Why AI without humanity is dangerous for diversity

So how can we imbue AI with a sense of humanity? It starts with making sure that our machines are not just learning to be intelligent, but doing it from a position of emotional intelligence (EI). To ensure that EI underpins the development of AI, it’s vital that we have the big ethical discussions about how it should (and should not) be used now, while the technology is still in its relative infancy and, even more importantly, before AI companies become too large and powerful to control—a situation we are already experiencing with the current tech giants because their businesses were poorly understood by the public and governments until it was too late.

One of the trickiest elements of AI is that, often, we don’t really know how artificial decision-making works. It’s a black box. We create AI by training it with data sets, but when it gives us a result, we don’t necessarily know how or why it came to that result. We can’t question its logic, we can’t probe its thinking, the result just is.

(There are, however, emerging technologies such as LIME that are helping us understand the decision trees of AI, and which may help to establish trust in the process. Hopefully these types of technologies and trust-building measures will be part of any discussions and resultant decisions around AI regulation.)

A crucial aspect of this is the diversity of the data. If the AI is involved in recreating human decision-making, or identifying humans, or if the decisions it makes directly affect our lives, then we want to make sure that it has been trained appropriately. This means using diverse data sets that span the full spectrum of humanity—from different physical characteristics and cultural behaviors to belief systems and thought processes. If machine learning doesn’t factor in diversity, then the AI we create will just reinforce—or, worse still, amplify—the biases we currently have in our society.

This problem of bias in AI is real and needs to be addressed at the most fundamental level. With more and more uses for AI being created every day, the risk that even seemingly benign applications may be infused with biases that could have harmful real world consequences is increasing exponentially. And once the companies behind the AI grow large, and the usage habitual, then we are in real danger of further institutionalizing our current biases if we don’t take adequate preventative measures.

And there is a further complication due to the demographics of AI researchers. A 2018 survey by WIRED magazine found that only 12% of AI researchers are women, indicating that data related to women, or women’s perspectives, wouldn’t be sufficiently represented in the training of the AI. Similarly, minorities are shamefully underrepresented. In short, the people creating the AI don’t adequately mirror our societies, and therefore won’t be fully equipped to train the AI on perspectives different to their own. If we don’t address this at a societal level, and by proxy at the government level, we will continue to deal with problems of AI discrimination, such as Amazon’s now-scrapped AI recruiting tool or flawed attempts to predict recidivism in US courts.

As AI becomes more complex and more prevalent within the infrastructure of our lives, there is the real risk that discrimination, even if it is through negligence, becomes more pernicious and likely more nuanced in its injurious effects.

The common objection to government involvement, and potential regulation, is that the cost will be too high because it will stifle innovation and competition. It will thus be harder for new companies and new services to emerge, and only the biggest and most well-resourced will be able to navigate the regulation, which will in effect institutionalize a different form of homogeneity. To avoid these common refrains, it is critical that discussions on how to approach AI’s development at the societal level are begun early and openly, taking in the views of all the key stakeholders.