Don't Let AI Bias Subvert Your Marketing Efforts

Relying too heavily on AI and omitting human oversight is a dangerous proposition.

Image created by ChatGPT-4/DALL-E

We welcome Brianna Blacet, our new co-editor to AI Marketing Ethics Digest. This is the first of many insightful articles you can expect from her.

The use cases for generative AI (GenAI) in marketing seem almost limitless. Marketers use ChatGPT to write marketing plans, outline and write blog posts and ebooks, analyze data to derive user insights, optimize websites for search engines, and so on. 

In most cases, AI has made marketers more efficient, allowing us to focus on tasks that require actual humans. Why not “hire” ChatGPT to write your SEO articles and social posts, use AI to plan next quarter’s campaigns, and leverage a chatbot for front-line customer support? Quick and easy, right? 

Not so fast. Relying too heavily on AI and omitting human oversight is downright dangerous. A dark stranger lurks inside the machine learning (ML) models that drive your AI’s decision-making: bias.

Why Is AI Bias an Issue?

To understand how bias can insert itself into automated systems, we need to start by recognizing that AI is only as good as its ML models, which are only as good as their training data. Models “learn” by identifying data patterns and anomalies. AI uses that learning to make its “decisions.” 

For example, an ML model may leverage large amounts of traffic data from a big city. The patterns it analyzes can help AI identify ways to ease congestion or prevent accidents. However, AI-generated decisions can go wrong if the ML model contains faulty data or inappropriately characterizes patterns. 

A recent example was Google’s launch of the image-generation feature used by its conversational AI tool Gemini (formerly Bard). Due to what the company referred to as “tuning problems” in its ML model, Gemini generated images of racially diverse Nazi-era German soldiers and non-white U.S. “Founding Fathers” — and even inaccurately portrayed the races of its own company co-founders. A giant gaffe, to say the least! 

“So what went wrong?” asked Prabhakar Raghavan, Google Senior Vice President of Knowledge & Information, in a Feb 23, 2024, blog post entitled “Gemini image generation got it wrong. We'll do better.” 

“In short, two things,” he explained. “First, our tuning to ensure that Gemini showed a range of people failed to account for cases that should clearly not show a range. Second, over time, the model became way more cautious than we intended and refused to answer certain prompts entirely — wrongly interpreting some very anodyne prompts as sensitive. These two things led the model to overcompensate in some cases, and be over-conservative in others, leading to images that were embarrassing and wrong.” 

Another high-profile example of ML model bias was Amazon’s attempt to use AI to review résumés from potential job candidates. The company trained its ML model on a large number of résumés written primarily by men. Unsurprisingly, the AI algorithm based on this dataset discriminated against résumés submitted by women candidates for technical jobs.

Where’s That Data Coming From?

If you’re using ChatGPT, join the club. But have you asked ChatGPT where it gets its data? Its answer: 

“It has been trained on a diverse range of Internet text, including books, articles, websites, and other publicly available written material. This vast dataset provides the foundation for ChatGPT's understanding of language and enables it to generate responses based on patterns and information present in the data it was trained on.” 

When you consider that this mountain of information reflects decades of real-world gender bias, racial inequality, socioeconomic discrimination, and other nasty things, it’s easy to see how these biases could creep into the AI’s responses. 

Google sums it up succinctly in a training course for developers: 

“When building models, it's important to be aware of common human biases that can manifest in your data, so you can take proactive steps to mitigate their effects.” 

The course covers several categories common to ML model bias, a few of which are particularly relevant to marketing:

• Reporting bias occurs when the “frequency of events, properties, and/or outcomes captured in a dataset does not accurately reflect their real-world frequency." A great example of reporting bias is pharmacological adverse events (ADEs). The FDA believes that its system for gathering information about ADEs, the Adverse Event Reporting System (AERS), receives reports for only about 1 to 10 percent of all ADEs. Because the vast majority of events go unreported, the data is inherently flawed.
  

• Selection bias happens when a dataset's examples are chosen in a way that does not reflect their real-world distribution. The Google developer course gives an example of a model trained to predict future sales of a new product based on phone surveys conducted with a sample of consumers who bought the product in the past. Consumers who opted to buy a competing product were not surveyed, and as a result, this group was not represented in the training data.
  

• Sampling bias is when data collection does not include proper randomization. An example might be surveying participants only from large metropolitan areas. These groups may have opinions or buying habits that differ dramatically from would-be participants in the middle of the country, which would produce biased, inaccurate results. 

How to Minimize Risk

While there is no way to completely eliminate ML model bias, there are things that you can do to mitigate your risks and use AI responsibly: 

Carefully consider your use cases.

There’s little harm in having ChatGPT write up your rough drafts of marketing emails or SEO content (as long as you, the human, take responsibility for the content of the final drafts).

However, it’s probably better to avoid using AI to create surveys intended to produce accurate insights — unless you’re feeding in the dataset. As discussed earlier in the section on selection bias, ChatGPT might be accessing an incomplete dataset. Knowing where your data is coming from can help you avoid potential pitfalls.

Think about the roles that gender, race, location, sexual preference, and socioeconomic status play in specific marketing activities.

As discussed in the Amazon example, applications of AI in hiring or recruiting can be dicey. In marketing, if you want to understand purchasing patterns to create automated promotions based on these user demographics, realize that a biased dataset will render derived insights useless. The same goes for pricing: a biased, incomplete, or non-randomized sample will lead your decision-making astray.

Meticulously examine your datasets.

As you choose where to use AI in your marketing activities, ask yourself, “Where will the data come from?” If the answer is “I don’t know,” take some time to think it through. Remember that the utility of AI is dependent on data quality. You don’t want to make marketing decisions based on outdated or unrelated data about your ideal customers.

Encourage your organization to codify its AI ethics policy.

Define the parameters for acceptable AI use in your environment. Ensuring everyone is on the same page is a good way to avoid potential problems.

Establish “fairness metrics” that make sense for your particular company or industry.

Fairness is subjective. It is a value judgment that defines the version of the world we want to create. For example, your company may want to ensure that men and women have equal pricing policies.

Perhaps your company caters only to one gender, so gender equality is a non-issue. In some cases, your company might want to increase the number of customers from underrepresented groups. Whatever the goals, be as intentional as possible and always have a human check that your systems produce meaningful results.

Be transparent.

Never underestimate the importance of customer trust. Being open and disclosing where and how you use AI is the equivalent of a love letter to your customers. If you send an email written with AI, don’t hesitate to include a note at the bottom that says, “This email was written with the use of AI. See anything wrong? Let us know!” The same goes for your chatbots.

Don’t try to slip them under the radar to convince your site visitors that they’re speaking to actual humans. Instead, the chatbot should be programmed to inform the site visitor that it is an automated system. Tell them how to contact a human, if necessary. In this age of “everywhere AI” and deep fakes, it’s becoming difficult to distinguish humans from machines. Your transparent disclosure can become a brand asset.

Let humans have the final say.

Human oversight is critical when using AI in marketing. Remember that you are ultimately responsible for conducting your marketing activities reasonably and unbiasedly.

What’s more, you will always be able to understand emotional nuance better than an AI, so you are a better judge of tone and word choice. You are more aptly equipped to address your customers respectfully and appropriately, especially concerning gender, race, location, sexual preference, and socioeconomic status. Never undervalue human empathy — it’s your greatest value add. 

Did you enjoy this article? Please let us know. If not, let us know. Either way, leave a comment and let us know.