Google strives to create bias-free AI

Google is attempting to right the errors of the past and spearhead the creation of truly bias-free AI tools

3Dec

Google has revealed that its new Gmail feature, which autocompletes sentences for users composing emails, will never suggest any gender-based pronouns. The adaptation forms part of a wider effort to curb inherent societal biases across Google's platforms and services that can be indirectly trained into machine learning (ML) algorithms.

The decision to employ the approach was reached by Paul Lambert, product manager at Gmail, when a research scientist typed "I am meeting an investor next week", and Google's autocomplete function, Smart Compose, suggested, "Do you want to meet him?". This illuminated the implicit bias ML algorithms pick up when being trained on datasets that reflect existing societal biases.


Visit Innovation Enterprise's Machine Learning Innovation Summit, part of the DATAx New York festival, on December 12–13, 2018


It was recently revealed that Amazon had scrapped a recruitment AI for accidentally programming in gender biases into the tool. Likewise, Google has been in hot water in the past for creating what have been considered "culturally insensitive" AI algorithms. Hence, this time, Lambert said the tech giant was taking no chances and removed all gender-based pronouns from the tool.

"Not all 'screw-ups' are equal. Gender is 'a big, big thing' to get wrong," said Lambert.

In an effort to develop more ethical and culturally sensitive AI, Google has also launched the "Inclusive Images Competition". In a September 2018 blog post, the tech giant admitted that while large datasets such ImageNet and Open Images have been vital towards the training of computer vision AI, they have been "found to be geographically skewed based on how they were collected". This lack of representation leads to skewing of machine learning results.

The crowdsource project is asking volunteers across the world to upload images from their surroundings in order to build up the cultural diversity of Google's ML databases.

"We hope that these datasets, built by donations from Google's global community, will provide a challenging geographically-based stress test for this competition," added the blog post.

"We also plan to release a larger set of images at the end of the competition to further encourage inclusive development, with more inclusive data."

Results of the competition will be announced on December 2 at the Conference on Neural Information Processing Systems in Montreal, Canada. 

Top tips to make the most of your business mobile app in 2019small

Read next:

Six tips to make the most of your business app in 2019

i