Open problems

The questions below are understudied, tractable for a single researcher with appropriate methods training, and speak directly to a criminologist's toolkit. Each is framed with what to do, who has done adjacent work, and how hard the data is to get. The list is not exhaustive. It is meant to be a productive starting point.

Q1 — Does ban evasion follow criminal-career patterns?

Trust & Safety operations spend enormous resources on ban-evasion detection, but almost no academic work has characterized ban evaders as a population. Do they show escalating violation severity, desistance over time, or specialization analogous to criminal careers? Method: partner with a platform's research team for anonymized longitudinal account data; apply criminal-career methodology (onset, frequency, duration, seriousness) to violation histories. Feasibility: moderate (data access is the bottleneck). Seed: Seto & Eke (2015); Blumstein & Cohen (1987) on criminal careers.

Q2 — How large is the platform "dark figure" of harm?

Platform enforcement is almost entirely reactive — driven by reports. Many harms are never reported. Estimating the gap between actual and reported harm would reframe T&S resource allocation and policy evaluation. Method: parallel survey design — population prevalence survey alongside platform report data for matched harm types and time windows; apply classical dark-figure estimation from victimology. Feasibility: tractable; the survey component does not require platform partnership. Seed: Skogan (1977); Cross, Smith, & Richards (2014) on underreporting in fraud.

Q3 — How do T&S enforcement decisions vary across demographic groups?

The labeling-theory critique of T&S is substantial but empirically underdeveloped. Critics assert that moderation systems disproportionately flag content from marginalized communities; the evidence base is thin. Method: audit methodology — controlled submission of equivalent content from accounts with varied apparent demographic signals; or analysis of disaggregated transparency-report data. Feasibility: tractable with audit design; moderate with platform data. Seed: Haimson et al. (2021) on differential moderation; Becker (1963) on labeling.

Q4 — Does deplatforming produce displacement or desistance among mid-level actors?

Deplatforming research has focused on high-profile individuals (Yiannopoulos, Jones, etc.). The mid-level actor — network member, not leader — is the modal deplatformed user, and whether they migrate, escalate, or desist is empirically unknown. Method: natural-experiment design exploiting sudden community bans; track coordinated networks across platforms before and after using OSINT. Feasibility: moderate. Seed: Jhaver et al. (2021); Ribeiro et al. (2020) on far-right migration after bans.

Q5 — What is the recidivism rate for T&S interventions, and what predicts it?

Criminologists have decades of actuarial risk-assessment literature. T&S strike systems (warning → temporary suspension → permanent ban) are built on intuition, not on validated risk scoring. Method: survival analysis of account histories after first T&S action; time-to-re-offense as the outcome, with account characteristics as covariates. Feasibility: moderate (requires platform data). Seed: Seto & Eke (2015); Andrews & Bonta (2010) on actuarial risk assessment.

Q6 — Is TFA classified correctly in T&S report queues?

Technology-facilitated IPV is often misclassified in T&S queues as generic harassment, because the context (ongoing relationship, coercive control) is invisible to the reviewer. Systematic misclassification means survivors receive generic enforcement responses rather than elevated safety-planning. Method: qualitative analysis of survivor accounts of platform reporting; matched comparison of queue outcomes for reported TFA vs. generic harassment. Feasibility: tractable (survey) to moderate (platform data). Seed: Dragiewicz et al. (2022); Freed et al. (2018).

Q7 — Do "friction" interventions reduce harmful behavior, and at what cost?

T&S extensively uses friction (confirmations, delays, identity checks) as harm reduction. The counterfactual — how much legitimate use is deterred — is almost never measured. Method: natural experiment exploiting friction rollouts (A/B tests platforms have already run) or regression-discontinuity around friction thresholds; measure both harmful-behavior reduction and behavioral changes in non-violating users. Feasibility: moderate. Seed: Clarke (1997) on situational crime prevention; Brauer & Tittle (2017) on deterrence and legitimacy.

Q8 — Do CSAM offenders show specialist vs. generalist patterns?

The criminological literature on whether CSAM-only offenders escalate to contact offending is contested and consequential for risk-stratified enforcement. T&S decisions about account removal vs. law-enforcement referral would benefit from a sharper empirical picture. Method: meta-analysis of existing longitudinal studies; or primary study using law-enforcement data (requires partnership). Feasibility: hard for primary data; moderate for meta-analysis. Seed: Seto et al. (2011) meta-analysis; Henshaw et al. (2025).

Q9 — What does CIB look like outside political contexts?

Coordinated inauthentic behavior research has concentrated on elections and political contexts. CIB is also used for commercial fraud, harassment campaigns, health misinformation, and financial manipulation. The non-political forms are nearly unstudied in peer-reviewed literature. Method: use publicly available CIB takedown archives (Meta CIB reports, Twitter Safety reports); apply network analysis and typological methods to non-political cases. Feasibility: tractable; data is publicly available. Seed: Cinelli et al. (2023); Starbird et al. (2019).

Q10 — How does T&S secondary trauma compare to other trauma-exposed professions?

Secondary-trauma research in T&S is new. The empirical base for comparing T&S workers to emergency services or social workers — and therefore for benchmarking appropriate organizational support — is almost nonexistent. Method: cross-occupational survey with standardized measures (PCL-5 for PTSD symptoms, ProQOL for compassion fatigue) across T&S workers, emergency responders, and social workers. Feasibility: tractable with appropriate IRB and organizational access. Seed: Spence et al. (2024); Craig & Sprang (2010).

Q11 — Does algorithmic amplification of extremism reflect supply or demand?

Chen et al. (2023) challenged the "pipeline" theory using YouTube-specific data. Replication across platforms with different algorithmic architectures (TikTok, Telegram, Reddit) would determine whether the finding is platform-specific or structural. Method: matched behavioral study using media diaries plus platform data. IRB protocol around extremist-content exposure is a methodological challenge. Feasibility: hard. Seed: Chen et al. (2023); GWU Program on Extremism (2023).

Q12 — What happens to T&S labor markets under regulatory pressure?

The DSA and OSA are creating simultaneous compliance-driven demand for T&S staff. Understanding how labor market dynamics respond — and whether supply of qualified workers tracks demand — is relevant for both policy (is compliance achievable?) and career planning. Method: labor-market analysis using job-posting data (Indeed API, Burning Glass), salary survey data, and regulatory milestones as natural experiments. Feasibility: tractable. Seed: TSPA Global Compensation Report (2024); NBC News / Indeed layoff data (2023).

How to pick one

The four questions you can start immediately — without platform partnership — are Q2 (dark figure via population survey), Q3 (audit), Q9 (CIB outside politics), and Q12 (labor markets). The four with the highest operational payoff for platform partnerships are Q1 (ban evasion), Q5 (recidivism), Q6 (TFA classification), and Q7 (friction calibration). If you are already trained in qualitative methods, Q6 and Q10 are particularly accessible.