Smart Software and Bias: Math Is Not Objective, Right?

December 12, 2016

I read “5 Unexpected Sources of Bias in Artificial Intelligence.” Was I surprised? Yep, but the five examples seemed a bit more pop psychology than substantive. In my view, the bias in smart software originates with the flaws or weaknesses in the common algorithms used to build artificially intelligent systems. I have a lecture about the ways in which a content creator can fiddle with algorithms to generate specific results. I call the lecture “Weaponizing Information: Using Words to Fiddle with Algorithms.” (Want to know more? Write benkent2020 at yahoo dot com. Be aware that this is a for fee presentation.)

This “5 Unexpected…” write up offers these ideas:

  • Data driven bias. The notion is that Stats 101 injunctions are happily ignored, forgotten, or just worked around. See what I mean? Human intent, not really mathy at its core.
  • Bias through interaction. The idea is that humans interact. If the humans are biased, guess what? The outputs are biased, which dominoes down the line. Key word: Human.
  • Emergent bias. This is the filter bubble. I view this as feedback looping, which is a short cut to figuring out stuff. I ran across this idea years ago in Minneapolis. A start up there was explaining how to let me do one thing to inform the somewhat dull system about what to present. Does this sound like Amazon’s method to you?
  • Similarity bias. Now we are getting close to a mathy notion. But the write up wanders back to the feedback notion and does not ask questions about the wonkiness of clustering. Sigh.
  • Conflicting goals bias. Now that puzzled me. I read the paragraphs in the original article and highlighted stereotyping. This struck me as a variant of feedback.

Math is sort of objective, but this write up sticks to some broad and somewhat repetitive ideas. The bias enters when thresholds are set, data are selected, processes structured to deliver what the programmer [a] desires, [b] ze’s boss desires,  [c] what can be made to run and sort of work in the time available, or [d] what the developer remembers from a university class, a Hacker News post, or a bit of open source goodness.

The key to bias is to keep the key word “human” in mind.

Stephen E Arnold, December 12, 2016

Comments

Comments are closed.

  • Archives

  • Recent Posts

  • Meta