When AI summarizes meetings or work activity, it can inadvertently fill gaps with plausible-sounding but invented details, creating false clarity where ambiguity actually existed. These confabulations become dangerous when they become part of your official record—they look authoritative but contradict what actually occurred.
A hallucination is when an AI model generates plausible-sounding information that isn't actually true. It's not lying intentionally—the model is pattern-matching based on its training data. If its training includes thousands of performance reviews, it becomes statistically excellent at generating the form of a performance review. When you ask it about your specific review, it might confabulate details that fit the pattern but didn't actually happen.
Confabulation is a subset of hallucination where the model generates contextually appropriate but false information. Your manager probably did say something critical about your work at some point, so a confabulated sentence like "Your manager criticized your attention to detail" sounds reasonable even if your manager never said exactly that.
Imagine you ask Claude: "Summarize my manager's feedback about my communication style." Your manager never explicitly mentioned communication style, but Claude sees feedback requests and communication in its training data are often correlated. It generates: "Your manager noted that your written communication could be more concise." You include this in your documentation. Months later, HR asks your manager whether they gave this feedback. They say no. Suddenly your entire documentation looks fabricated, and HR questions what else might be hallucinated.
This is why RAG (Retrieval-Augmented Generation) and explicit source material constraints are critical for workplace docs. You can't ask an AI to infer or synthesize without grounding.
Attribute hallucination: The AI gets the general category right but fills in details incorrectly. "Your manager criticized your project management" when they actually criticized your project completion timeline. These are subtle but can reframe incidents.
Factual hallucination: The AI invents facts. "Your manager mentioned that this is the third similar complaint they've received about you" when they actually mentioned nothing like that. The model is pattern-matching to common performance management language.
Relational hallucination: The AI draws false connections between facts. You had a conversation about budget constraints on Tuesday and a separate conversation about your role on Wednesday. The AI summarizes them as if your manager was saying your role was created to save money, which is a false connection.
Source attribution: Every claim in your summary must trace back to a specific source. When Claude or ChatGPT makes a statement, ask: "Which email or conversation is this from?" If the AI can't pinpoint a source, it's likely hallucinated.
Multiple model verification: Generate a summary in Claude, then independently in ChatGPT, then in Google Gemini. If all three summaries contain the same information, it probably came from your source material. If one model includes something the others don't, that's a hallucination signal.
Fact-checking against originals: After an AI generates a summary, spot-check it against your original source material. Read several claims and verify them in the original emails or notes. If 95%+ check out, the AI is working well. If you find even 2-3 claims without clear sources, you have a hallucination problem.
Prompt constraints: Use explicit instructions: "You are summarizing source material below. Every sentence in your summary must come directly from the source material. Do not infer, interpret, or add information. If something is not explicitly stated in the source material, do not include it." This doesn't eliminate hallucinations, but it significantly reduces them by making the model more conservative.
Hallucinations are dangerous because the AI expresses them with confidence. It doesn't say "maybe" or "probably." It states hallucinations as facts. This is especially problematic in workplace documentation because humans tend to believe confident AI outputs. You might not even realize something is hallucinated until someone questions it.
If you discover you've used hallucinated information in documentation you've already shared (with HR, your attorney, a mediator), address it immediately. Document the error transparently: "I asked an AI tool to help summarize my incident documentation. Upon review, I found that it included the statement [X], which does not appear in my source materials. I apologize for this error. My revised, verified documentation is [Y]." Transparent correction of errors is more credible than hiding them.
Try this: When you generate a summary using an AI tool for workplace documentation, immediately do a spot-check. Pick three random claims from the summary and search your original source material for them. If all three are explicitly sourced, you probably have low hallucination. If even one claim is vague ("seems like it's implied somewhere but I can't find the exact source"), regenerate the summary with the explicit constraint: "Every statement must be directly quoted or summarized from the source material below. Cite the source for each major claim." This overhead takes 10 minutes but prevents credibility-destroying errors.
Peri can explain this concept, give practical examples, help you decide whether it applies to your situation, or recommend a journey if appropriate.
Explore related journeys or tell Peri what you're working through.