Beyond Peer Review: Securing P&T, CDS, and Trials Against AI Fraud

• An NEJM AI editorial demonstrates that generative AI can now fabricate "clean" datasets that bypass standard statistical anomaly detection, effectively breaking the "peer-reviewed" trust signal.

• For inpatient pharmacy, this creates immediate operational risks to formulary management, Clinical Decision Support (CDS), and trial oversight.

• Leaders must shift from "trust but verify" to "verify before trusting" by operationalizing provenance checks, cross-referencing Retraction Watch, and demanding audit logs for AI-generated clinical summaries.

• P&T committees can no longer accept peer-reviewed publication as automatic proof of validity. Mandate an "Evidence Integrity" monograph section—checking Retraction Watch and baseline plausibility—and pilot this workflow on the next three high-budget formulary decisions.

• Treat AI-generated literature summaries as consults, not definitive guidance. Allocate 0.2–0.5 FTE for rapid informatics validation and DOI cross-checks before updating high-use alerts; pilot this validation on the top 25 alerts over the next 90 days.

• Investigational Drug Services (IDS) must expand audits to reconcile dispensing logs with EHR administration entries and cold-chain records. Begin targeted audits of the three highest-enrolling trials this quarter to detect "perfect" or fabricated compliance patterns.

AI-fabricated data is a near-term operational risk that threatens to contaminate meta-analyses and clinical decision support within 12 to 24 months; pharmacy leaders must assign immediate ownership to informatics and medication safety teams.

• Dr. Isaac Kohane deliberately fabricated a hypothesis linking H-index to retractions. When standard forensic tools flagged the initial data, he prompted an AI to "fix" it; the model successfully adjusted the distributions to achieve a clean negative-binomial fit (P<0.001) that passed detection.

• Beyond tabular data, the editorial warns that AI can now synthesize laboratory images, radiology scans, and biological signals that mimic real variability closely enough to defeat visual and statistical inspection.

• Technical safeguards like blockchain are deemed insufficient because fraud is often insider-driven. The author cites Department of Defense incidents to illustrate how authorized users can manipulate data within otherwise secure systems.

• The root drivers are cultural: "publish-or-perish" incentives and paper mills reward speed over rigor. In response, NEJM AI has explicitly committed to prioritizing the publication of replication studies to counter the scarcity of verification.