There was the voice recognition software that struggled to understand women, the crime prediction algorithm that targeted black neighbourhoods and the online ad platform which was more likely to show men highly paid executive jobs.
Concerns have been growing about AI’s so-called “white guy problem” and now scientists have devised a way to test whether an algorithm is introducing gender or racial biases into decision-making.
Mortiz Hardt, a senior research scientist at Google and a co-author of the paper,said: “Decisions based on machine learning can be both incredibly useful and have a profound impact on our lives … Despite the need, a vetted methodology in machine learning for preventing this kind of discrimination based on sensitive attributes has been lacking.”
The paper was one of several on detecting discrimination by algorithms to be presented at the Neural Information Processing Systems (NIPS) conference in Barcelona this month, indicating a growing recognition of the problem.
Nathan Srebro, a computer scientist at the Toyota Technological Institute at Chicago and co-author, said: “We are trying to enforce that you will not have inappropriate bias in the statistical prediction.”
The test is aimed at machine learning programs, which learn to make predictions about the future by crunching through vast quantities of existing data. Since the decision-making criteria are essentially learnt by the computer, rather than being pre-programmed by humans, the exact logic behind decisions is often opaque, even to the scientists who wrote the software.
“Even if we do have access to the innards of the algorithm, they are getting so complicated it’s almost futile to get inside them,” said Srebro. “The whole point of machine learning is to build magical black boxes.”
To get around this, Srebro and colleagues devised a way to test for discrimination simply by analysing the data going into a programme and the decisions coming out the other end.
“Our criteria does not look at the innards of the learning algorithm,” said Srebro. “It just looks at the predictions it makes.”
Their approach, called Equality of Opportunity in Supervised Learning, works on the basic principle that when an algorithm makes a decision about an individual – be it to show them an online ad or award them parole – the decision should not reveal anything about the individual’s race or gender beyond what might be gleaned from the data itself.
For instance, if men were on average twice as likely to default on bank loans than women, and if you knew that a particular individual in a dataset had defaulted on a loan, you could reasonably conclude they were more likely (but not certain) to be male.
However, if an algorithm calculated that the most profitable strategy for a lender was to reject all loan applications from men and accept all female applications, the decision would precisely confirm a person’s gender.
“This can be interpreted as inappropriate discrimination,” said Srebro.
The US financial regulator, the Consumer Financial Protection Bureau, has already expressed an interest in using the method to assess banks.
However, others have raised concerns that the approach appears to side-step any requirement for transparency about how decisions made by algorithms are actually reached.
Alan Winfield, professor of robot ethics at the University of the West of England, said: “Imagine there’s a court case for one of these decisions. A court would have to hear from an expert witness explaining why the program made the decision it did.”
Winfield acknowledged that an absolute requirement for transparency is likely to prompt “howls of protest” from the deep learning community. “It’s too bad,” he said.
Noel Sharkey, emeritus professor of robotics and AI at the University of Sheffield, agreed. “Machine learning is great if you’re using it to work out the best way to route an oil pipeline,” he said. “Until we know more about how biases work in them, I’d be very concerned about them making predictions that affect people’s lives.”