Two takes on Algorithmic Audits

From external to internal audits in a 6-year span.

Photo by Christian Fregnan; Edited by me.

Algorithmic audits is to audit results of algorithmic systems. Though it has been mentioned here and there when discussing algorithmic accountability approaches, it was less discussed compared to transparency and explanations. Available literature are mostly from two groups, “project: auditing algorithms” and Google.

Before 2020, most of the algorithmic audit papers are written by the former group in 2014, they focused on external audits and proposed 5 sets of designs in Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms. Since the audit is performed by external forces, having no access to the actual algorithms, the suggested designs are different in how to get the war materials for audits and who will perform the audit.

1. Code Audit: Performing a “code audit” is for researchers to obtain a set of original or similar codes.
2. Non-invasive User Audit: A non-invasive user audit is to audit users’ self-reported answers regarding their interactions with systems. Non-invasive audit is to asks users to answer traditional science format survey, therefore, the authors acknowledged that it could be a stretch to consider it as a type of audit. The design has the advantage of not perturbing the platform yet since the audit design only relies on users’ submission, if the samples are not enough to reflect the overall performance of such algorithm, the audit result would be flawed. It would also suffer from having no control over the quality of user-submitted materials, thus making the validity an insurmountable problem. Cognitive biases and unreliable memory would introduce even more doubt toward the audit results.
3. Scraping Audit: A scraping audit is to audit the materials “scrapped” by researchers. The scrapping is done by researchers repeatedly issue queries to a platform and to observe the results. Compared to non-invasive user audits, scraping audits are more intrusive and more likely to be found violating platforms’ terms of services. This design could gather audit materials in a more efficient manner, yet since researchers might have different access to the platform, such as APIs, or distinct behavioural patterns, it is possible that the auditing results would be less authentic considering how different researchers and users interact with the system.
4. Sock Puppet Audit: A sock puppet audit is researchers using computer programs to impersonate users to make queries, such as creating multiple accounts
5. Crowdsourced Audit: The crowdsourced audit is very similar to the sock puppet audit; it only differs in one respect: in the former design the tester is a human. The audit materials be gathered by actual users recruited on crowdsourcing platforms such as Amazon Mechanical Turk.

Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing at https://arxiv.org/pdf/2001.00973.pdf

The recent focus on FAT* ML definitely brings more attention to algorithmic approaches, therefore, resulting Google proposing an internal audit framework, SMACTR, in 2020. Since there’s no need to worry about not being able to access audit materials, SMACTR focuses more on the execution details. It lists out 5 stages for performing an internal audit.

1. Scoping: Define the scope of the audit. Set out the ethical expectations and principles for such AI system to follow.
2. Mapping: Identify individuals that contribute to the system and document their participant for future accountability.
3. Artifact collection: Make sure every material are prepared.
4. Testing: The actual testing.
5. Reflection: Reflect whether or not AI is in conflict with the principles set forth.

Above is a way too simplified note, if interested please give the paper a read, it’s very interesting and definitely worth some thoughts.

Please let me know if there are ever new ways for algorithmic audits! They are so hard to find!! Or maybe because they being hard to achieve and the fact audit results might be false makes it less appealing?

Christian Sandvig et al., “An algorithm audit” in Data and Discrimination: Collected Essays. (Washington, DC: New America Foundation, 2014) 6.

Christian Sandvig, Kevin Hamilton & Karrie Karahalios, Auditing Algorithms: Research Methods for Detecting Discrimination on Internet Platforms (May 2014), online: Data and Discrimination: Converting Critical Concerns into Productive Inquiry <https://pdfs.semanticscholar.org/b722/7cbd34766655dea10d0437ab10df3a127396.pdf>.

Inioluwa Deborah Raji et al., “Closing the AI accountability gap: defining an end-to-end framework for internal algorithmic auditing” in Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency (New York: ACM New York, 2020) 33.