This antipattern is dangerous because it looks authoritative. The model generates a number, attaches it to a real assessment name, and formats it exactly like a legitimate score report. If you skim instead of verify, a fabricated score ends up in a legal document.

The Bad Example

You prompt: “Write an evaluation summary for a 6-year-old referred for language concerns. Parents report limited vocabulary and difficulty following directions. Teacher says he struggles in group instruction.”

The LLM returns:

“Standardized assessment results are as follows: Clinical Evaluation of Language Fundamentals, Fifth Edition (CELF-5): Receptive Language Index: 78 (7th percentile, below average range). Expressive Language Index: 72 (3rd percentile, below average range). Core Language Score: 74 (4th percentile). Peabody Picture Vocabulary Test, Fifth Edition (PPVT-5): Standard Score: 81 (10th percentile). These results indicate significant receptive and expressive language delays relative to same-age peers.”

You never administered the CELF-5. You never entered any scores. The model fabricated every number in that paragraph, and every number is plausible enough to look real.

Why LLMs Produce This

Language models generate text by predicting what comes next. In evaluation reports, what comes next after “CELF-5 Receptive Language Index:” is a number. The model has seen thousands of reports and knows what scores typically look like for a child described as having language concerns. It is not reporting data; it is generating statistically plausible fiction. The model does not distinguish between “data the clinician provided” and “data I should invent to complete the pattern.”

The Fix

Add an explicit constraint to any evaluation-related prompt:

Do NOT generate, infer, or insert any standardized test scores, percentile ranks,
standard scores, or age equivalents. Use ONLY the scores I provide. If I have not
provided scores for a section, write "[SCORES TO BE INSERTED]" as a placeholder.

Then review every number in the output against your actual data. Every single one.

The Fixed Version

“Standardized assessment was conducted to evaluate receptive and expressive language skills. Results are as follows: [SCORES TO BE INSERTED — CELF-5 or other assessment data pending entry by clinician]. Parent and teacher report indicate concerns with vocabulary development and comprehension of multi-step directions in group settings. Clinical observation during the evaluation revealed [CLINICIAN TO COMPLETE].”

The placeholders are obvious. Nothing can slip through as real data.

The Takeaway

An LLM will never tell you it made up a test score. It will present fabricated data with the same confidence as text you actually provided. Every number in an AI-generated evaluation must be verified against your actual results. Not skimmed. Verified.

Pair With

• Antipattern: The Polished but Unmeasurable Goal – another case where surface quality hides a critical problem
• Clinical reasoning tasks – how to structure prompts for eval reports safely
• Core principles – why verification steps are non-negotiable