Author(s)
Source
UCLA Law Review, Vol. 66, pp. 54-141, 2019
Summary
Algorithmic decision-making may perpetuate bias if the data used to train the system reflects bias. Thus far, regulators and the courts have not addressed algorithmic bias effectively.
Policy Relevance
Self-regulation and whistleblower protections could support algorithmic accountability.
Main Points
- For members of certain groups, such as the less wealthy, an algorithmic mistake can lead to disastrous denial of employment, housing, or insurance.
- Automated decision-making initially appears free from human prejudice and irrational biases, but algorithmic models are the product of fallible creators.
- Bias in big data generally arises from one of two causes:
- Errors in data collection, which leads to inaccurate depictions of reality.
- The algorithm is trained using data that reflects bias (for example, data on promotions are collected from an industry that systematically promotes men over women), and thus perpetuates the bias.
- An algorithm labelled black defendants as future criminals at twice the rate of white defendants; nonetheless, the Wisconsin Supreme Court upheld the algorithm’s use in sentencing, even though trade secret law barred examination of the algorithm in court.
- Two foundational concepts in antidiscrimination law come into conflict in cases involving big data.
- Anticlassification principles suggest that the very act of classification risks unfairness.
- Antisubordination theory uses classifications to remedy inequalities.
- The law cannot address algorithmic bias without adopting antisubordination theory, but this is constitutionally infeasible.
- Private firms and organizations such as the Association for Computing Machinery are formulating ethical standards for artificial intelligence; to be most effective, these standards require regulatory oversight.
- Human impact assessments could guide firms in addressing algorithmic bias; such assessments should include:
- A substantive commitment to the fair treatments of all races, cultures, and income levels.
- Structures that promote oversight of programmers (who develop the algorithm) by controllers (those responsible for compliance).
- The assessment should examine the algorithm, its output, and the training data.
- Algorithmic civil rights concerns require a new approach to trade secret law, because accountability is impossible without transparency.
- Whistleblowing involves the disclosure of an organization’s wrongful practices by a member of the organization; whistleblower protections could help address algorithmic bias.
- Whistleblowing is especially effective when there is information asymmetry.
- Whistleblowing is appropriate when the government relies on private entities to carry out public functions.
- The Defend Trade Secrets Act of 2016 immunizes whistleblowers from liability under trade secret law for making confidential disclosures to regulators; similar protections could immunize whistleblowers from trade secret liability to promote algorithmic transparency.