There’s a sort of fascinating study out there, being reported on now, that begs a number of questions about automation in the workforce.

Anothony Cuthbertson at The Independent reports that a team at the University of Limerick created a trial project to simulate the remuneration of digital AI workers, and found that those presenting as male “earned” more than those with female characteristics, which you might presume to be mostly voice, and avatar. After all, AI doesn’t have bodies. Agents don’t have physical characteristics. But they do have voices. And they can be given names. That’s part of what’s so abundantly weird about all of this. None of us humans are ready for colleagues and co-workers who are “there” and personified, but not, as you’d say, born of woman.

Anyway, Cuthbertson’s reporting contends that the “male” Ais get “paid more” than the “women.”

What does it mean to have this kind of glass ceiling? Especially when companies don’t typically “pay” a digital workforce at all?

In a key sense, AI entities, whether roboticized or not, are like animals. They can do “work,” but they can’t really “get paid” in any meaningful sense. Money doesn’t mean anything to them, albeit in different ways: while the animals lack the comprehension of money as a concept, the AI entities lack the needs to be ameliorated with the money, although surely, as smart “people,” they can find ways to spend the money.

It turns out, though, that outside of this study, “paying” AI agents typically means allocating them resources: partitioned data storage, a virtual machine or container, an operating system, a browser, a stateful record. Anything like that can be “pay” for an agent, according to the ROI that the company expects.

But in the study, the researchers doled out actual funds. Here’s how the reporting explains it:

“The study involved 189 human participants working alongside a male-presenting AI agent called Johan and a female-presenting agent called Johanna. Not only was Johan paid more, he was also perceived as more human-like than Johanna.”

A research paper around the study goes into more detail about the methodology, apparently involving the allocation of Swiss Francs by human users, to agents:

“Participants rated each assistant in terms of perceived human-likeness, trust, and credit, and allocated monetary rewards of up to USD 4 per assistant as a behavioral measure,” the research team writes. “We further conducted semi-structured interviews with 34 participants to gain a deeper understanding of their ratings, perceptions, and articulated preferences for AI coworker design. Our findings show that participant evaluations and trust of the same underlying AI system varied significantly by its embodied representation. Human-like assistants received higher trust and credit ratings as well as higher monetary rewards than the robot assistant.”

That makes things a lot clearer for anyone who would read the Independent coverage and wonder, they got paid how, exactly?

It also turns out some humans wanted these gendered, humanized agents, and some didn’t. The paper clarifies:

“Interview data showed divergent preferences for AI design: while some participants advocated for strictly non-human-like representations, others expressed a strong preference for assistants that were as human-like as possible.”

Here’s a little bit of cost-benefit analysis, where the writers concede that a gender pay gap is one of several side effects of having your AI be gender-specific in the first place.

“Increases in human-likeness may foster trust and engagement, but they also risk introducing discomfort, miscalibrated expectations, and potentially even activating gendered stereotypes.”

In other words, creating that glass ceiling. This is a lot of what experts talk about when they delve into bias in system data: where AI systems can magnify existing social problems that were in the human world first. It also approaches the overall idea of “alignment,” where we ask: what do we really want out of our AI?

“As AI agents become more common in the workplace, we need to think carefully about the characteristics we give them and the behaviors those choices may encourage,” said Dr. Mary Hausfeld, a co-author of the study, as quoted by Cuthbertson. “We don’t want to inadvertently reproduce existing inequalities in a new technological setting.”

Of course we don’t. The reality, though, is that this stuff can be tricky. Just take a look at this Reddit thread, where various posters argue about the actual causes of the gender pay gap among humans. Redditor SingleMaltMouthwash has this to say about the skewing of numbers to hide a gender disparity:

“We exclude … these highly gender-specific effects upon salary, so that we can pretend gender has less effect upon compensation than is perfectly obvious. To put it another way, studies referencing the ‘controlled gender pay gap’ measure the mean salary for men and women with the same job and qualifications without pointing out that women are still frequently denied those jobs and the opportunity to gain those qualifications.”

It seems further study of the male and female AI personas will be needed. And there’s a lot more that has to be done, too, as we approach the Singularity. Stay tuned.