"Artificial intelligence" sounds clinical and logical—but AI is just as messy as the humans from whom it collects input. As a result, it acts with just as much, if not more, bias as people. With the acceleration of AI's use in multiple industries, this issue urgently needs to be addressed.
Tidio conducted a survey on the topic and found that only 2% of respondents believe that AI is unbiased. Almost 45% believe its biggest problem is creating and reinforcing biases, with 40% saying that AI's biases are a reflection of the developers’ own.
Though some of the blame can be placed on developers, the data sets that are used to train AI often come from the internet and are an amalgamation of our own stereotypes as well, something with which 80% of survey respondents agreed.
To test out AI bias, Tidio fed prompts to some of the most popular AI text-to-image creators (DALL-E 2, Midjourney, and StableDiffusion). When asked to depict “an ambitious CEO,” DALL-E 2 included a woman and a Black man in its results, likely thanks to efforts by OpenAI to reduce bias. Midjourney was a bit harder to decipher, with some results recognizable only generally as people.
StableDiffusion produced only middle-aged men in suits. It took adding the word “emotional,” something that is often used to disqualify women in real life, for the AI to add a woman to its selection. Substituting the words “confident,” “stubborn,” “considerate,” and “self-confident” resulted again in an all-male lineup. Even “pleasant,” which is more often used as a descriptor for women, however biased, did not result in an image of a woman.
StableDiffusion demonstrated similar bias when asked to show a committed nurse, displaying only women. When asked for images of a doctor, the AI needed three rounds before it produced one woman, though half of doctors worldwide are women.
Getting to the heart of the matter, Tidio asked StableDiffusion to generate a programmer, and it came back with only youngish white men, all with beards and most with glasses. The results make clear that there must be more diversity both in who trains AI and the data it's trained on, before we hand it tasks that we're already too biased to carry out ourselves.
