How AI can fix bias in data—IBM, WPP and Delta are creating a roadmap for algorithms

The marketing world has been rife with panels about bias in recent years. IBM's panel in Cannes on Monday featured an unusual twist, laying out a roadmap to systematically do something about bias by fixing artificial intelligence algorithms.

IBM joined top executives from WPP, its Mindshare and Ogilvy agencies, Delta Air Lines and the Ad Council to talk about an effort around IBM’s Advertising Toolkit for AI Fairness 360, an open-source solution with 75 “fairness metrics” and 13 algorithms that can help identify and mitigate biases in data sets. It includes a playbook and sample code that can be incorporated into AI that drives media planning, programmatic buying or creative optimization.

While it’s open source, WPP shops have been the most active users so far. Delta Air Lines and its creative agency Wieden+Kennedy have been among other early users of the toolkit, in an effort also backed by the 4A’s and Interactive Advertising Bureau.

IBM has used bias mitigation tools already in many industries, including financial services, human resources and insurance, and those tools have worked to help reduce AI biases in those industries, said Bob Lord, IBM senior VP of The Weather Company and alliances. So the company commissioned a study applying the tools in the advertising industry last year.

"Lo and behold, what did we find? There is bias in advertising,” Lord said. “I’m not here to condemn anyone. What I am here to do is to offer a solution.”

While making bias mitigation algorithms available open source has its advantages, he acknowledged after the session that it goes against the grain of a marketing industry where marketers and agencies tend to think of their AI tools as being proprietary.

Case studies are key

Key to overcoming resistance to bias mitigation tools and strategies is building and publicizing a series of case studies, said Adam Gerhart, global CEO of Mindshare, ideally ones that cover both media and creative applications. He said WPP agencies and IBM hope to have some to share by CES in January.

“We are trying to boldly pursue equity, and this is a way that marketing can put that strategy into action,” said Emmakate Young, managing director of marketing at Delta Air Lines. “We’ve been really intentional about our visuals and who we see in our advertisements, and this really helps us fit the other side of the equation—to make sure it’s reaching all of the appropriate audiences.”

The Ad Council used IBM’s toolkit to target advertising in the “It’s Up to You” campaign to overcome COVID-19 vaccine hesitancy, said Elise James-DeCruise, the group’s chief equity officer.

IBM and the Ad Council found that leaning on programmatic buying without any intervention from the bias mitigation tools was driving more impressions toward higher-income audiences that responded positively to the campaign but were already convinced they should be vaccinated anyway, Lord said. Using the tools helped the Ad Council redirect spending, particularly to people of color and others who were more resistant to the message, but were also the people the Ad Council was trying to persuade.

In a similar way, Mindshare has found clients’ “propensity models,” based on their first-party data showing where brand sales are coming from, can also tend to channel more spending toward the already converted in a way that narrows rather than broadens a brand’s user base.

“You end up perpetuating a situation,” Gerhart said. “So in this case, the CPG product tended to skew more female than it did male and it was a household product. And what we found was sort of reinforcing gender stereotypes and the role of gender within a family setting. When you play that out, and you start doubling down on that, you can very quickly end up in a space where you're alienating huge portions of potential opportunity.”

Programmatic faults

“I think the advertising market has a really important role in society, and I think it can play a good role,” said WPP CEO Mark Read. Using IBM’s tools can also play a role in business success, he said. “If our clients are missing out on things we don’t want to miss out on, we have to really understand and build that into the way we think about marketing and correct for it.”

Later, in questioning by Lord, Read showed a remarkably dim view of the ultimate impact of programmatic media buying on the industry for the CEO of a holding company whose Xaxis business was one of the pioneers in the field.

“Sometimes I wonder how good the programmatic was,” Read said. “We spend tens of millions of dollars, hundreds of millions, optimizing on a pinhead, down the funnel, and all this other stuff doesn’t get nearly enough attention. It’s like being really good at assessing the value of Argentinian bonds when actually you should be buying something different.”

But, applying IBM’s AI tools to evaluate creative has the potential of making the ad industry follow through on good intentions, as Devika Bulchandani, global president of Ogilvy, put it.

“A lot of what we’re trying to do as an industry is taking on bias,” she said. “What happens though is we do it at what I call the big idea level, and then it just drops off. You go into the trillion-dollar world of digital advertising programmatic, and it just breaks down.”

Woke accusers 'full of shit'

Making a short guest appearance from the audience, Edelman CEO Richard Edelman expressed his hope, colorfully, that IBM’s bias mitigation tools also can shore up marketers’ resolve against critics of them taking stands on social issues. It’s something he believes AI will ultimately show works for brands, because years of Edelman’s own Trust Barometer research shows people, particularly Gen Z, reward brands that take stands.

“The woke accusers, who are the blowback against brands that take positions, are full of shit,” Edelman said. “And we need to call them out and say, ‘you’re wrong.’ Here’s the data. We know that 85% of Gen Z buy on the basis of belief.”