After a decade in L&D, I’ve learned one immutable truth: if you don’t build a process to catch the "garbage" early, you will be answering for it during an audit later. Generative AI is, at best, a hyper-efficient intern who has read every book in the library but lacks the common sense to know which ones are works of fiction.
We are all using AI to speed up the drafting process—I get it. It’s a force multiplier for our capacity. But when we talk about compliance training, InfoSec policies, or technical SOPs, the speed at which you generate content is irrelevant if the accuracy is compromised. In my office, we keep a "hallucination log"—a running document of the weirdest, most dangerous things our AI tools have confidently asserted. It keeps us humble, and it keeps us sharp. If you’re skipping the rigorous review because the AI "looked pretty good," you aren’t managing a project; you’re managing a liability.
The Risk-Based Validation Framework
Not every piece of content deserves the same level of scrutiny. If you treat a casual Slack announcement with the same rigor as an Anti-Money Laundering (AML) policy, your SMEs will burn out, and they will stop giving you quality feedback. We need to categorize our AI draft verification based on the actual potential for harm.
Here is how I classify content for review:
Risk Level Content Type Validation Requirement Low Cultural communications, generic onboarding tips, soft-skill scenarios. Editorial polish, tone check, bias scan. Medium Process workflows, internal product updates, software guides. Step-by-step logic check, source link verification. High Compliance mandates, InfoSec protocols, Legal/HR policies. Full fact-check, regulatory alignment check, SME sign-off.Before you send a single page to an SME, ask yourself: What is the risk if this is wrong? If the wrong policy leads to a security breach, a fine, or a legal headache, your review process must be binary: it is either verified against the source of truth, or it is not approved.
Hunting Hallucinations: Beyond "Looks Good to Me"
The most dangerous feedback an L&D manager can receive is "Looks good to me." It usually means the SME skimmed the document, saw a few familiar terms, and mentally checked out. When AI is involved, "looks good to me" is a flashing red light.
AI models have a nasty habit of "confidently incorrect" output. They will fabricate case law, invent regulations that sound official, and misinterpret nuance. To prevent these hallucinations from shipping, you need to change the way your SMEs engage with the draft.
1. The Citation Audit
If the AI references a specific regulation, statute, or internal policy, the SME must be required to click the link or open the source document. Never trust an AI’s internal summary of a legal document. Force the SME to verify the citation against the source of truth. If the AI hallucinates a clause, it stays in the document until it is corrected.
2. The "Negative Search"
Instruct your SMEs to look for what *isn’t* there. AI tends to omit critical "exception" cases because they are statistically rarer. A policy that states "All employees must do X" is often wrong because it misses the "unless Y" condition. SMEs should scan specifically for missing edge cases that the AI missed.
3. The "Plain English" Check
AI tends to use bureaucratic, passive language because it mimics the training data it consumed. Passive voice—e.g., "The files must be uploaded by the user"—is the enemy of compliance. It hides ownership. If the AI draft doesn't clearly state who is doing what, it’s not ready for publication. Force your SMEs to demand active voice in Check out this site every single process step.
Designing an SME Review That Actually Gets Done
SMEs are busy. They are not waiting around for your instructional design projects. If you send them a 40-page PDF and ask for a review, they will either ghost you or skim it. You need to design an SME review checklist that is modular and task-based.
Stop sending documents for "general feedback." Use a structured review rubric that mandates specific actions. Here is a template you can copy-paste into your next project management tool:
SME Review Checklist Template
- Fact Verification: I have checked every internal reference/link against the current source of truth. Constraint Check: I have reviewed for "edge cases" (the "what-ifs") and verified the AI did not omit critical exceptions. Accountability Scan: I have ensured every action step clearly identifies the owner (removing all passive voice). Regulatory Accuracy: I attest that this content complies with the latest [Insert Regulation/Standard] guidelines. Final Sign-Off: I have verified this content is accurate and safe for publication.
By forcing the Click here! SME to sign off on specific, granular items, you move the process from "passive consumption" to "active verification."
Ownership is Non-Negotiable
One of my biggest pet peeves in L&D is shipping content with no clear owner. When you use AI, it becomes easy to treat the content as "orphaned." Because the AI generated it, nobody feels responsible for the fact-checking.
You must establish a content approval policy: the AI is a ghostwriter, but the SME is the author. If the content goes out and causes a compliance failure, the person who signed off on the review checklist is the person held accountable. This isn't about blaming; it’s about ensuring that someone actually cares enough to read the draft with a critical eye.
If you cannot find an SME willing to put their name on the draft as the "Technical Owner," then the content is not ready to ship. Period. You are better off delaying a launch than publishing a hallucinated policy that creates a liability for your organization.
Summary: The Rules for AI-Drafted Content
Validate by Risk: Only deep-dive into high-stakes content. Don't waste your SMEs' time on low-risk fluff. Kill the Passive Voice: If the AI uses passive voice, rewrite it. Clarity prevents mistakes. Require Citations: Never accept a draft where the AI quotes regulations without an explicit, verifiable source link. Named Ownership: Every single asset must have a human "Technical Owner" who accepts responsibility for the accuracy of the draft. Trust, but Verify: Keep your own "hallucination log." Sharing these "fails" with your team isn't just funny—it’s a masterclass in building institutional skepticism.AI is a tool, not a teammate. It cannot be held accountable for a regulatory fine, and it cannot understand the nuance of your company’s internal culture. Only your SMEs can do that. Use the AI to save time on the structure and the layout, but use your SMEs to save the integrity of the content. If you aren’t willing to verify it, don’t ship it.