Name: Artificial Intelligence Has A Bias Problem
Uploaded: 2018-06-25
Duration: 5 min 8 s
Description: These companies are working to find a solution.

There's this excitement and this hype around artificial intelligence, which is leading a lot of investment into this space. It's been called the new electricity. It's one of the drivers of the Fourth Industrial Revolution. It is going to completely revolutionize a lot of aspects of our lives. It is transforming everything that we're doing. We're relying on artificial intelligence to make more and more decisions in our daily lives including deciding who gets access to things like a job or a loan or parole. And so if we see Racial and gender biases for example creeping into these systems that has life altering consequences. [ MUSIC ] What people are worried about is that these machine learning models will end up giving preferences to some groups over others in a way that's not justifiable. We're seeing existing societal biases, whether it's gender bias or racial bias Showing up in a lot of artificial intelligence systems, like online translators or in the criminal justice system algorithms that are being used to help judges make parole decisions. A recent study just came out, that was looking at some of the top facial recognition systems, and it found racial bias and gender bias in how accurate the system was in identifying Faces. When you look at overall accuracy it's pretty good, upwards of 90 plus percent. But then if you break down by subgroups, by race and gender, the accuracy actually gets lower, and that's a big, big problem. Problem, and so we really have to address that. There is certainly a lot of confusion about what machine learning is. Machine learning, really, is just a fancy term for different statistical techniques powered by computers to build models that make predictions about something. Any system that is making those sort of predictions is going to have. Biases, right, and to be biased isn't necessarily a bad thing. It's that you wanna understand what the biases that are built into your system are. And so when we're thinking about using data about historic loan repayment or historic Reoffending where in the criminal justice system we have to understand that those data are collected under these set of conditions that I think everybody agrees were discriminatory. And so, those data themselves embed those practices of discrimination and so if we're trying to learn from them and use them to predict the future. What we basically do is making the future carry all the baggage of the past. So we need to be really attentive to all of the historical biases that are in those data. The issue around bias in artificial intelligence has to do with the data that's put into it. But it also has to do with the questions that were asked. Asking. Pymetrics uses neuroscience and artificial intelligence to help companies hire in a more predictive and fair way. Artificial intelligence doesn't have to have bias in it. So long as it's explainable, and so long as it's audited. We actually build custom models for all the clients that we work with. And we profiled their high performers using our platform. And then we take that data and we use AI to figure out what are the features that driving high performance in that particular role. So whenever we build a new algorithm, we will essentially test for statistical differences between genders and ethnicities. And if we do find them audit AI will actually go in and actually de weight or remove some of those features that are causing the bias. Until the algorithm is not basically preferencing either gender or ethnicity. We build consumer lending technology that enables banks to offer expanded access to credit. We try to build models that are far more accurate than traditional models based only on credit score and income. These models enable us to understand many more signals of someone's credit- worthiness. And As a result, we have found that we are able to approve about twice as many people as a traditional bank can, with similar loss rates. It's a bias if you are for some group just giving them loans even though they don't really pay you back. And it's also a bias if you are not giving loans to a group that would have also paid you back. And both of those are bad things, and you want to minimize both of those. There's a lot of bias that's inherant in society, and just in what we do as humans. And there's no way to purely reduce that completely, or have a perfect system. So it's really critical that we ensure that everyone can be part of shaping it. To ensure that the benefits will be shared and that we mitigate some of the risks associated with the development of the technology. We need to make sure that our technical systems are bent and shaped in a way that protect our committments to a just society. Are commitments to freedom of information and freedom of expression. Are commitments to privacy and those are things that we can do. We can use machine learning. We cand use AI in ways that are equitable. And just, and fair. And help us have a kind of society we want. [ BLANK_AUDIO ]

These companies are working to find a solution.