Assessing T&S Policies

Problematic Political Ads - Case Studies

Overview

Consider the following controversial political ad examples that expose tensions, inconsistencies, and edge cases in platform policies. These cases are designed to:

  1. Show where Policy A and Policy B give different results
  2. Highlight situations where policies produce seemingly inappropriate or controversial outcomes
  3. Reveal ambiguities and enforcement challenges

Policy definitions: Political Ad Policies

Case 1: The Devil Horns Ad

A political action committee creates an ad featuring a stylized image of the opposing candidate with:

  • Devil horns superimposed on their head
  • An American flag burning in the background
  • Text overlay: "Don't let 'Candidate X' destroy America"
  • Dramatic apocalyptic imagery (crumbling buildings, dark skies)

The ad is targeted to voters in swing states who have previously engaged with content about patriotism and national security.

--

Policy A (Meta-based) - Likely Outcome: ✅ APPROVED

  • Reasoning:
    • Stylized imagery and negative campaigning are explicitly allowed
    • Not a direct call for violence
    • Not voter suppression
    • Custom audience targeting is permitted
    • Dramatic imagery is allowed under free political speech
      Note: Controversy: Many would argue this ad demonizes the candidate and could contribute to hostile climate, but policy prioritizes political speech

--

Policy B (Google-based) - Likely Outcome: ✅ APPROVED

  • Reasoning:
    • False or misleading claims about opponents are allowed
    • Not direct violence incitement
    • Not voter suppression
    • Would be subject to basic targeting only (can't microtarget)
  • Controversy: Same concerns as Policy A; the ad is inflammatory but doesn't violate technical policy provisions

Note: Both policies approve this controversial ad. The main difference is targeting precision.

Case 2: The "Mental Illness" Attack

A campaign creates an ad stating:

  • "Transgender activists are mentally ill and have no place teaching our children"
  • Features images of a transgender school teacher
  • Calls for legislation to ban transgender individuals from teaching positions
  • Text: "Vote 'Candidate X' to protect our kids"

The ad uses custom audience lists of parents with school-age children and targets users who have engaged with content about education policy.

--

Policy A (Meta-based) - Likely Outcome: ✅ APPROVED (as of January 2025)

  • Reasoning:
    • Post-January 2025 policy explicitly allows "allegations of mental illness or abnormality when based on gender or sexual orientation, given political and religious discourse about transgenderism"
    • Policy allows "content arguing for gender-based limitations of military, law enforcement and teaching jobs"
    • Falls under "religious beliefs" exception
    • Custom audience targeting permitted
      Note:
  • Controversy: This represents a major policy shift. Previously would have been removed as hate speech. Now explicitly permitted despite concerns about:
    • Real-world violence against LGBTQ+ individuals
    • Creating hostile environment
    • Spreading misinformation about transgender identity

--

Policy B (Google-based) - Likely Outcome: ⚠️ UNCLEAR

  • Reasoning:
    • Google's policy doesn't explicitly address attacks on protected characteristics in political ads
    • Not technically voter suppression or violence incitement
    • Would likely be approved under "political speech" but enforcement uncertain
  • Controversy: Lack of clear policy guidance on hate speech in political context creates enforcement ambiguity

Note: Policy A explicitly permits this after January 2025 changes; Policy B is ambiguous. This case highlights how recent policy changes normalized anti-LGBTQ+ content.

Case 3: The AI Deepfake

An ad features a highly realistic AI-generated video showing:

  • The opposing candidate appearing to say: "I don't care about working-class families"
  • The candidate never actually said this
  • The ad includes a small disclaimer at the bottom: "This ad contains AI-generated content"
  • Targets working-class voters in specific zip codes using custom audience data

--

Policy A (Meta-based) - Likely Outcome: ✅ APPROVED (with disclosure)

  • Reasoning:
    • AI-generated content depicting false statements by candidates IS allowed
    • Advertiser checked the disclosure box for synthetic content
    • Small disclaimer present (even if hard to notice)
    • Custom audience targeting permitted
    • False claims about opponents are allowed
      Note; Controversy: Even with disclosure, many viewers may believe the video is real, especially if disclaimer is small or easy to miss

--

Policy B (Google-based) - Likely Outcome: ✅ APPROVED (with disclosure)

  • Reasoning:
    • Synthetic content must be disclosed - advertiser complied
    • False claims about opponents are allowed
    • Would use basic targeting only (less precise than Policy A)

Note: Controversy: Same as Policy A - disclosure may be insufficient to prevent deception

Key Insight: Both policies allow deceptive AI content with minimal disclosure requirements. The disclosure is technically compliant but may be functionally inadequate.

