jackVerified account

We build our policies and rules with a principle of impartiality: objective criteria, rather than on the basis of bias, prejudice, or preferring the benefit to one person over another for improper reasons. If we learn we failed to create impartial outcomes, we work hard to fix.

In the spirit of accountability and transparency: recently we failed our intended impartiality. Our algorithms were unfairly filtering 600,000 accounts, including some members of Congress, from our search auto-complete and latest results. We fixed it. But how did it happen?

Our technology was using a decision making criteria that considers the behavior of people following these accounts. We decided that wasn’t fair, and corrected. We‘ll always improve our technology and algorithms to drive healthier usage, and measure the impartiality of outcomes.

Bias in algorithms is an important topic. Our responsibility is to understand, measure, and reduce accidental bias due to factors such as the quality of the data used to train our algorithms. This is an extremely complex challenge facing everyone applying artificial intelligence.

For our part, machine learning teams at Twitter are experimenting with these techniques and developing roadmaps to ensure present and future machine learning models uphold a high standard when it comes to algorithmic fairness. It’s an important step towards ensuring impartiality.

Looking at the data, we analyzed tweets sent by all members of the House and Senate, and found no statistically significant difference between the number of times a tweet by a Democrat is viewed versus a Republican, even after our ranking and filtering of tweets has been applied.

Also, there’s a distinction we need to make clear. When people follow you, you’ve earned that audience. And we have a responsibility to make sure they can see your tweets. We do not have a responsibility, nor you a right, to amplify your tweets to audiences that don’t follow you.

What our algorithms decide to show in shared spaces, like search results, is based on thousands of signals that constantly learn and evolve over time. Some of those signals are engagement, some are the number of abuse reports. We balance all of these to prevent gaming our system.