AI’s Not-So-Hidden Gender Gap
A new study discovers that AI writes less polished workplace correspondence when the prompt sounds like it comes from a woman.
The manager prompted AI to write an email for a leader, then send it. She checked it off her to-do list. Phew, that was easy.
Not so fast. A new study out of Johns Hopkins suggests that chatbots provide less sophisticated workplace correspondence when prompts used language demonstrated to be associated with women: Overall, responses written using male-associated prompts were deemed more competent 69% of the time and smarter 68% of the time. Overall, outputs from female-associated prompts were less complex and less formal, and used language at a lower grade level. Experts are alarmed but not surprised. “Any AI that is built on a preexisting system is going to perpetuate the biases of that system,” says Jenna Young, senior partner and global solution architect at Korn Ferry.
The study is far from the first to reveal disparities AI has propagated in the workplace, whether in its research results or its hiring and pay suggestions. But it is the latest warning of how, in ways both large and small, the technology continues to impact careers.
In this case, the disparity emerged because AI detects subtle gendered differences in language: Its bias is triggered by the wording of a prompt. The researchers emphasize this isn’t something a user can simply fix, writing that “because these patterns are culturally embedded and outside conscious control, users cannot easily avoid them through strategic self-presentation.” The bias appeared across all four AI models tested in the study, three of which are widely used in the business world. While concerning for women, the lesson is pertinent to everyone. “There are biases built in, and we know that to be true,” says Karena Man, senior client partner in the Technology and Digital practices at Korn Ferry. “That means that you cannot rely on chatbots to represent your intelligence and who you are.”
The researchers caution against simply telling women to write prompts differently—this “places an unfair burden on users,” they note. The long-term fix will have to come from how models are built. In the meantime, experts say that organizations will need to instill AI judgment in their workforce. “The results point to the need for increased capability in judgement of AI outputs,” says Janet Mertens, managing director for research at Korn Ferry’s AMS business. She encourages employers to invest in AI-literacy and -fluency training.
For now, experts say that AI best practices are essential. “Taking the output of AI as gospel and hitting ‘send’ is simply not the right way to use AI,” says Bryan Ackermann, head of AI strategy and transformation at Korn Ferry. “The AI may well say something wrong, stupid, poorly formed, or less considered.”
One way to potentially dodge this problem is to train AI on your particular writing voice. Large language models learn from examples, so rather than describing your style, provide bots with examples of your professional writing in different formats—emails, reports, memos—noting which are strong examples and which are less ideal. “You need to take the time to train the bot to reflect your branding, messaging, and what makes you memorable,” says Man.
While experts hope that current and future large language models will be adjusted to avoid these gendered language biases, they say that the solution in the short term is simple. “There are a million reasons to say this,” says Ackermann. “Check the work before it goes out.”
Learn more about Korn Ferry’s AI in the Workplace capabilities.
