Amazon scrapped a recruiting AI that learned to penalize women's resumes

Amazon trained a recruiting model on a decade of resumes that skewed male and the model learned to downrank resumes that included the word women's, women's chess club, or all-women's colleges. The team scrapped the project.

Amazon · Incident Oct 10, 2018 · Indexed May 13, 2026 · 2 sources

If you train on a biased corpus, you ship a biased model. The corpus is the policy.
What
Amazon trained a recruiting model on a decade of resumes that skewed male and the model learned to downrank resumes that included the word women's, women's chess club, or all-women's colleges.
Incident date
Oct 10, 2018
Who
Amazon
Failure mode
Policy Violation
AI surface
Agentic Workflow
Severity
High

What happened

In 2018, Reuters reported that Amazon had quietly scrapped an internal recruiting AI it had been building since 2014. The model had been trained on a decade of Amazon resumes, which were predominantly male. The model learned to penalize resumes that mentioned "women's clubs," "women's chess," or graduates of all-women's colleges. Amazon scrapped the project rather than fix it.

The case is the foundational example of bias amplification in hiring AI and is cited in every regulator-facing conversation about AI fairness since.

What broke inside the model

Failure path · mode profile · Policy Violation
  1. 01 · TriggerA prompt pushes against a deployment boundary.
  2. 02 · Model stepThe model produces the disallowed output.
  3. 03 · Control gapNo enforcement blocks it at generation time.
  4. 04 · FailureThe output crosses the policy line.
  5. 05 · ConsequenceA limit the business set is breached in public.

The output crosses a policy boundary the deployment had defined.

The model learned what it was shown. The training data was biased; the model reproduced the bias. The mechanism is identical to the Apple Card case: features that correlate with a protected attribute produce divergent outcomes by protected attribute, even when the protected attribute is not an input.

Public visibilityHigh
Regulatory exposureNone
Customer impactClass-wide
Financial impactUnknown
Time to disclosureMonths
  1. PressAmazon scrapped sexist AI recruiting toolbbc.com
  2. PressAmazon scraps secret AI recruiting tool that showed bias against womenreuters.com
Permalinkhttps://failureindex.ai/failures/amazon-recruiting-ai-bias-against-women
CitationAI Failure Index. "Amazon scrapped a recruiting AI that learned to penalize women's resumes" (FI-0011). Realm Labs. https://failureindex.ai/failures/amazon-recruiting-ai-bias-against-women (indexed May 13, 2026).
Share cardA branded image of this record for posts and slides.

Data fields CC-BY 4.0, prose citation permitted. Incident ID FI-0011. Full dataset at /data.

Note from Realm Labs, the Index steward

How Realm would have caught this

Controls for this failure mode
  • Prism
  • OmniGuard

Realm reads the model's decision representation against a fairness policy. When the model's feature attribution clusters on terms that correlate with a protected attribute, Prism flags it and OmniGuard requires either a remediation or a human review before the decision is rendered. The model can still be used; the operator just gets the runtime signal that lets them remediate before the headline.