Google has been pushing ahead and breaking barriers with its work bringing powerful AI to consumer products. Google’s Assistant grabs all the headlines with John Legend’s voice and the full-on human impersonating Duplex feature, but there are plenty of other Google products packing a powerful AI punch. Google Photos’ Live Albums show just how good AI features can actually be and Gmail’s Smart Compose can take the strain out of having to write boring emails all the time. The Smart Compose feature is now highlighting some of the advanced issues facing Google’s AI efforts as the company leaps forward.
Google has removed gender pronouns from Gmail’s Smart Compose feature to avoid AI gender bias
Smart Compose works by analyzing past emails and using them to predict what a user will say in the email they are writing. As if to show just how in-sync Gmail is with its users, Smart Compose can even go as far as finishing their sentences for them. This sounds great and, in many cases, it works without a problem.
The system isn’t infallible though as predictions can only ever be made based on the data that is being fed into the algorithm. If an email includes a gender-neutral word like “doctor” or “teacher,” there is nothing there to give Smart Compose an indication of what gender that doctor or teacher may be. If Smart Compose offers a pronoun like “him” or “her” in a predicted reply, there is a chance it will pick the wrong gender. That’s a public relations nightmare waiting to unfold.
This problem first came to Google’s attention back in January when a research scientist at Google was writing an email about meeting an investor. Gmail suggested the pronoun “him,” which wrongly assumed that the investor was male. Speaking to Reuters about the problem, Gmail product manager Paul Lambert said that he and his team had been working on a solution to the problem but that they couldn’t find a reliable enough answer to roll-out. This led to the decision to remove gender pronouns from Smart Compose altogether. Lambert was eager to show though, that this decision will affect less than one percent of all Smart Compose predictions.
The issue raises a fundamental limitation in how AI learns and improves itself. As all learning comes from past data, an inherent discrimination in that data will present itself in the actions of the AI. Even if it isn’t discrimination per se, there is no reason to expect that just because a data set includes more male investors than female investors there is anything to say that the current investor in question will be male or female. This provides a problem for AI developers who don’t want to be seen as perpetrating the biases found in the data they are using to train their AI. And it’s a problem for society at large as AI is given more and more responsibility over our daily lives.
For a relatively small issue like predicting email sentences, this is not a serious problem. As AI begins to do more and more jobs like sorting through job applications, however, a problem like this could be huge. With that in mind, it is important to remember that Google couldn’t fix this problem, they just removed the issue causing it.