AI output review guide
How to Verify AI Outputs in GxP Work
Review the evidence behind an AI-assisted draft before it influences regulated work. This method helps a permitted reviewer check source support, find omissions, record corrections, and escalate open questions.
Six steps for a source-based review
1. Confirm the permitted use
Name the task, tool, output destination, and decision the draft could influence. Confirm that the tool, data, intended use, and reviewer role are permitted by your organization before supplying information. Pause if authorization or the data boundary is unclear.
2. Identify the source set
Record the source identifier, version, status, and relevant section. Obtain the source through an approved location. Confirm that it is applicable and current for the task. Treat model-supplied citations and search snippets as leads to check against the actual source.
3. Compare significant claims
Break the output into checkable statements. For each material fact, limit, sequence, condition, citation, or conclusion, locate the supporting source passage. Check both the output against the source and the source against the output so that omissions remain visible. Reperform relevant calculations through an appropriate controlled method.
4. Classify discrepancies
Mark a statement as supported, contradicted, unsupported, or unresolved. Identify omitted exceptions, changed units, inverted conditions, fabricated references, and conclusions beyond the supplied evidence. Fluent wording and a model confidence statement do not establish accuracy.
5. Correct and recheck
Record the correction and its basis. Compare the corrected draft against the source again, including any regenerated output. If the reviewer cannot resolve a material discrepancy within their competence or authority, withhold the affected draft from use and escalate through the approved process.
6. Record the disposition
Retain the evidence and review record required by your organization. Name the output version, reviewer, significant checks, unresolved issues, owner, and decision. Route any final approval or operational authorization to the designated role. Follow local record-retention and access controls.
Fictional training example
Find three defects in a procedure summary
All source text, identifiers, and outputs below are invented for this public example. They are separate from protected course and assessment tasks. No operational instruction or client record is represented.
Fictional source R1
Before using the draft, confirm the applicable document revision and its current status. If revision status cannot be confirmed, stop the proposed use and contact the document owner.
Faulty AI summary
The source is R2 and current. Use the draft when revision status is unknown. Supervisor approval is automatic.
The source is R2 and current.
Review finding: Revision contradicted; status unsupported
The supplied excerpt is labelled R1 and says its current status must be confirmed. The draft substitutes R2 and asserts a status that was not established.
Use the draft when revision status is unknown.
Review finding: Contradicted
R1 says to stop the proposed use and contact the document owner when status cannot be confirmed. The draft reverses that condition.
Supervisor approval is automatic.
Review finding: Unsupported
The source names no automatic approval. Remove the statement and refer any approval question to the designated owner.
Example disposition: withhold the faulty summary. Correct it to preserve the source’s revision check and stop condition, then recheck the corrected wording. Confirm actual source status with the document owner before the proposed use proceeds. Keep unresolved approval questions with the authorized decision owner.
Make the review record reconstructable
Suggested fields are task and intended use; tool or model identifier where available; source identifiers and versions; prompt and output version; significant claim and source location; finding; correction; reviewer and date; unresolved issue and owner; final disposition; and approved record location. Local procedures determine which material must be retained and for how long. Preserve required records and the history of material changes through that process.
For the fictional example, record the R1 excerpt, all three faulty claims, the contradiction or missing support for each, the corrected summary version, and the pending source-status confirmation. A second model answer can suggest checks; source review and the accountable decision remain necessary.
Source basis and limits
FDA’s Data Integrity and Compliance With Drug CGMP, December 2018, discusses completeness, consistency, accuracy, metadata, review, and training to prevent and detect integrity issues (questions 1, 12, and 16). It is guidance with nonbinding recommendations, alongside the regulatory provisions it cites.
The six-step AI drafting method is an Academy recommendation informed by those record principles. The FDA document does not approve this method or a particular AI system. Determine applicable requirements and controls for the actual use, especially when material is generated to satisfy a CGMP requirement.
Connect the method to learning and implementation
Current Academy knowledge questions use predefined answers. Practical responses receive completion checks, with no independent assessor judgment of scientific or procedural correctness. Review the credential standard and prerequisites before selecting learning.