One study published earlier this year analysed images on Google,Wikipedia and the Internet Movie Database (IMBD), specifically looking at what genders predominated when they searched for different occupations – such as "farmer", "chief executive officer" or "TV reporter". The findings were stark. Although women were underrepresented overall, gender stereotypes were strong. Categories like "plumber", "developer", "investment banker" and "heart surgeon" were far more likely to be male. "Housekeeper", "nurse practitioner", "cheerleader" and "ballet dancer" tended to be female.

The latest study, however, took this a step further. Rather than just showing the extent of gender bias in online imagery, the researchers tested whether exposure to these images had any impact on people's own biases. In the experiment, 423 US participants used Google to search for different occupations. Two groups searched by text, using either Google or Google News; another group used Google Images, instead. (A control group also used Google, but to search for categories unrelated to occupations, like "apple" and "guitar"). Then all participants were given an "implicit association test", which measures implicit biases.

The more biased images AI models themselves spit out, the more we see; the more we see, the more implicitly biased we become ourselves.