How Marketers Are Navigate AI's Language Gap

When Mike Sotelo, vice president of digital content and experience at ad agency Alma, experimented with OpenAI's generative artificial intelligence tool GPT-3 in Spanish earlier this year, it wasn't long before he came across a language gap. Sotelo entered two prompts: “How does a guey [a Mexican-Spanish slang term for a man] look in Miami?” and, “How does a tio [a Spain-Spanish slang term for man] look in Miami?” While the AI model's response correctly identified "tio" as a male, it did not recognize the term "guey" and instead provided a gender-neutral term in its response. Generative AI tools are far from perfect, often lacking the most up-to-date information or cultural nuance, but they are improving at a clip. For now, limitations in understanding regional linguistic differences are causing headaches for multicultural agencies and potential pitfalls for marketers trying to reach a wide, diverse audience. Agencies are turning to more manual processes to verify and train these AI models to avoid alienating consumers. Without requisite and intentional training, these problems will only self-perpetuate. "The language gap can create risks for brand executions with varying impacts across industries," said Andrew McKernan, group vice president of content and consumer experience strategy at Razorfish. A travel and hospitality brand wanting to interact with visitors who speak French and Spanish can’t rely on an English-language experience with supplemental translation. "We need to ensure that we are addressing people with authenticity and cultural competence," he added. A predominantly English-language internet A plethora of ethical concerns including racial bias, hallucinations (when AI generates inaccurate information) and the fear of job loss surround generative AI tools that are predominately trained on text and data in English extracted from the Common Crawl—a nonprofit, open repository dataset that contains billions of webpages across the internet. Nearly 60% of the data used to train ChatGPT’s 3.5 series comes from Common Crawl, according to sessions from machine learning conference NeurIPS 2020. The vast expanse of the internet is dominated by a handful of languages. While generative AI is empowering marketers to accomplish tasks with increased efficiency and speed, the language disparity may cause those reaching non-English-speaking audiences to generate lower-quality results from AI tools, according to Adolfo Corujo, partner and Deep Digital CEO at communications and digital marketing agency LLYC. “The tools aren’t prepared to be used as intensively as in the U.S., and that's a huge problem for the marketing departments,” said Corujo. “[Marketing agencies] in Spain, for example, will need more human intervention [to assess the output] compared to the U.S.” This language gap poses a problem for marketers in the U.S. as well. As of 2022, about 42 million people in the U.S. spoke Spanish as their first language, per Statista. That's nearly 13% of the country's total population. Training AI models on relevant phrases The problem, according to Michael Olaye, ‪svp and managing director of strategy and innovation at R/GA, is the underrepresentation of people within the industry. “Most of the AI tools are being built by white men, so bias will exist,” he said. R/GA's brand partners actively seek to reach non-English-speaking audiences in the U.S. and other regions like Latin America. The agency trains generative AI models in languages such as Spanish, French, German and Chinese. These models aren't trained on isolated words but on relevant phrases for each language, receiving feedback through reinforcement processes. Across agencies, Olaye said, this task is either delegated to people in the U.S. who are native speakers of a particular language, or to countries where the languages are native. “It's straightforward in practice, but in reality, it's quite challenging,” he said. Meanwhile, involving diverse human intervention to review and assess AI's content output is becoming even more critical. “Brands, agencies and the teams that work with AI need more diversity,” said Olaye, “whether you're training [AI] or you're using them to create [ads].” At Ogilvy Paris, nearly 99% of the prompts provided to generative AI tools are in English, said executive creative director David Raichman, because English is the primary trained dataset for all versions of AI, so it should generate the best results. When a campaign needs to include the French language, the agency trains AI models with specific French elements, a task that on average takes a day and falls under the remit of the agency's copywriters. A self-perpetuating problem Despite manual workarounds, the AI language gap is only expected to grow. Tools like ChatGPT consistently churn out an ever-increasing volume of content each day, which operates as a feedback mechanism to further train the model. In September alone, ChatGPT saw 1.5 billion visits across desktop and mobile web, according to Similarweb. The desktop traffic alone increased by 5% compared to August. This surge in AI activity intensifies the divide between English and other languages, said Corujo. “It’s essential to recognize that the variation of words can dramatically impact the degree in which [marketers] are solving the problem," said Raj Manocha, chief client officer at research and data firm Sago. "Certain words create more inclusivity and can help to align more with client segments. This becomes critical to ensuring AI is more client-centric and fits deeper with your brand identity.” Being transparent about biases It's worth pointing out that there's no such thing as a perfect, entirely unbiased AI model. However, according to Sasha Luccioni, AI researcher and climate lead at startup Hugging Face, transparency about these biases is crucial in tool creation. During Luccioni's tenure at Bloom, the first open source language model, her team spent months reaching out to historically underrepresented communities and created avenues for them to consensually share their data with Bloom via hackathons or licensing agreements. “We managed to get a lot of African and Indic languages represented in the data that way,” said Luccioni. This story is part of the Power of Tech Marketing special feature.

AI News

How Marketers Are Navigating AI's Language Gap

Biases will exist when white men are building most AI tools

Trishla Ostwal

About

Subscriptions

Events

Publications

WORK SMARTER - LEARN, GROW AND BE INSPIRED.

Subscribe today!

Trishla Ostwal