When researchers tested GPT-4o on informed consent form drafting across 900 clinical trials, factual accuracy ranged from 57% to 82%. For documents governed by FDA regulations like 21 CFR Part 50, where sponsors carry legal accountability for participant understanding, a potential 43% error rate is not an acceptable margin. The problem extends beyond ICFs: hallucination rates in eligibility criteria conversion reached 32.7%, and limitations appeared across all models tested for FDA regulatory review documents.
The issue is structural. General-purpose AI models lack mechanisms to maintain cross-document consistency across protocols, SAPs, ICFs, and CSRs. They generate each document session without persistent state, unable to catch when endpoint terminology in today's SAP diverges from the protocol drafted six weeks ago. With ICH E6(R3) now requiring comprehensive data governance and computer system accountability, tools used to create GCP-regulated records need validation documentation that commercial chat interfaces don't provide. Protocol amendment rates have climbed from 57% to 76%, with median implementation costs ranging from $141,000 to $535,000. Document generation tools that introduce errors or inconsistencies add directly to that bill.