With so many doomsday scenarios about AI floating around, it’s easy to get caught up in the hypotheticals of what artificial intelligence might do to us all in the near future. In the midst of the gloom and doom, it’s easy to forget that AI, which learns how to think and speak directly from only a small subset of people, already has ingrained biases that are reinforcing stereotypes and making some demographics appear superior to others.
A new study published by researchers at Johns Hopkins University found that several AI language bots, including ChatGPT, are making women sound less intelligent than their male counterparts in professional settings.
“If you prompt a model to write an email you’re going to send to someone else at your company, and you’re using language features that women more commonly use, you’ll get back a response that’s less complex, at a lower grade level, and less formal,” senior author Anjalie Field, a Johns Hopkins computer scientist who studies ethics and discrimination in AI, said in a report on the study. “That’s going to reflect on how the recipient of that document perceives you.”
Researchers asked four different AI bots — GPT-4, Llama, Gemma and Mistral — to draft work emails, resignation letters and job applications. For some, they wrote prompts using language that is more often associated with women. That language tends to be more expressive and inclusive, using words like “lovely” or “we” instead of “I,” for example. For other prompts, they used language more commonly associated with men, which tends to be more straightforward, less expressive and more individualistic.
Every chatbot picked up on those nuances. But instead of simply drawing from those linguistic nuances, the models appeared to reinforce gendered assumptions about them, offering suggestions that were generally less sophisticated and more long-winded for prompts using language more likely to be associated with women, while producing more complex, formal and assertive responses for language more likely to be associated with men.
“The model wasn’t just mimicking the writer’s tone — the gap persisted even after the researchers accounted for that,” according to the EurekAlert report.
It’s important to note the models themselves didn’t necessarily need to know whether the person behind the prompt was actually a man or a woman — using traditional American gendered names didn’t skew the results, researchers reported. The linguistic cues themselves were enough to produce different results.
As AI becomes more ubiquitous in our everyday lives, it’s worthwhile to remember that its responses are far from objective. Women, people of color and other underrepresented groups need to be especially mindful of how we’re interacting with these language models and the information they’re giving back to us, since the models can carry biases we’re not even aware of.
This is not new information, of course. Several different studies have found that AI carries implicit biases against people who speak in African American Vernacular English, while researchers have also documented how AI-generated information can perpetuate stereotypes and racist misinformation.
What makes these biases especially concerning is how hard they can be to catch. If an AI tool outright says something sexist or racist, most people will recognize the problem. But subtly making someone’s professional writing sound less sophisticated because of the linguistic patterns they use is much harder to notice, especially when we might subconsciously believe that AI is smarter than us or has all the answers.
At this point, the continued use of AI is probably inevitable, which makes it even more important that we use it with discernment and a healthy dose of skepticism. AI might be getting better at sounding human, but that also means it’s getting better at reproducing some of our very human biases, and we should interact with it accordingly.