media outlets usually don't have the full-fledged capacity to produce their own AI avatar, therefore they use the products of different SaaS (Software as a service) companies. 

Commenting on this, Sadeque said, "I don't think any media has yet developed the technical capabilities where they can build a completely AI presenter from scratch, who can present news in a completely new language fluently through natural gestures."

He added, "This kind of power is only in the hands of big tech giants, who have invested billions of dollars behind it. Let's take OpenAI, which is associated with Microsoft, they have a lot of investment behind them."

However, since the amount of data processed here is small, the new language has some linguistic complexities, so facial abnormalities, very slow speech or slurred speech will remain to some extent. Further processing is required to normalise it.

Is AI a threat to news anchors?

AI news presenters like 'Aparajita' only convert text to audio. But generative AI is much more complex. To replace a human news presenter, AI needs the understanding and processing of natural language to rearrange questions and answers coherently and humanly, making it more challenging compared to current AI avatars.

"AI is now far from replacing human news presenters. AI is yet to develop spontaneous emotion in face and voice—it literally and symbolically lacks 'life.' Maybe in the distant future something could emerge, but in the near future I don't see that happening," said Foysal Morshed, who has been a news presenter for years on different national television channels.

The primary hurdles in this domain are data adequacy and high-performance computing. While there are chatbots like ChatGPT for English, there is a notable absence of similar chatbots for languages like Bangla due to the scarcity of data and suitable processing capabilities. 

However, the rapid advancement of AI makes researchers and scientists struggle to keep up with its trajectory. What seemed unimaginable just five years ago – in terms of what AI can do – has become common, even for the general public. 

As AI models become more adept at analysing data and learning independently, predicting the scope of AI's future capabilities becomes increasingly challenging. 

What might an AI-powered newsroom look like?

A fully AI-powered newsroom has the potential to revolutionise journalism in numerous ways. Konstantinos Antonopoulos, a journalist at the Al Jazeera Media Institute, envisions several key tasks that AI can perform to assist journalists in their work. 

First and foremost, AI can enable real-time transcription, automating the conversion of audio to text. This advancement eliminates the need for manual transcription, saving journalists hours of tedious work and allowing them to focus on more critical aspects of their reporting.

Additionally, AI can facilitate the seamless translation of content into various languages. This capability enables news organisations to deliver the same news to a global audience rapidly and efficiently. 

Furthermore, AI can simplify complex language, transforming intricate reports into more accessible and comprehensible content. This makes news easier to understand for a broader audience, fostering greater engagement and comprehension.

AI's ability to generate captivating headlines and optimise content delivery is another significant advantage. By analysing the reading patterns and preferences of the audience, AI can tailor content to suit individual readers, increasing their interest and engagement with the news. 

Moreover, AI's connection to various databases allows it to verify the authenticity of news stories, reducing the need for manual fact-checking and enhancing the accuracy of reporting.

The AI-powered newsroom also can gauge reader sentiment and opinions about the content. By analysing user feedback, news organisations can better understand their audience's reactions and tailor their reporting accordingly. 

Risks of using AI in journalism

The survey conducted by the World Association of News Publishers shed light on the perceived risks of using generative AI in journalism. Surprisingly, only 38% of surveyed journalists believe AI poses a threat to their job security. Instead, they are more concerned about other risks, such as misinformation (85%), plagiarism or copyright infringement (67%) and data security (46%).

While the threat of job losses to AI is not entirely trivial, it remains a valid concern. To explore the capabilities of AI in reporting, Tom Clarke, editor of the science and technology department at SkyNews, experimented using Python programming. 

He employed ChatGPT as two separate AI entities: an AI reporter and an AI editor. The AI reporter and editor collaborated to create a script for a TV news show featuring multiple stories.

The AI reporter demonstrated competence in generating logical and informative story ideas, offering excellent advice on scriptwriting and suitable footage. However, it also exhibited some limitations, including the propagation of misinformation and the reliance on existing story ideas from different media sources. The AI reporter struggled to produce completely new and creative content.

In addition to spreading false and fictitious information, AI in journalism poses several other risks. It may misinterpret sarcasm or metaphors, oversimplify language at the expense of crucial details, and trap users in filter bubbles by analysing their behaviour to make content suggestions. 

Furthermore, there have been reports of virtual avatars being exploited to spread fake news and propaganda in countries like Venezuela.

While it is undeniable that AI has changed the way newsrooms operate and made certain tasks more manageable, the rapid pace of AI development makes it hard to predict the future. 

Professor Beckett of the JournalismAI Project at the London School of Economics emphasises that reporters can use AI to offload simple and time-consuming tasks, allowing them to focus on more interviews or creative and opinionated storytelling, which mechanical AI cannot excel at.

Miraz Hossain contributed to this report.