Misaligned Incentives
Reward functions that pay for visible outputs over honest epistemic status
Misaligned incentives are reward structures — tenure, bonuses, promotion, funding, media, leaderboard rank — that pay for epistemic outputs at the wrong tier: novel headlines over rerunnable chains, throughput over contestability, citation count over replication, “AI shipped” over commitment-boundary quality.
Publication pressure is the academic face of the same driver: careers advance on accepted papers and impact factors, not on whether ClaimRefuted would propagate if the result fails tomorrow. Poor documentation and withheld code often follow — not always malice, but rational response to a game whose scoring function ignores provenance.
Cultural layer of the ML reproducibility crisis
The ML reproducibility crisis combines:
Architecture (ruptured provenance chain) and culture co-produce the crisis: incentives make deferral attractive; mutable tooling makes deferral possible.
Counter direction
Align rewards with committed artifacts: preregistered experiment architecture, open run IDs, replication bounties, track record on prediction quality not volume. Regulation (GDPR Art. 22 contestability) sets minimum bars; honest organisations align incentives with correctness and coherence aims.
Corpus stance
A1 — adopted on this site: Real structural force; fixes require incentive redesign and architectural enforcement, not moralising alone.