Joy Boulamwini's Gender Shades brings attention to racial and gender biases in certain commercially available facial recognition technologies, and although the biases in those programs have largely been addressed, many still exist and need to be dealt with. These biases are mainly due to the training datasets of the models, which include disproportionately white and male data.

One example was found in the image generation AI, Midjourney. When prompted with certain job titles, it was much more likely to generate men than women. It also favored younger people for certain non-specialized job titles.

Predictive policing tools, which are supposed to anticipate crime hot spots and allocate police patrols accordingly, have been shown to be more likely to flag black neighborhoods as potential high-crime areas. This reinforces existing biases and contributes to the feedback loop wherein people in heavily policed areas are arrested more, leading to the area being even more heavily policed, and so on. Many of these tools have started using victim reports to curtail this, but research has shown that, for various reasons, this data is also skewed. Without significantly more data, it is unlikely that these tools will become unbiased to the point that they can be used without fear of making things worse.

In all, though there is still a long way to go, significant work has been done toward eliminating bias in AI technologies across the field.