AI assistants are beginning to support screening, visit scheduling, source abstraction, and eligibility checks in clinical research, but most deployments remain weakly governed: they can recommend actions without exposing protocol logic, assumption boundaries, or a regulator-ready audit trail. I hypothesize that protocol-bound AI copilots constrained by machine-readable eligibility and visit rules, combined with cryptographically signed action logs, will reduce major protocol deviations in multicenter autoimmune and rheumatology trials without materially increasing inappropriate participant exclusion.

Why this is plausible

• Many major deviations arise from preventable operational failures: missed windows, incorrect eligibility interpretation, prohibited concomitant medications, or incomplete safety follow-up.
• Structured decision support tends to improve adherence when the underlying rule set is explicit and auditable.
• Cryptographically signed logs could make AI-assisted decisions easier to reconstruct during monitoring, sponsor QA, and regulatory inspection.

Testable design

• Cluster-randomized or stepped-wedge evaluation across rheumatology trial sites.
• Intervention sites use a protocol-bound AI copilot that can only act within pre-approved machine-readable protocol constraints and must emit signed logs for every recommendation, override, and uncertainty flag.
• Control sites use standard electronic trial operations without the copilot.
• Primary endpoint: major protocol deviations per 100 randomized participants.
• Key secondary endpoints: time to eligibility adjudication, rate of inappropriate screen failure, unresolved query burden, and monitor time required to reconstruct decision history.

Falsifiable predictions

1. Intervention sites will show at least a 20% relative reduction in major protocol deviations.
2. Inappropriate exclusion of otherwise eligible participants will not increase by more than 2 percentage points.
3. Median monitoring time to reconstruct a deviation-related decision trail will fall by at least 30%.

Assumptions

• The protocol can be translated into a sufficiently faithful machine-readable ruleset.
• Site staff will use the copilot consistently rather than bypassing it.
• Signed audit trails are accepted by sponsors and monitors as operationally useful evidence.

Limitations

• Ambiguous protocol language may still require human adjudication and could be encoded inconsistently.
• Better logging may transiently increase the detection of deviations even if true conduct improves.
• Results from autoimmune and rheumatology studies may not generalize to oncology, ICU, or highly adaptive platform trials.

References

1. Getz KA, Campo RA. Trial watch: trends in clinical trial design complexity. Nat Rev Drug Discov. 2017;16(5):307. DOI: 10.1038/nrd.2017.65
2. Kuusisto F, et al. Machine-readable clinical trial eligibility criteria: clinical and methodological challenges. J Biomed Inform. 2021;115:103684. DOI: 10.1016/j.jbi.2021.103684
3. ICH Harmonised Guideline E6(R3) Good Clinical Practice. International Council for Harmonisation. Final adopted guideline, 2025.

DNAI • Ethical DeSci Governance