AI Don’t See Color

AI Don’t Work

Examples of biased AI products brought to market are increasing alongside the adoption of artificial intelligence and machine learning (AI Machines) in products. Some examples include hand washing stations not recognizing darker skin, facial recognition software not recognizing darker skin, and dermatology healthcare products… you guessed it, not recognizing darker skin. Puzzlingly, under further investigation of how these broken products came to market, the correction does not rely on highly technical fixes that only PhD’s in data science could catch. Instead, the correction boils down to the makers not testing on enough, if any, people with darker skin. Simple sample bias. The samples they used to test and build the product did not include a holistic sample of the product’s intended use; all skin colors [read darker skin colors]. AI Machine developers planned to use a product on every skin tone, but mainly tested on lighter skin tones. How could such a simple mistake be made by some of the brightest minds in AI and data science and at such large scales?

I wonder, whether our socialization to adopt color blindness contributes to expert scientists’ blind spots for not seeing color; even when their products are supposed to. 

AI Don’t See Color

There was and still is a push to convince ourselves that we are all the same; but we are not and that’s okay. What is not okay, is using our differences to exclude, harm, and shame. The idea of colorblindness became encoded into modern society to address the issues of exclusion, harm, and shame. At the end of the 1990s, colorblindness grew as the dominant response to accusations of prejudice. The speaker would often sidestep accountability for prejudiced behavior by claiming to not see color at all. This way, the actor’s decisions and actions were not based on race and thus preventing problematic behavior or systems from being labeled as racist. You can’t make a decision based on color if you can’t see color was the assumption.

But to be clear, these claims were made by people who absolutely could “see color” metaphorically and likely physically. These false claims of colorblindness had the effect of silencing legitimate concerns about the effects of prejudice, ignorance, and systemic blind spots to race. Colorblindness created a world that falsely assumed raced based decisions, actions, and systems were not possible. While baseless, these assumptions function today as a collective bug in our society’s psyche that auto-incorrects to pretending not to see or be influenced by factors that exist in reality. We live in a world where people are encouraged to pretend they don’t see skin color; even when they do see color and even when skin color is relevant.

AI Don’t See Color But My Product Does

If an engineer is designing a tool that uses pigment to make decisions, and that engineer has been encoded with ignoring the differences in pigmentation; what happens to the technology? (See the examples in the first paragraph of this post.) To find the solution, teams need not look at the code, algorithm, or model, they simply must review the monolithic data set to which they trained their machine. Our country’s insistence on avoiding accountability for racial prejudice has contributed to our brightest minds producing broken handwash stations and false arrests. It is basic data science that broad data pools usually improve the accuracy of the proposed technology if the intended use is broad. However, if a scientist has been socialized to pretend to not see skin color, the scientist could be more likely to not catch the monolithic sample data pool. She will likely be proud of the amount of data points, and not see the lack of variance within the data points. This leads to the tool being marketed for use to the general public, instead of limiting market use to the fair-skinned public on which the technology was tested.

AI Opener

Technology would benefit from us moving away from the pretend world of the metaphorically colorblind and moving toward the real world, which involves seeing differences and learning how to productively and ethically engage with those differences. This framework creates a sense of urgency for companies to implement robust diversity, equity, and inclusion programs to de-bug colorblind philosophy out of their employees, products and services. This would ensure the products they make and the services they provide work properly for everyone. The issue was never our ability to see people’s differences, it was our inability to stop harming people based on those differences. We can only solve that problem by living in reality and appreciating this world for its true and many colors. If we are going to create for the world, we must see the world.