Case 4: The Immigrant Invasion Ad

An ad shows:

  • Images of migrants at the border
  • Text: "Illegal immigrants are criminals poisoning our communities"
  • Dramatic music and red alert-style graphics
  • Call to action: "Vote to stop the invasion"
  • Uses lookalike audiences based on people who donated to immigration restriction groups

--

Policy A (Meta-based) - Likely Outcome: ✅ APPROVED (as of January 2025)

  • Reasoning:
    • Post-January 2025 policy allows attacks on immigrants
    • "Claims about immigrants' character or fitness for participation in society" now permitted
    • Not direct violence incitement
    • Lookalike audience targeting allowed
    • Falls under political discourse about immigration
      Note: Controversy:
    • Uses dehumanizing language ("invasion", "poisoning")
    • Research shows this rhetoric correlates with hate crimes against immigrants
    • Previously might have been restricted; now explicitly allowed

--

Policy B (Google-based) - Likely Outcome: ✅ LIKELY APPROVED

  • Reasoning:
    • Doesn't explicitly call for violence
    • Political speech about immigration policy
    • False/misleading claims allowed
    • Can only use basic targeting (not lookalike audiences)
      Note: Controversy: Same concerns about dehumanizing rhetoric

Key Insight: Both approve, but Policy A allows more precise targeting of receptive audiences. The language is inflammatory but doesn't cross the line to direct violence incitement.

Case 5: The Polling Place Confrontation

An ad from a political activist group states:

  • "Patriots: Show up at polling places on Election Day"
  • "Document suspicious voters and ask them for their papers"
  • "Don't let them steal our election"
  • Features images of confrontations at polling places
  • Targets users who have engaged with election fraud conspiracy content

--

Policy A (Meta-based) - Likely Outcome: ❌ REJECTED

  • Reasoning:
    • Violates prohibition on voter intimidation
    • "Coordinated calls to interfere with voting or election processes" is prohibited
    • Could be interpreted as encouraging confrontation at polls
    • Likely flagged as attempting to suppress voting through intimidation
  • Notes: This should be removed, but enforcement may be inconsistent

--

Policy B (Google-based) - Likely Outcome: ❌ REJECTED

  • Reasoning:
    • Explicitly violates prohibition on "content encouraging others to interfere with democratic processes"
    • Telling people to confront voters at polling locations is specifically called out as prohibited
    • "Calls to incite physical conflict... at polling locations to deter voting" is banned
  • Notes: Clear policy violation

Note: Key Insight: Both policies reject this. However, subtle variations (e.g., "observe" instead of "confront") might create grey areas.

Case 6: The False Victory Claim

On election night, before votes are fully counted, a candidate runs ads stating:

  • "Victory! I've won the election!"
  • "Despite fake news media, the real numbers show I won"
  • "Don't believe the lying media when they say I lost"
  • Uses all available targeting to reach maximum audience

--

Policy A (Meta-based) - Likely Outcome: ⚠️ LIKELY REJECTED (but historically inconsistent)

  • Reasoning:
    • "Premature victory claims" before official certification are prohibited
    • However, claiming election fraud/stolen election is now allowed (fact-checking removed)
    • Ambiguous whether "real numbers show I won" crosses the line

Note: Controversy:

  • Enforcement has been highly inconsistent
  • Similar content appeared in 2020 and 2024 elections
  • Policy was supposed to prevent this but failed in practice

--

Policy B (Google-based) - Likely Outcome: ⚠️ UNCLEAR

  • Reasoning:
    • Previously would have been removed
    • As of June 2023, false claims about election outcomes are no longer prohibited
    • However, "premature" claims might still violate voter suppression rules
    • Policy now prioritizes "free expression" over accuracy
  • Controversy: The 2023 policy rollback makes enforcement unclear

Note: Key Insight: Both policies have been weakened on election misinformation. What was once clearly prohibited is now in a grey area.

