Accountable Algorithms

Privacy and Security and Innovation and Economic Growth

Article Snapshot


Solon Barocas, Edward Felten, Joanna Huey, Joshua A. Kroll, Joel R. Reidenberg, David G. Robinson and Harlan Yu


University of Pennsylvania Law Review, Vol. 165, No. 3, pp. 633-705, 2017


Computers now make many decisions formerly made by humans. Procedures used to oversee human decision-makers cannot be applied to computers. This article describes technological tools to help developers design algorithms consistent with social goals.

Policy Relevance

Fairness and accountability should be built in to computerized processes from the start. Policymakers and computer scientists should work together to ensure accountability.

Main Points

  • Computers use algorithms to approve loan applications, target travelers for search, grant visas, and more; the public and the courts are ill equipped to ensure that algorithms are fair.
  • Some assert that transparency would promote accountability, but this would be ineffective.
    • Firms could disclose their source code, but only experts could understand it.
    • Users could game the system; for example, disclosure of code intended to target tax audits could enable cheaters to evade detection.
  • Randomization improves some computerized processes, but complicates the evaluation process.
    • The Roomba floor cleaner moves in random patterns, so programmers need not program every type of possible motion.
    • It is hard to detect when a corrupt developer skews “random” results.
  • Computer decisions should be made with “procedural regularity,” that is, the same procedure should be applied to everyone.
  • Many technological tools can improve automated decisions, including software that tests every possible outcome, and the use of encryption to “seal” files.
  • The U.S. State Department uses a lottery to grant “green cards” to permanent residents, and some question the fairness of the process.
    • The State Department should publish commitments to its source code and proofs of the code’s fairness.
    • The State Department could work with a trusted third party to ensure that random selections are truly random, and to audit compliance with its commitments.
  • Technological tools can ensure that machine learning systems such as those used to target police searches do not discriminate by race.


Get The Article

Find the full article online

Search for Full Article