Consumer insights data can help combat bias in AI

Welcome to AI2, where I discuss trends in artificial intelligence and share my opinion on what they mean for the consumer insights and market research industry.

Have you ever heard of WEIRD data? Not unusual data, but WEIRD: Western, educated, industrialized, rich and democratic. It’s often used to refer to psychological findings that are based on research done specifically on western university students. The data isn’t representative of a global population, because it’s from a very non-representative sample of people.

There is a very valid concern that AIs are biased because they are not trained on representative data. They’re trained on internet data — much of which can be classified as WEIRD.?

I participated in an ESOMAR panel last year where we discussed using AI in Africa. As part of that conversation, we talked about this problem. If much of the data AI is trained on comes from other parts of the world, can it be relevant to the African population? How can anything produced in partnership with AI be useful to Africa or other non-western markets??

Negative sentiment from the internet is used to train AI

You may remember Tay, a Twitter bot from Microsoft in 2016 that had to be taken down in less than 24 hours after launch because it became incredibly racist and offensive. Designed to learn from the humans it interacted with, it started reflecting back the awful things some people online were saying to it.?

All of this negative commentary is used to train AIs because it all exists on the internet. This means that the AI has been largely trained on data of people talking about different marginalised groups rather than data from the people in those groups. And unfortunately, there are negative and inflammatory statements about those groups of people in that data.?

So the AIs don’t really understand what those people think or feel or what would appeal to them.?

There are certainly solutions that have been developed to address this inherent bias, but they can sometimes lead to less-than-ideal results. For example, Google Gemini’s efforts to correct bias initially produced inaccurate images of historical figures, such as rendering America’s founding fathers as non-white.?

Consumer insights data is part of the solution

I believe that consumer insights data is critical for combatting the bias inherent in AI. Consumer insights data tells us how people think and feel, directly from them. It is data from people in their own words, not data about them from others. And we researchers collect this data from all types of people, in all types of countries, on lots of different subjects, all the time!

The one concern in this solution is that a traditional market research study may only contain data from one or two people from a single demographic. But if you have data from multiple studies all in one place, you have a much richer data set that can train the AI to understand that group.

Then, you can work with the AI to design new products and ads that appeal to these different groups of consumers. You can run those ideas by synthetic respondents and get data that is representative of those populations.?

Overall, if you’re intentional about getting the data that helps you understand different groups and populations, you help your AI understand them too.?

Further reading

The Kindle edition of my book, The Consumer Insights Revolution, is currently $0.99 on Amazon. If you’re interested in establishing that centralised place for your consumer insights data to help minimise the bias in your AI efforts, this book is a good place to start.?