You Are a Political Junkie and Felon Who Loves the Sound of Blenders:
 Machine Learning Taste Publics

• In Event: Algorithms, Public Culture, and Platform Responsibility

Abstract

Consider three cases: First, during the last presidential election, Google searches for “Iran,” “gay marriage,” and “Medicare” provided some people with Obama’s political positions, even though their query did not contain the word “Obama,” Obama’s positions were not normally the top results, and the positions of Obama’s opponent received no preferential display (Angwin, 2012). Second, the Shazam song identification algorithm has determined that the song “Bangarang” by the artist “Skrillex” is equivalent to the sound of an Oster household all-purpose blender (Wang, 2003; electrolemon, 2013). Third, searches for African-American names on Reuters.com produced disproportionately more advertising suggesting that the person with that name has an arrest record (Sweeney, 2013).

Each of these three cases are examples of unanticipated results produced by machine learning (ML) when it is operating as designed. ML algorithms couple statistical learning or data reduction with a decision, meaning that an ML algorithm does not make judgments on the basis of an explicit rule written by its designer, but instead learns from a set of data to produce new decision criteria that no human has ever vetted. This paper considers the cultural, social, and political consequences of widespread ML curation, arguing that ML tends to routinely surface associations that, while they may be “successful” or valid in some sense, should not be employed in the context where they appear.

ML programmers know that machines can learn from data produced by humans, but do not acknowledge how humans learn from the media. ML systems falsely equate success with the satisfaction of individual preferences, discounting both system-level effects and feedback. This means that widespread use of ML algorithms has the potential to reinforce and expand the ugliest and least desirable human behaviors and to advantage particular taste publics (Gans, 2008) at the expense of others. In the ML world, cultural products suffer if they involve patterns that are difficult to classify or if they produce affect that is not correlated with other affect (sometimes called “The Napoleon Dynamite Problem”). Finally, formerly protected categories (Gandy, 2012) of expression lose protection because ML systems are not aware of them.

Author

• Christian Sandvig, U of Michigan, Ann Arbor