Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms

From AcaWiki
Jump to: navigation, search

Citation: Christian Sandvig, Kevin Hamilton, Karrie Karahalios, Cedric Langbort (2014/05/22) Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms. Data and Discrimination: Converting Critical Concerns into Productive Inquiry (RSS)




Algorithmic manipulation can be illegal or socially problematic or worthy of scrutiny. Outlines research methods for algorithmic scrutiny, in particular to detect discrimination.

Describes how "audit study" used in non-computer scrutiny, eg by researchers or activists sending fake communications or in person "testers" to employers or real estate agents. Describes ethical challenges of audit design.

Outlines potential methods for auditing computer mediated systems:

  1. Code Audit (Algorithm Transparency): unlikely code would be disclosed, disclosure could help spammers, code is complicated, behavior also depends on data
  2. Noninvasive User Audit (survey users about interaction with system): traditional social science method, but hard to infer causality, correcting sampling problems expensive relative to other methods, self-report may introduce insurmountable validity problems
  3. Scraping Audit: researchers might fall afoul of US Computer Fraud and Abuse Act or site Terms of Service, claim no randomization or manipulation thus difficult to infer causality
  4. Sock Puppet Audit: instead of hiring actors/"testers" as in classic "audit study", use computer program to impersonate users -- includes manipulation, but similar legal problems as previous. Note actually hiring actors would work, but be expensive as detecting discrimination by a computer system might require large number of tests.
  5. Crowdsourced Audit / Collaborative Audit: Uses hired users rather than programs in previous, but uses crowdsourcing (eg mechanical turk) or volunteers aided by collaborative auditing software to keep relatively inexpensive.

Argue for "regulation for auditability" (cf for transparency or misbehavior) including CFAA and scholarly guideline reform, and shift to "accountability by auditing" and funding for research/institutions to perform these.


Finally, a shift of perspective to “accountability by auditing” implies a normative discussion about what the acceptable behavior of Internet platforms ought to be. While some instances of discrimination are obviously unfair and illegal, algorithm audits might produce situations where a comparison to standards of practice are required (as is done in financial auditing) yet no consensus or “accounting standard” for algorithms yet exists. Discovering how algorithms behave on the Internet might then lead to a difficult but important discussion that we have so far avoided: how do we as a society want these algorithms to behave?

Theoretical and practical relevance:


"questions that often have yet to be asked, such as: How difficult is it to audit a platform by injecting data without perturbing the platform? What is the minimum amount of data that would be required to detect a significant bias in an important algorithm? What proofs or certifications of algorithmic behavior could be brought to bear on public interest problems of discrimination? A program of research on auditable algorithms could make great contributions via simulation and game theoretic modeling of these situations."