Publicis hopes new Arabic-language AI tool will reverse Anglophone bias in Middle East

The Open AI tool currently supports a handful of languages – including many of the most widespread languages on the planet: Chinese, English, Spanish, Portuguese, French and Japanese. However, Arabic, and its many dialects, spoken by more than 420 million people worldwide, is not currently included.

Copy created using generative AI tools can, of course, be translated – but this is a time-consuming activity for brands and agencies (and leaves one relying on translation software, or the services of a human translator). Some 30% of marketers say translation is the single most time-consuming part of copywriting, according to a survey this week released by digital asset management firm Bynder.

Accuracy is one of the key factors behind Large Language Model (LLM) development in languages other than English (which, given the amount of material available, is the cheapest to develop). The Lince Zero LLM, for example, was released back in July with just that aim.

Regarding Arabic, one of the most promising efforts in this area is Jais, an Arabic LLM developed by AI company Core42 and the Mohamed bin Zayed University of Artificial Intelligence in Abu Dhabi. A new partnership deal between Core42 and Publicis Groupe’s Middle East business is now set to give the agency group’s creatives access.

Bassel Kakish, chief executive officer of Publicis Groupe Middle East & Turkey, tells The Drum that the ultimate aim is to “leverage the power of Jais” to create more personalized marketing messages for Arabic-speaking audiences. “Combined with our Epsilon tech, this means true Arabic hyper-personalization at scale… allowing us to be at the forefront of reaching the consumer in the dialect they are most familiar with,” he says.

Kakish hopes the tool will enable the firm to pursue those goals while dodging potential landmines, too. Using LLM not fine-tuned for Arabic, or the differences encoded in its various dialects, could lead to creative output that misses the mark, misrepresents a brand’s aims – or simply causes outright offense. “Biased or culturally insensitive content may result in lost market share to competitors who better understand and cater to the local audience,” he says.

He adds: “Developing a more sensitive LLM ensures that the generated content is culturally and contextually appropriate, which is essential in the field of advertising and marketing.”

“We recognize the importance of preserving and enhancing the reputation of the brands and advertisers we work with. Inaccurate or insensitive content can damage a brand's reputation, potentially leading to negative publicity, reduced customer trust, and even legal issues. By addressing bias in content generation, we can mitigate these risks efficiently and effectively.”

Publicis’ access to Jais, he says, will be a particular focus for its Sapient business in the region. In return, feedback from Publicis’ 3,600 staff based in the region will be relayed back to the LLM’s developers, allowing them to tweak the tech. Andrew Jackson, chief AI officer at Core42, says the partnership will aid efforts to “make [Jais] not only more culturally sensitive, but also highly attuned to the specific needs of those within the marketing field.”

The partnership is also designed to steer the development of the LLM towards a product more useful for advertising, Kakish says that the firm will be working to develop marketing use cases for Jais in the near future, including copywriting outputs in different Arabic dialects and AI-generated voice content. It’ll be putting in place a region-wide QA team before deploying the tool on client projects “for a long time,” he says.

“We are bringing the experience of 100s of copywriters, content creators, and creatives with experience across a wide range of Arabic dialects, to QA the LLM’s output based on the input given,” he adds. “We aim to leverage Jais to make digital channels more impactful and effective for our customers and clients.”

The Anglophone stance of most AI developers has a secondary risk. If English becomes the primary language of AI-enabled marketing and business, the use of other languages could be discouraged in the long-term. That could damage businesses in the region as well as linguistic diversity, Kakish notes.

“As the world embeds more and more AI in all it does, if no large language model in Arabic existed – one that is accurate, precise and unbiased – then the region would have become less competitive as a whole,” he says. “Globally, the development of LLM models is led by English as a first language. Our region is unique, and language plays a key role in its culture. So, besides the commercial aspect, preserving the Arabic language nuances and subtleties, all while driving innovation is important.”