The social network Twitter has recently launched a new initiative with which it seeks to evaluate and analyze its artificial intelligence systems in order to better “educate” users, thereby addressing the growing concerns about the biases exhibited by its algorithms.
The objective? Putting some of your machine learning systems under much deeper scrutiny, paving the way for changes to any model.
It would be, specifically, an initiative whose objective is not only to increase the transparency of the different artificial intelligence (AI) systems used by Twitter , but to improve the fairness of its algorithms , offering users a kind of “algorithmic choice” on the different technologies that may affect them directly.
In this sense, Twitter has committed to taking responsibility for the algorithmic decisions of the platform, and decided to appoint a working group a few weeks ago to lead this initiative, in which it seems to have put all its efforts. Not surprisingly, this group has members from across the company, and would be managed by the team in charge of ethics, transparency and responsibility of Twitter.
We must bear in mind that, every day, around 200 million users use this social network. A platform that relies on machine learning models for a wide variety of tasks and functions , from organizing content based on its relevance to identifying posts that violate terms of use.
As experts say, when the social network uses machine learning, hundreds of millions of tweets a day can end up being affected. In this sense, Twitter itself has been concerned about the possibility that a system originally designed to help may start differently than expected at first.
To see the results of the analysis we will have to wait a few months. But when we can access the assessment, we could learn what the potential harm would be from three of the platform’s main algorithms: an analysis of the racial and gender bias of image cropping; an assessment of the fairness of the recommendations that appear in our Timeline in different racial subgroups; and an analysis of the content recommendations for different political ideologies in a total of seven countries.
The first of the analyzes, for example, is especially related to the problem that the social network faced when it discovered that the use of clipping images of the platform automatically favored white faces (a bias that they have already talked about many AI specialists).
In this sense, although the company recognized that the way in which the photos were automatically recovered could cause damage, in reality they did not show evidence of racial or gender bias.
In any case, there is no doubt that the decision is very important, since by committing to carry out a much more in-depth analysis of its platform, the social network has also committed to giving users greater visibility in the future .
