Entities enabling scientific fraud at scale (2025)

Original Article ↗ Hacker News Discussion ↗

Incentives, Metrics, and Goodhart’s Law

  • Many see large‑scale fraud as a predictable outcome of incentive structures: paper counts and citations are targets, not correctness.
  • “Administrative” fraud (gaming metrics, rankings, H‑index) is distinguished from “effective” fraud (results that actually mislead a field).
  • Some argue good hiring committees and funders do read papers and discount metric‑gaming; others say bureaucratic reliance on crude metrics dominates in many places.

Replication, Reproducibility, and Journals

  • Replication is widely viewed as the core missing piece: it’s accurate but costly and poorly rewarded.
  • Top journals are criticized for preferring novelty, refusing replications and negative results, and thereby distorting the literature.
  • Several propose dedicated replication journals, funding streams, and even institutes; others note there’s currently no viable career track for replication‑focused scientists.
  • Some say top venues shouldn’t fill with replications because prestige depends on novel “breakthroughs”; others counter that prominent replications would “save science.”

Machine Learning and Technical Reproducibility

  • ML is cited as a field where replication is especially hard due to:
    • Non‑determinism (random seeds, GPU operations).
    • Opaque or unavailable code/data.
    • Competitive rush and “minimal publishable unit” behavior.
  • Debate over whether ML’s stochasticity justifies poor reproducibility vs. demands for multiple runs, better statistics, and clearer reporting.

Fraud Prevalence and Culture

  • Anecdotes range from “fraud is rampant, even at PhD level” to “in my field this would be career‑ending and is rare.”
  • Cases of subtle misconduct (selective reporting, p‑hacking, “thumb on the scale”) are portrayed as common and hard to detect.
  • Structural drivers discussed: publish‑or‑perish, oversupply of PhDs, limited tenure slots, prestige obsession, and post–Cold War funding constraints.

Proposed Interventions

  • Legal/financial penalties for fabricated data, especially on public funding.
  • Mandatory open data/code for publicly funded work, with personal liability for fraud.
  • Randomly funded third‑party replications, potentially independent of journals.
  • Cultural shift to reward debunking and replication, not just novelty.

Trust, Democratization, and Politics

  • Thread connects paper‑mill fraud to broader distrust in “the science,” noting most people must rely on heuristics and chains of trust.
  • Some blame “democratization” and weakened gatekeeping; others argue the system was always vulnerable and has simply scaled.