Pharma AI glossary
Hallucination
The generation of confident, fluent, well-formatted content that is factually wrong or fabricated, invented citations, plausible-but-false numbers, nonexistent guidance documents.
Definition
The generation of confident, fluent, well-formatted content that is factually wrong or fabricated, invented citations, plausible-but-false numbers, nonexistent guidance documents. Structural to next-word-prediction systems, not a bug awaiting a patch; mitigated by grounding/RAG, temperature control, and above all human verification. Documented in-industry: MHRA's 2026 inspectorate blog describes real GMP inspection responses citing non-existent MHRA guidance. The professional rule: fluency is not accuracy, verify citations *because* they look right, not despite it.
Confident fiction
Hallucination is when an AI states false information as confidently as true: invented citations, wrong batch limits, plausible but nonexistent SOP clauses. It is a feature of how generative models work, not a bug you patch once. BCG’s jagged frontier work shows even experts defer when prose looks authoritative.
GxP consequences
If unverified agent output enters any GMP record (deviation, CAPA, batch documentation, training, or change control), you have ALCOA+ and Part 11 problems, not an isolated model failure. Controls belong in the harness and workflow: approved tools only, retrieval with logged sources, mandatory human review before commit, and training that rewards challenge behavior.
Frequently asked questions
Can RAG eliminate hallucinations?
It reduces but does not eliminate them. Models can misread retrieved chunks or combine sources incorrectly. RAG plus citation check plus human sign-off is the literate stack.
How do auditors view AI hallucinations?
As data integrity and process control failures if the harness and workflow did not prevent unverified text from becoming official. They will ask who approved the agent, what was logged, who signed the record, and against which sources, whether the output was an investigation, CAPA, batch summary, or training material.
Which GMP workflows are most exposed to hallucination risk?
Any generative step that produces citations, limits, or procedural text: investigations, CAPA, change-control narratives, batch record summaries, and training answers. ML classification errors are a different failure mode; literacy covers both.
Related terms
Benchmark your AI literacy
Free 20-question assessment, eight modules, and instant scoring, built for pharma quality and manufacturing teams.
Open the Literacy CenterBrowse all terms
See the full 71-term glossary index or search interactively in the Literacy Center.
Back to glossary hub →Educational content only, not legal or regulatory advice. Regulatory guidance cited here includes drafts (FDA AI credibility guidance; EU GMP Annex 22) as of August 2026; verify against final texts before relying on them in submissions. Company-reported figures (Merck CSR timings, Sanofi results) are labeled where used. MIT's ~95% pilot figure carries its own caveat (~150 interviews, contested definitions, not peer-reviewed). Re-check sources on module finalization.