The BBC reported, recently, that several major supermarket chains are trialling the use of automated technology to scan a customer’s face and verify their age, when they are buying alcohol. The reason for using this technology for age verification (as opposed to relying on cashiers to do so, in line with legal requirements) is, according to the article, because:
“Waiting for age approval at the self-checkout is sometimes frustrating for shoppers,” Robin Tombs, chief executive of Yoti, the company providing the technology, said. “Our age-verification solutions are helping retailers like Asda meet the requirements of regulators worldwide and keep pace with consumer demands for fast and convenient services, while preserving people’s privacy.”
I.e., as is often the case, this automated age-verification technology is being deployed in the name of improved customer service. This claim prompted me to search whether there is any published evidence suggesting that customers would, indeed, prefer using this specific type of age-verification technology over staff. And, funny enough, I found a masters’ dissertation which actually tested an age verification system developed by Yoti (possibly the same mentioned in the BBC article?).
The title of the dissertation is “The feasibility of face scanning as an age verification tool from a technical- and UX-perspective” and was authored by Emma Lindmark. And, to be 100% clear, I think that it is a solid and very interesting dissertation!
The study found that the system accurately guessed the age span of the customer on around 33-36% of the occasions (29/89 and 32/89). Moreover, the system would underestimate the age in around 56-63% of the times, and overestimate it in the remaining 4-8%.
On average, this high rate of false positive error (i.e, under-estimating the customer’s age) did not seem to exasperate the participants in this research – indeed, they reported feeling positively about such an event:
However, this is one of those cases where focusing on the average can be very misleading.
First, a false positive error (i.e., under-estimating the customer’s age) would have different impacts for different customers. For instance, it wouldn’t make much difference for those aged, say, 40. However, for those customers just above the age limit it would result in them having to prove their age to the cashier before they could proceed with their purchase.
Second, performance is likely to vary by gender and ethnic markers (this is not disclosed in the dissertation mentioned – only performance by age group). The paper “Automated classification of demographics from face images: A tutorial and validation”, authored by Bastian Jaeger, Willem W. A. Sleegers and Anthony M. Evans found that automated age classification systems are more accurate for male than female faces; and for Black and White faces than for Hispanic or Asian ones.
For instance, it is possible that the system might over-estimate the age of faces with beards, which could result in more women than men being wrongly flagged for manual age checks; or that women with short hair and no make-up end up being classified as male and, in the absence of facial hair and other markers of age, be deemed as under-age males. It is also possible that those Hispanic and Asian customers that are wrongly flagged for manual age checks are also those already feeling discriminated against, or those who can’t really afford the extra time for additional checks.
To be clear, I am not criticising this study (indeed, it is a really interesting masters’ dissertation!); and I am not suggesting that Yoti has not conducted additional research about this system (indeed, I suspect that they did). Rather, my thoughts here are about research underpinning the deployment of customer interface technology, in general.
Technology can help some customers navigate daily discrimination and stereotyping but, if it is not carefully designed with a diverse customer base in mind, it can hurt more than help disadvantaged customers. All too often, analysts fail to break down the analysis by group, or even to collect data about those groups most negatively impacted by the technology. If you are involved in the deployment of customer interface technology, you need to conduct solid – and extensive – research focusing on those customers most likely to be negatively impacted by the technology!
Now that my curiosity has been sparked, I would really like to know: Are you aware of any research on customer satisfaction with this type of technology which specifically looks at the gender and/or ethnic angles?