Skip to the lesson

Data and machine learningLesson 5 of 6

Part 1 of 3, 0 done

14 min

Biased data, biased predictions

Amazon built a résumé-scoring tool that taught itself to mark down the word "women's". Nobody programmed that. It came from the data.

Ticket 01 · The guess

Who does the voice app understand?

In a 2020 study, researchers tested the speech-to-text systems of five big tech companies on recorded interviews with Black and white Americans. For white speakers, the systems got about 19%doi.org ·2020 of words wrong, on average.

For Black speakers?

50%
0%60%
Words to know

Spot something wrong?

Tell us what looks off. No name or email needed.

What kind of problem?

0 / 1000

We'll also send this page, the lesson, ticket 1, so we can find it. Nothing else. Privacy

·
Sources and every number (1)

Sources

  • Racial disparities in automated speech recognition (PNAS 117(14): 7684–7689) · PNAS (Allison Koenecke and others), 2020

Every number

  • 19%doi.org ·2020 · Average word error rate for white speakers across five companies' speech-to-text systems, shown as a percent of words (PNAS, 2020).
  • 35%doi.org ·2020 · Average word error rate for Black speakers across the same five systems, shown as a percent of words (PNAS, 2020).
  1. Lessons›
  2. The school›
  3. Data and machine learning
  • ·Before this: Fraud alert: two ways to be wrong
  • ·Next: Because you watched…
Skip to the lesson

Data and machine learningLesson 5 of 6

Part 1 of 3, 0 done

14 min

Biased data, biased predictions

Amazon built a résumé-scoring tool that taught itself to mark down the word "women's". Nobody programmed that. It came from the data.

Ticket 01 · The guess

Who does the voice app understand?

In a 2020 study, researchers tested the speech-to-text systems of five big tech companies on recorded interviews with Black and white Americans. For white speakers, the systems got about 19%doi.org ·2020 of words wrong, on average.

For Black speakers?

50%
0%60%
Words to know

Spot something wrong?

Tell us what looks off. No name or email needed.

What kind of problem?

0 / 1000

We'll also send this page, the lesson, ticket 1, so we can find it. Nothing else. Privacy

·
Sources and every number (1)

Sources

  • Racial disparities in automated speech recognition (PNAS 117(14): 7684–7689) · PNAS (Allison Koenecke and others), 2020

Every number

  • 19%doi.org ·2020 · Average word error rate for white speakers across five companies' speech-to-text systems, shown as a percent of words (PNAS, 2020).
  • 35%doi.org ·2020 · Average word error rate for Black speakers across the same five systems, shown as a percent of words (PNAS, 2020).
  1. Lessons›
  2. The school›
  3. Data and machine learning
  • ·Before this: Fraud alert: two ways to be wrong
  • ·Next: Because you watched…