Blog

• Proxies for protected classes that sneak into model features

  Excluding race, sex, and other protected attributes from a feature set does not end the fairness question. Proxy variables can still carry the same information into the score.

• Sampling plans that boards misunderstand in bias audits

  Why sampling design determines whether a bias-audit finding can be reconstructed, and the questions boards should ask before accepting a headline ratio.

• What evidence a bias audit should leave behind for counsel

  The documentary trail a careful bias audit should leave so counsel can reconstruct scope, measurement, independence, and remediation when questions arrive later.

• Adverse impact versus disparate treatment in model review language boards understand

  How boards and counsel should keep adverse impact and disparate treatment distinct when reviewing automated decisions, and what follow-up each theory requires.

• How boards are starting to engage with model risk

  A guest post from the Stator AI team on what boards are now asking about model risk in practice, and how audit teams can prepare for the questions that show up at executive level.

• How to scope a bias audit without slowing the ML team

  What a defensible algorithmic bias audit looks like when the goal is regulatory compliance and the constraint is an ML team that still needs to ship product.

• What NYC Local Law 144 actually requires from your team

  A practitioner view of what New York City Local Law 144 actually demands of organizations using automated employment decision tools, separating the headline summaries from the obligations that show up in compliance reviews.