Artificial Intelligence and the Problem of Bias
When we think of artificial intelligence, we think of it as objective, impartial and unrelentingly logical. But we need to remember that, at one point, AI is programmed to learn by humans. To learn, machines are fed data sets and they can be full of historical or human bias.
Machine learning is something we need consider in the wider community, because as forms of AI increasingly thread into our day-to-day lives, if you’re not male, or white, there could be some problems.
A very easy example first. Take Pokémon Go. When Pokémon Go was released, users in New York found the gyms and PokeStops appearing more in predominantly whiteneighbourhoods.
Turns out Pokémon Go was using a crowdsourced dataset from a previous augmented reality game. The people who wrote the algorithms weren’t a diverse group and so their bias ended up in the game.
If diverse groups are required to help create unbiased products then it’s worrying when you consider how unwelcoming the tech industry is to women and people of colour.
Another example is LinkedIn. Women on LinkedIn (a business and employment-oriented service) found they weren’t seeing high-paying jobs as frequently as men. That’s because LinkedIn was selecting men to see the jobs.
Anu Tewary, chief data officer for Mint at Inuit, explained the problem to TechRepublic: “… it was biases that came in from the way the algorithms were written. The initial users of the product features were predominantly male for these high-paying jobs, and so it just ended up reinforcing some of the biases.”
Joy Buolamwini carried out a study of various AI-powered facial recognition software and found that they performed best with white, especially male, faces. When it came to women of colour, there were 34% more errors in recognition. Buolamwini found that when using examples of the darkest-skinned women, the face-detection systems could get their sex wrong close to half the time. This error rate was happening because when building the software, the engineers fed their algorithms primarily images of white males.
Buolamwini highlights that when you train your software with a biased data set, you end up with a biased result.
This is a link to an example where AI learned to associate “woman” with “kitchen” using research image collections.
Machine learning reinforces – it magnifies. If a photoset generally associates women with kitchens, software trained to study the association and the labels assigned will end up creating an even stronger association.
Just look how long it took Tay to turn from Microsoft’s upbeat twitter AI bot to foul-mouthed racist telling feminists to burn in hell. (Note: less than a day). This was the result of jokey-troll tweets but still it makes you question how will we ever teach AI using public data without supplying our public racism and inequality?
What we feed AI matters. Who programs AI matters.
These are low-level examples because AI is still in early stages. But as AI-based systems take on more complex tasks, as we embed more AI into our daily lives, we risk embedding sexism, racism and all our prejudices.
What if we were to use AI for diagnosing in healthcare? Machines can parse loads of information very quickly. But if we use current data about symptoms and treatment, we could end up with incorrect or dangerous analysis.
For example: women can experience heart attack symptoms differently to men – but may be misdiagnosed because the male symptoms are the “typical” ones.
Women with endometriosis can take years to get a correct diagnosis because there’s not enough information about the condition. Not to mention pregnant or menstruating women are often left out of medical trials.
This is evidence of biased data. Could feeding this skewed data to a learning machine just exacerbate the inequalities women experience when it comes to healthcare? I would say, yes, it could. We need to consider how we might accidentally contaminate new systems as we expand our thinking to utilise the benefits AI will provide us.
We like to imagine using artificial intelligence in our machines will help us be more logical and less prejudiced. But are we really freeing ourselves from bias? Or are we embedding it for future generations?
By: Tee Linden