Case 7: The Targeted Ethnic Attack

In a local election with large immigrant population, an ad states:

  • "[Ethnic Group] voters are destroying our neighborhood"
  • "They don't share our values and shouldn't have a say in our community"
  • Shows unflattering images of people from that ethnic group
  • Uses custom audience targeting to reach residents of specific neighborhoods

--

Policy A (Meta-based) - Likely Outcome: ⚠️ AMBIGUOUS - Likely Rejected but Uncertain

  • Reasoning:
    • Race and ethnicity are still protected characteristics under base policy
    • However, immigration status-based attacks are now allowed
    • Ambiguous whether this is "ethnic" discrimination (prohibited) or "immigration" discourse (allowed)
    • If framed as "immigrants" rather than ethnic group, might be approved
    • Enforcement would depend on exact wording

Note:

  • Controversy:
    • The line between ethnic discrimination and immigration discourse is blurry
    • Ads can be reworded to exploit this ambiguity
    • Custom targeting makes this especially harmful to specific communities

--

Policy B (Google-based) - Likely Outcome: ⚠️ UNCLEAR

  • Reasoning:
    • No explicit policy on ethnic/racial attacks in political ads
    • Not technically voter suppression
    • Not direct violence incitement
    • Likely would depend on human review
  • Controversy: Lack of clear policy creates inconsistent enforcement

Note:
Key Insight: Both policies have ambiguities around ethnic attacks that aren't direct violence incitement. Clever wording can exploit these gaps.

Case 8: The "Groomers" Ad

An ad attacking LGBTQ+ school board candidates:

  • "Stop the groomers from getting near our children"
  • Features photos of LGBTQ+ candidates with ominous music
  • Text: "They want to indoctrinate your kids"
  • Links LGBTQ+ identity with child predation
  • Targets parents using custom audience lists from school districts

--

Policy A (Meta-based) - Likely Outcome: ⚠️ AMBIGUOUS (Likely Approved post-2025 changes)

  • Reasoning:
    • Post-January 2025, allows content about LGBTQ+ individuals and their "fitness" for roles involving children
    • "Mental illness or abnormality" allegations allowed when based on sexual orientation
    • Might be approved as "political discourse"
    • However, "groomer" rhetoric has been linked to violence against LGBTQ+ people
    • Enforcement highly uncertain

Note:

  • Controversy:
    • "Groomer" is a slur falsely associating LGBTQ+ people with pedophilia
    • This rhetoric has preceded real-world attacks on LGBTQ+ individuals
    • Policy changes may have emboldened this type of content

--

Policy B (Google-based) - Likely Outcome: ⚠️ UNCLEAR

  • Reasoning:
    • No explicit policy on LGBTQ+ attacks in political context
    • Not technically violence incitement (though may inspire violence)
    • Would likely depend on individual review
  • Controversy: Dangerous rhetoric in policy grey area

Note:
Key Insight: Policy A's 2025 changes may have opened door to this type of content. Both policies struggle with indirect incitement to violence.

Case 9: The Microtargeted Conspiracy

An advertiser creates dozens of variations of an ad, each tailored to specific audiences:

  • To Latino voters: "Your opponent wants to deport your family"
  • To white rural voters: "Your opponent is giving your jobs to immigrants"
  • To Black voters: "Your opponent supports policies that target your community"
  • Each version uses custom audiences, voter file data, and interest targeting
  • No single message seen broadly; each group sees different (contradictory) claims

--

Policy A (Meta-based) - Likely Outcome: ✅ APPROVED

  • Reasoning:
    • Custom audience targeting explicitly allowed
    • False/misleading claims about opponents permitted
    • Each individual ad doesn't violate policy
    • Ability to show different messages to different groups is a feature, not a bug

Note:

  • Controversy:
    • Prevents public accountability (no one sees all messages)
    • Allows contradictory claims to different audiences
    • Makes fact-checking nearly impossible
    • Enables targeted manipulation
    • This is exactly what Cambridge Analytica did

--

Policy B (Google-based) - Likely Outcome: ⚠️ PARTIALLY PREVENTED

  • Reasoning:
    • Microtargeting is prohibited, limiting precision
    • Can still target different geographic regions or basic demographics
    • False claims allowed but less precisely targeted
    • More likely that contradictory messages would be noticed

Note:

  • Controversy: Basic targeting still allows some audience segmentation with different messages

Key Insight: This case highlights the danger of microtargeting + lack of truthfulness requirements. Policy A enables this; Policy B partially prevents it.

