Although AI-driven marketing technology (martech) can be powerful, improper training, weak algorithm designs and hidden prejudices will quickly derail any campaign.
How can you identify and mitigate bias before it impacts your brand?
Dove, the personal care company owned by Unilever, just became the first brand in its industry to pledge not to use generative AI in advertising. An accompanying two-minute video depicts search terms like “most beautiful woman” and “perfect skin” returning AI-generated models.
The video cuts from the uncannily flawless faces to real women, highlighting the impact of impossible beauty standards. This move is part of the Dove Campaign for Real Beauty, a 20-year campaign focused on showcasing all body, skin and hair types.
Dove may be the first to commit to such a pledge, but it likely isn’t the last. Its decision comes after various other brands have come under fire for experimenting with AI-generated content. Critics and consumers alike claim these tools favor certain stereotypes and demographics.
The Washington Post’s study on bias in generative content seemingly proves Dove’s point. The leading generative models — including Midjourney, Stable Diffusion and DALL-E — favored thin, light-skinned women when asked to generate a “beautiful” woman.
Nearly 90% of Midjourney’s images depicted light-skinned women. DALL-E and Stable Diffusion were barely better, with only 38% and 18% featuring dark-skinned individuals, respectively. This evidence suggests the potential algorithmic bias in martech is significant.
Algorithmic bias in martech can influence consumers’ perception of a company. If it is glaring or consistent enough, it may impact a brand’s revenue, reputation or conversion rate:
Marketers who disregard the potential implications of algorithmic bias in martech may lose website traffic, anger customers and miss out on potential sales.
Generally, generative AI’s algorithmic bias is visible if you search for it. However, all kinds of models can produce prejudiced output — and it is often challenging to recognize. You should know how to identify it so you can mitigate it.
Your go-to method should be to audit your model’s data feed before integrating it into your martech. Minimizing the amount of stereotypes, assumptions and unfounded beliefs that get into training prevents it from learning to be unfair or bigoted.
A human-in-the-loop system is one of the most effective methods because it pairs one of your team’s marketing professionals with an AI. This person reviews its output for prejudice, ensuring only high-quality, relatively bias-free material makes it to consumers.
Another method involves analyzing your industry's historical and current state of bias. Since some use cases have a higher prejudice potential to specific communities than others, assessing the broader scope of prejudice can help you determine its likelihood.
No matter what type of AI you integrate into martech, you should ensure you develop a feedback loop. Since machine learning models learn based on feedback and interactions, inadequately addressing biased output reinforces it over time, perpetuating the problem.
Bias will remain ever-present in marketing because it is a byproduct of societal norms and modern expectations. Prioritize fairness and inclusivity in your practices to minimize it:
In today’s age, prioritizing fairness and inclusivity will get your brand far. Listen to customers’ concerns about AI and prejudiced martech to inform decision-making.
While nothing can be truly bias-free, you can come close if you regularly audit your model’s output, review your martech for accuracy and listen to feedback from your customers.
In an age of high AI concerns, diligence and consistency will result in better business outcomes.