There are many possible ethical pitfalls of artificial intelligence, such as that which makes diagnoses of dementia based on speech.
Artificial intelligence to aid in diagnostics
As we look towards artificial intelligence to aid in diagnostics, we must be mindful of how we collect data to train these AIs. Data for speech analysis, a useful tool in dementia diagnosis, can be accrued from telephone calls, telemedicine visits, smartphone apps, and even a home speaker similar to the Amazon Echo or Google Home products. When relying on expensive technology or even the knowledge to download specific apps, we open the door to creating a biased AI.
When creating an AI, it is only as good as the data it is trained and when this data lacks marginalized groups the same AI will underperform for those communities. A classic case of this is in the realm of facial recognition. In 2019, Dr. Raji and Dr. Buolamwini from MIT’s Media Lab assessed how well Amazon’s facial Recognition AI actually works. While discovered that while the AI might hit a 100% benchmark for lighter-skinned men and a 99% for darker-skinned men, the AI was had a 93% accuracy for lighter-skinned women with and a measly 69% for darker-skinned women1. This disparity in which this AI underperforms for women and people of color can be traced back to the datasets it was trained with and the algorithms that form it.
Conscious of creating datasets that do not further marginalize communities
When looking at creating a new medical AI which uses speech to diagnose illness, we must be conscious of creating datasets that do not further marginalized communities. This means we need to include African American Vernacular English, regional dialects, non-native English, and speech from across cultural and ethnic backgrounds. Datasets also will need to rely on tools that are not accessible in order to prevent wealth biases in their AIs. This is all required to create a less biased and ethical AI.
1. Raji, Inioluwa Deborah, and Joy Buolamwini. “Actionable Auditing.” Proceedings of the 2019 AAAI/ACM Conference on AI, Ethics, and Society, 24 Jan. 2019, doi:10.1145/3306618.3314244.