Retailers are adopting artificial intelligence almost twice as much as other industries. Once out of reach for everyone but the largest retailers, AI is now readily available for ecommerce merchants and powering everything from attracting customers to understanding their needs and delivering frictionless experiences.
As organizations gobble up data and AI becomes more mainstream, the retail industry is asking new questions about bias in AI results. There’s no shortage of AI bias stumbles, including some from the largest, most advanced and best resourced companies. Two recent examples: Twitter’s simple AI image-cropping tool showed racial bias; and an Amazon AI-recruiting program showed bias against women.
In ecommerce, AI-driven personalization is particularly ripe for bias with geography, past behavior and social network activity as key data sources. Research done by the Wharton School of Business illustrated that targeted ads and dynamic pricing can be unfair and perpetuate bias. Images remain a significant avenue for bias as well. A Carnegie Mellon study, for example, found clear sexism in AI-generated female images.
Let’s get to the root of what causes AI bias and how to best tackle it.
Understanding the Root Causes of AI Bias
AI bias is when a machine generates information or takes action that may be unfair, discriminatory or harmful to individuals, the business and ultimately society. Humans programming AI algorithms and selecting data sources will always insert their bias, whether intentionally or not. With technology organizations concentrated heavily in particular geographies and predominantly staffed by men, combined with a dearth of women in the field of AI, it’s easy to see how it happens.
There’s a lot of attention and finger pointing at AI algorithms, which are the formulas for decisions. But algorithms are a tool, and the tools rely on the data input. Machine learning has an inherent challenge in combating data bias due to the massive volume and diversity of data it can use.
To really solve this problem, we need to look just as closely at the data entering algorithms as that leaving it. Data is not free from bias when it enters the algorithm — humans decide what to include or exclude. And those sources likely already have inherent biases based on culture and human decision-makers with little or no awareness of potential biases. In fact, just 18% of companies have a clear data sourcing strategy in place for AI initiatives, never mind a focus on bias in that data.
Remedying Bias in AI
Although we’re at the nascent stages of understanding and mitigating bias in AI across the technology industry, here’s where we need to concentrate efforts to deliver unbiased outcomes:
- Human awareness and interaction — Developers, users and executives have to watch for and flag bias. Humans need to intervene, from the onset of data selection and throughout the whole cycle of development and deployment. What’s the data input? What’s the potential bias? How does that impact decisioning? And it’s critical for the industry to openly share both problems and resolutions to move everyone forward.
- More diversity in developers and technology companies — Everyone has some level of bias. Since ultimately humans must intervene and manage these algorithms, then it follows that more diversity in the people will bring more diverse perspectives and ultimately reduce the weight of biases from a particular group.
- Identifying and understanding data sources and information bubbles — Social networks are a huge data source for AI, especially in ecommerce. They’re also infamous for putting users into information bubbles, where algorithms feed what users want to hear based on past behavior. Merchants use social networks (as they should) to find and market to customers, but left unchecked, they are simply reinforcing biased behavior. If you’re going to mine these networks, you need to understand your role in perpetuating bias, and also call on the social giants to do their part.
- Deploy AI to augment (not excuse) human effort — With AI becoming more mainstream, there’s a lot of temptation to turn it loose and gain all the benefits. But ultimately, you’re limiting yourself. The answer is to use AI in the places it helps the most, govern it and free up humans to do valuable work. In ecommerce and retail as a whole, that often creates offers, new products and other activities that machines will never replace. You cannot simply outsource to AI. It will free your time. It will give you an advantage. But it cannot run and manage itself.
- AI to better AI — Google, one of the front-runners in AI, has no shortage of controversy around some of its efforts. But it’s also doing work to actively listen to cultural sensitivities and address them even in the simplest of ways. Last year, Google released AI to actually address gender bias in Translate. AI may actually be useful in flagging bias, but humans still need to get to the root cause and fix bias before it enters the system. Microsoft has announced efforts along these lines, with tools to help developers catch bias in algorithms.
- Continuous governance efforts — Data is a huge source of bias, and it’s skyrocketing in volumes and characteristics. Ecommerce thrives on recent — even instant — data for AI decisioning. Ninety percent of the world’s data is less than two years old. Facebook alone generated four petabytes of data per day in 2020. And that social data is riddled with bias. You will need governance strategies that deal specifically with bias and adapt them as fast as the data — the greatest source of bias — changes.
What can’t you do? Blame the machine. Machine data is ultimately human data. Algorithms are instructions from humans. The machines are ultimately nothing more than a reflection of ourselves.
Alexandra Faynburd is the CTO of Fast Simon, which develops AI-powered shopping optimizations merchants. She invents and produces algorithms for cloud-based ecommerce optimization. Prior to her current position, she was a software engineer at Microsoft. Faynburd earned her BASc and master’s degrees in computer science from Technion.