In today’s marketing and communications landscape, transparency rules. With so many different systems digging deep into consumer data to find individuals’ explicit, behavioral, and derived preferences, it’s critical that marketers are able to explain their systems’ conclusions. Smart marketers who are confident that their recommendations or predictions are accurate, and can explain how they got there, will use this capability to tell the potential customers: I think you’ll like this message because of X, Y, or Z.
In addition to being able to communicate honestly and transparently with consumers, something that I think should be more central to modern communication, it’s also becoming a regulatory and ethical necessity to be able to explain why a particular group is getting a specific message or offer (not just what the offer is).
Let’s look at the emerging field of explainable artificial intelligence (XAI) and how we can apply it to address some of these issues.
AI, EXPLAIN THYSELF
All eyes are on AI as the hot new tool, and rightfully so. AI is able to find patterns in noisy data, pick out unique targets for campaigns, or even design campaigns that do not appear obvious to marketers using traditional campaign strategies.
You can’t assume that using AI to find patterns to optimize your marketing actions is always a perfect science. Why these recommendations appear is as important as the recommendation itself. Known to have biases, AI training sets can result in recommendations that can have negative or unintended effects on your customers and/or prospects, including reinforcing racial or gender biases, potentially withholding favorable terms from segments of customers, or worse. You have a duty to ensure that you utilize AI in a manner that protects your company’s most important asset—your brand and its equity.
And even if an AI model is completely ethical—meaning it is built responsibly, based on accurate data, and designed to be free of bias — when it is deployed, the impact on a certain group of people may become more obvious than others. This may be an unanticipated consequence. As a result, additional testing, monitoring, and a proactive feedback loop must be provided for model re-training.
That’s why, for every recommendation you get from AI, you need it to be explained if it’s going to impact customers, and your communication and engagement with them.With the field of AI moving so fast, however, the science of building XAI lags the creation of AI systems themselves.
There is no universal, one-size-fits-all AI bias testing tool, unfortunately. Some projects, such as the What-If Tool from Google, provide libraries that AI teams can use to explore a model’s sensitivity to inputs. If you don’t have data ethicists on hand, a thorough third-party review of models and outcomes is also a valuable check against biased AI results.
Furthermore, a big part of moving towards XAI in a marketing organization is keeping a tight watch on the data you are feeding into your systems—because all AI does, after all, is draw conclusions from the data you give it. Exercise good data lineage practices and you can move to XAI more readily.
DATA AT THE CORE
Data lineage is about knowing your data’s history and quality—making sure it comes from reliable and verifiable sources. The goal is to enable XAI tools to illustrate which data goes into their outputs—turning the “black box” of AI into a glass box.
In addition to being able to justify your AI-based conclusions to customers, a growing number of regulations also demand explainable AI. The newest is the E.U.’s AI Act, which lays down requirements for AI used in the EU. The act is centered around classifying AI risk—focusing on categories like people’s health, safety, their fundamental rights, and the environment. For some high-risk applications, such as those that impact voters or the users of social media platforms, the act mandates that AI decision-making processes should be transparent and traceable (explainable).
GETTING STARTED
For marketers preparing their business to deliver explainable results from AI to customers, there are two departments with which you should be communicating.
First, make sure the teams responsible for AI projects—often under the CTO’s or Data and Analytics office, but possibly elsewhere—are on top of XAI science. They should be following a crawl/walk/run/sprint approach on very well-defined scopes using documented tools—not boiling the ocean and trying to invent XAI from scratch. Starting small and proving value will build trust for subsequent phases. With a well-defined scope and the use of well-known and widely used algorithms, it is possible to achieve explainability.
Second, your company will benefit from clear and published data quality policies. This is the purview of the chief data officer, if you have one. AI amplifies the benefits of having good (reliable, consistent) data. AI output will also exacerbate data quality problems if you don’t have clean data.
There are many data quality tools on the market. At our company, we use proven, industry-standard cleansing and transformation methods on data sets before running AI processes on them.
With a good handle on explainability and high-quality data, you can address current and emerging regulations designed to keep your company compliant, ethical, and fair, and customers safe from bias and exploitation. And you can also reveal to your users why they’re getting the communications and offers from you that they are—and give them a chance to participate and improve how you’re connecting with them.