Every Cinco de Mayo, my friends in media and advertising and I share the worst ads we can find pegged to this long-misunderstood date. In our hall of shame: detergent promising a “fiesta like there’s no “mañana” in the laundry room, a “Flyo de Mayo” for you and your “amigos” to go somewhere “hot” and mayonnaise to “put the mayo” in your celebration. 

I don’t latch onto Cinco de Mayo because it’s a convenient target (it usually is), but for a more important reason. Cultural phenomena in the Latino sphere are regularly misunderstood, co-opted and distorted by marketers and advertisers, and this tendency is top of mind for me for two reasons. First, it’s Hispanic Heritage Month. Second, large language models (LLMs) such as ChatGPT will supercharge the “Cinco de Mayo effect” of misrepresenting and stereotyping Latinos in media and advertising—unless creatives are brave enough to combat it.

With every prompt or query, LLMs betray the linguistic and cultural biases they absorb. Reported biases in content generated by AI include underrepresentation of minorities, perpetuation of stereotypes and botched attempts at capturing linguistic nuance. And that’s just in written results.

A recent analysis by Bloomberg of more than 5,000 images created with Stable Diffusion revealed that the language-based image generator took racial and gender disparities to extremes. Image sets for high-paying jobs were dominated by faces with lighter skin tones, while people with darker tones dominated prompt results for “fast-food worker” and “social worker.”

For designers and strategists in creative agencies, the incentive to look past these shortcomings are literally baked into the product UI. Need some inspo? Enter some terms into the text field. Not quite what you had in mind? Keep pumping in prompts. The ease of the transaction overlooks the harm we exacerbate, wittingly or unwittingly, by letting the results influence our work. It’s not like the product architects are riding to the rescue any time soon. 

Consider that Sam Altman, CEO of the ChatGPT developer OpenAI, has acknowledged the tool’s linguistic bias. OpenAI’s report card for GPT-4 concedes that it was “mostly designed, built, and tested primarily in English and with a US-centric point of view,” that “the majority of [their] pre-training data and alignment data is in English” and that their models had not been robustly tested for multilingual performance.

See A Celebration of Creative Excellence for Hispanic Heritage Month

“Any good Spanish results are a bonus,” wrote an OpenAI developer on the company’s support forum, noting that the models are “purposely trained to use mostly English.” For reference, some calculations show that more than 58% of the content on the internet is in English, while content in Spanish accounts for a little more than 4%.

And while the algorithm-based magic that LLMs perform can feel mysterious, it’s easy to understand what creates biased results: biases in the datasets on which the models train. We consequently face the danger of many stories and perspectives being buried, regardless of their cultural origins.

Latinos bring a variety of identities, voices, dialects, slangs and languages to the table. The heritage being celebrated this month in the United States is, in essence, that impossible sum of contexts and contradictions, encompassing the lived experience of people hailing from more than 20 countries.

So, as concerned and well-meaning creative partners for our clients, what can we be doing to combat these very real dangers?

First—and it’s painful to have to put this in writing—it’s a good idea to have Latinos represented on the creative and strategic teams crafting the communication.

Second (and this pairs rather well with the first): We must acknowledge that the risk of having a message about and for Latinos land flat—or be flat-out incorrect—has increased substantially in the generative AI era. Attuning your teams to bias in generative AI results is not a step best left for final QA. It has to be part of the project kick-off.

Counterintuitively, for all the efficiencies that gen AI is helping drive across creative industries, content creators will need to work harder in this first age of massive gen-AI adoption to get Latino stories right. As my Cinco de Mayo hall of shame readily proves, Latino stories and identities are already difficult to grasp, and much harder still for machines trained on limited data sets. We all will need to learn to use these tools the right way by developing culture-aware prompts and socializing them across our organizations.

Hispanic Heritage Month is an opportunity to bring people together, typically in school and office environments, to learn about the traditions and identities that comprise Latinidad. Let’s make this year’s edition the first where humans and machines come together not just to tell truer Latino stories, but better ones.