Pitfalls of Binary Classification

Policies: Trust & Safety Class/Quizzes/Political Ad Policies
Cases: Trust & Safety Class/Quizzes/Assessing T&S Policies

Learning Objective(s)

  • Classify easy examples of online harms according to a platform policy
  • Identify borderline cases where the determination of harm is ambiguous
  • Diagnose how and why policies begin to breakdown on these examples
  • (System) Criticize how automated content moderation mechanisms handle such cases
  • (System) Propose measurement strategies to track such failures
  • (System) Propose alterations to content moderation mechanisms to account for the identified borderline cases

Active Learning Strategy Design

Learning Objective #1: Enforcing platform policies

  • Strategy Name: "Set It Up" (problem solving process)
  • Description: consider a list of 3 example cases against a reference policy (given) to determine is the obvious violation and which is not a violation
  • Expected Outcomes: Platform policies are understood as a process.
  • Assessment Method: Clickers

Learning Objective #2: Identifying edge cases

  • Strategy Name: think-pair-share
  • Description: discuss and action a series of borderline cases using the same policy
  • Expected Outcomes: Borderline cases determination is improved by discussion
  • Assessment Method: Clickers

Learning Objective #3: Diagnosing policy issues

  • Strategy Name: contrasting cases
  • Description: given 2 example policies for the same issue, determine where and why each policy is problematic for the previous examples (ambiguity in policy, binary classification where multiple thresholds are needed, false positives, false negatives,...)
  • Expected Outcomes: each small group may discuss a different set of issues
  • Assessment Method: Discussion

Active Learning for system based questions is left to the assignment

Active Learning for the class in general

Online Trust & Safety tackles a range of online harms.

  • To decide which to cover, use Dotmocracy.

Trust & Safety is about people! Case's presented in the class could be simply portrayed using screenshots in a slide deck, but there is opportunity for more interaction.

  • Fishbowl method: like charades, have a small number of students act out certain cases, and the rest of the class makes the determination based on a platform policy.
    • Extremely risky: have to be very careful about which topics, cases, and platform policies are chosen for this. Students must have the ability to choose another case or refuse to participate.

The lines drawn by platform policies are subject to intense decision making processes (e.g. decisions to take down President Trump's social media accounts during Jan 6th Capitol Riot).

  • Structured debates on safety (public harm) vs censorship (freedom of speech)
    • Question: how do the decisions made by different platforms reflect the tradeoff between safety and censorship?

Design Questions

What active learning strategies to use for homeworks?
How much time to set aside for this in practice?