Case 10: The Voter Fraud Hotline

An ad encourages viewers to:

  • "Report suspected voter fraud to our hotline"
  • "If you see something suspicious at polls, document it"
  • Shows examples of "suspicious behavior" that are actually normal (people helping elderly voters, non-English speakers, etc.)
  • Provides a phone number
  • Doesn't explicitly call for confrontation

--

Policy A (Meta-based) - Likely Outcome: ⚠️ AMBIGUOUS

  • Reasoning:
    • Doesn't explicitly call for confrontation or intimidation
    • Could be framed as "election integrity" efforts
    • However, showing normal behavior as "suspicious" could suppress voting
    • "Coordinated calls to interfere" might apply
    • Enforcement would be inconsistent
      Note:
  • Controversy: Chilling effect on legitimate voters without explicit intimidation

--

Policy B (Google-based) - Likely Outcome: ⚠️ AMBIGUOUS

  • Reasoning:
    • Not technically "instructing" interference
    • Could be argued as voter education (though misleading)
    • Doesn't explicitly tell people to create long lines or confront voters
    • Grey area between observation and intimidation

Note:

  • Controversy: Same as Policy A - subtle intimidation that doesn't explicitly violate policy

Key Insight: Both policies struggle with subtle forms of voter intimidation that don't explicitly call for confrontation.

Cases Where Policies Differ Significantly:

  1. Case 2 (Mental Illness): Policy A explicitly permits (post-2025); Policy B unclear
  2. Case 9 (Microtargeting): Policy A enables; Policy B restricts
  3. Case 4 (Immigration): Both approve but Policy A allows more precise targeting

--

Cases Where Both Policies Fail:

  1. Case 1 (Devil Horns): Both approve inflammatory demonization
  2. Case 3 (AI Deepfake): Both allow with minimal disclosure
  3. Case 6 (False Victory): Both weakened enforcement on election misinformation
  4. Case 8 (Groomers): Both struggle with indirect violence incitement

--

Ambiguous Cases (Enforcement Inconsistent):

  1. Case 7 (Ethnic Attack): Depends on exact wording
  2. Case 10 (Fraud Hotline): Subtle intimidation in grey area

Discussion Questions

  1. Policy Effectiveness: Which cases show where truthfulness requirements would matter most?
  2. Targeting vs. Content: Is it worse to have precise targeting of harmful content (Policy A) or broad distribution of harmful content (Policy B)?
  3. Recent Changes: How do Policy A's January 2025 changes affect marginalized communities? What's the trade-off between "free speech" and safety?
  4. Enforcement Gaps: Which cases reveal that written policies don't match actual enforcement?
  5. Indirect Harm: How do policies handle content that doesn't directly incite violence but creates conditions for violence?
  6. Microtargeting: Why is Case 9 particularly problematic? How does it undermine democratic discourse?
  7. AI Content: Is disclosure sufficient for AI-generated content, or do we need stronger restrictions?
  8. Protected Characteristics: Should political speech exceptions exist for attacks on protected groups? Why or why not?

Notes for Instructors

Learning Objectives Addressed:

  • ✅ Classify cases according to platform policies
  • ✅ Identify borderline cases with ambiguous determinations
  • ✅ Diagnose how and why policies break down
  • ✅ Criticize automated content moderation on edge cases
  • ✅ Propose measurement strategies for tracking failures
  • ✅ Propose alterations to address borderline cases

Activity 1: Set It Up (LO1 - Classification)

  • Use Cases 1, 3, 5 (clear outcomes)
  • Students classify: Approve/Reject under each policy
  • Use clickers for real-time feedback

Activity 2: Think-Pair-Share (LO2 - Borderline Cases)

  • Use Cases 2, 7, 8, 10 (ambiguous)
  • Pairs discuss and justify their determinations
  • Compare reasoning

Activity 3: Contrasting Cases (LO3 - Policy Diagnosis)

  • Compare how Policies A and B handle Cases 2, 4, 9
  • Identify: ambiguity, binary classification issues, false positives/negatives
  • Small group discussion on different policy failures

Real-World Context:

These cases are inspired by actual ads that ran on Meta and Google platforms:

  • Case 1: Based on actual anti-Harris ads (2024)
  • Case 2: Based on anti-trans political ads (2023-2025)
  • Case 3: Common deepfake concern across platforms
  • Case 4: Common immigration ad rhetoric
  • Case 8: "Groomer" rhetoric seen in 2022-2024 school board races
  • Case 9: Cambridge Analytica-style tactics

Key Teaching Points:

  1. Policy ≠ Enforcement: Written policies often fail in practice
  2. Grey Areas: Most controversial content lives in ambiguous spaces
  3. Harm Beyond Violence: Indirect harm is real but harder to regulate
  4. Recent Backsliding: Both platforms weakened protections (2023-2025)
  5. Targeting Amplifies Harm: Same content is worse with precise targeting

Course: CSPedagogy / Trust & Safety
Related: old/Trust & Safety Class Old/Quizzes/Political Ad Policies, Active Learning Resources