Taking the control back – An adventure in developing personalized content moderation
This is a compelling CHI honorable-mention because the paper’s value comes from a rare combination of lived need, working system building, and reflective analysis. Its strongest claim is practical and situated: personalized moderation can be made effective for one severe harassment case, but the paper is careful enough to show why that success does not automatically generalize.
Axes Lens
Rare contribution shape, typical evidence profile. The point here is not a score. It is to show what kind of claim the paper makes, and whether the evidence pattern is unusual or baseline in this 268 -review set.
Contribution shape
- Knowledge form
- technical knowledge typical · 50/268
- Novelty type
- system architecture typical · 35/268
- Abstraction level
- system typical · 61/268
- Generalization target
- user population typical · 75/268
- Validation mode
- field deployment typical · 9/268
Evidence profile
- Evidence strength
- strong typical · 158/268
- Claim alignment
- strong typical · 231/268
- Overclaim risk
- medium typical · 210/268
Review Summary
This paper is best read as a field-grounded systems contribution with a strong critical-design flavor. The abstract and reported outcomes support a clear story: a victim of sustained harassment built an automated, personalized, collaborative anti-harassment system, then had to rebuild it as platform APIs changed and tightened. That is the real novelty here. It is not merely that moderation should be personalized; rather, the paper demonstrates what it takes to operationalize that idea under adversarial, platform-constrained conditions, and it surfaces the maintenance burden and power asymmetries that standard moderation discussions often understate. The evidence for effectiveness is concrete but bounded: the paper reports that 81% of 231 harasser accounts were suspended or deleted by July 15 2023, later rising to 96% of 267 accounts by October 11 2023. Those numbers make the validation persuasive as a field deployment, but they do not convert the work into a broadly validated general solution. The paper is appropriately cautious about scope, noting that the current system does not address networked or coordinated harassment campaigns and that different threat models would require different handling. That limitation matters because it marks the boundary between a powerful personal defense system and a universal moderation architecture. As a CHI contribution, the paper’s strength is the combination of artifact, lived experience, and critical reflection: it offers technical insight into personalized moderation while also making a normative argument about user control and platform responsibility. The main risk is overgeneralization, not lack of substance. In short, this is a strong, situated contribution whose significance lies in showing both the promise and the limits of victim-centered moderation in practice.
What Changed
Canon before
Prior CHI moderation work and platform safety systems generally emphasize platform-level, one-size-fits-all interventions rather than victim-specific control over harassment handling.
Departure from common sense
The paper’s framing pushes against the default assumption that moderation should be uniform and centrally administered. It argues that a victim facing sustained harassment may need individualized controls and that the platform’s standard intervention model can be inadequate for that situation.
Actual novelty
The contribution is not a new moderation theory in the abstract; it is a concrete personalized anti-harassment system built by the victim herself and iteratively rebuilt under platform API changes and restrictions. The novelty lies in the lived, end-to-end design and maintenance of a personalized moderation workflow, plus the critical analysis that emerges from that process.
Evidence
The paper presents a personalized anti-harassment system developed in response to a real sustained harassment campaign on Twitter. The abstract states that the system was constructed to regain control and that the experience of developing and re-developing it under API changes highlights design issues. The paper also reports concrete outcomes: 81% of 231 harasser accounts suspended or deleted by July 15 2023, rising to 96% of 267 accounts by October 11 2023. The paper itself notes limits around networked/coordinated harassment and the fact that the author did not preserve harassing content.
“ Rather than relying on a manual journal, this work primarily uses the logs generated automatically by my anti-harassment system, which documents metadata for every Twitter interaction (including follows, comments, likes, reposts, and quotes) I recei”
actual novelty · Share on · confidence 0.75
“ Current online moderation follows a one-size-fits-all approach, where each intervention is applied in the same way to all ”
departure from common sense · Current online moderation follows a one-size-fits-all approach, where each intervention · confidence 0.30
“ During member checking, one person noted that the current system design does not address networked or coordinated harassment campaig”
limitation · Share on · confidence 0.70
“ The median time for Twitter to take actions on the reported accounts is 14 days, and the statuses of the harasser’s accounts at July 15, 2023 and October 11, 2023 are shown in Fig”
validation scope · Share on · confidence 0.65
Limits
Method limits
The evidence is grounded in a single, deeply contextualized case and in the author’s own deployment experience. The paper does not establish comparative efficacy against alternative moderation strategies, nor does it provide broad experimental generalization across users, platforms, or harassment types.
Deployment limits
The system’s reported effectiveness depends on a specific harassment pattern and on platform-specific affordances and restrictions. Repeated API changes and restrictions are part of the deployment story, so portability to other platforms or future API regimes is uncertain.
Boundary conditions
The paper itself indicates that the current design does not address networked or coordinated harassment campaigns and that other threat models may require different handling. It also reflects a case where the author assessed no physical threat and chose not to preserve harassing content.
Position in field
This sits at the intersection of personal informatics, safety tooling, and content moderation, but its strongest contribution is as a critical, practice-grounded account of personalized moderation under real adversarial conditions rather than as a general-purpose moderation framework.