W3110881698 Working papers AI × management and organizations

Biased Programmers? Or Biased Data? A Field Experiment in Operationalizing AI Ethics

Bo Cowgill · Fabrizio Dell'Acqua · Sam Deng · Daniel Hsu · Nakul Verma · Augustin Chaintreau

Also recorded as: doi:10.2139/ssrn.3615404 · ssrn:3615404

ssrn abstract page · 2020-06-24

Recorded claims
2
coded from the inspected source
Evidence classes
1
ssrn-abstract-page
Recorded limitations
2
stated, never hidden

From the hash-bound evidence releases

Recorded claims

Each claim states what the inspected source says, at the recorded location—bounded by its scope and graded by its confidence. Nothing here is a synthesis across works.

A field experiment evaluated 8.2 million algorithmic predictions of math skill from roughly 400 AI engineers under randomized changes to incentives, training data, ethics awareness, and technical knowledge, then audited out-of-sample predictions for about 20,000 subjects; the abstract attributes most bias to training data and reports that one-third of the benefit of better data operated through greater engineer effort and incentive responsiveness.

confidence: high ssrn-abstract-page apw-e1-ssrn-backfill

Scope: Universe and denominators: approximately 400 AI engineers, 8.2 million predictions, and about 20,000 audit subjects; unit: engineer/algorithmic prediction; period: not displayed; condition counts are not displayed.

The authors also evaluate practical managerial or policy interventions—including technical advice, reminders, and improved incentives—for reducing algorithmic bias, establishing an organizational mechanism beyond general AI-ethics discussion.

confidence: high ssrn-abstract-page apw-e1-ssrn-backfill

Scope: Universe: participating engineers and their algorithms; unit: engineer effort and prediction performance; period and intervention-arm denominators are not displayed; no numerical effect is asserted in this claim.

Structured relationships

In-corpus citations

Only source-supplied cites relationships whose two endpoints are admitted are shown. Invocation evidence remains a separate relationship.

Cites (0)

No in-corpus outgoing citation is available in the current enrichment coverage.

Cited by (0)

No in-corpus incoming citation is available in the current enrichment coverage.

Boundaries

Limitations & independence

Recorded at coding time, carried with the work forever. A claim without its limits is not evidence.

Recorded limitations

  • Abstract-level evidence only; full text was not inspected.
  • Condition counts, study period, uncertainty, and outcome-specific denominators are not displayed; the SSRN page rounds the engineer and subject counts.

Source independence

ssrn-only