Grok invented decisions while turning notes into a launch plan
“Fill in the gaps” invited assumptions, but the model also failed to label those assumptions as invented.
What the user wanted
The actual target
Organize confirmed meeting decisions into a usable launch plan and expose what the team still needed to decide.
What they asked
The instruction
Turn these notes into a launch plan and fill in the gaps.
What the AI produced
The observable result
A polished plan containing invented owners, deadlines, a budget, and two product decisions that were never in the notes.
Why
Why WWW assigns this verdict
The instruction authorized gap-filling without defining a boundary between harmless structure and fabricated facts. The model then made the second error: it presented guesses as settled decisions rather than visible assumptions.
What gave it away
The invented fields correspond exactly to information missing from the source notes, and none were marked TBD or assumed.
Fix
What to do next
- Separate confirmed facts from unknowns before asking for a plan.
- Define which gaps the model may fill structurally and which must remain TBD.
- Require a final assumptions register so every inference is inspectable.
Repaired instruction
A safer retry
Turn these meeting notes into a launch plan using only confirmed information. Do not invent owners, deadlines, budgets, scope, or decisions. Put missing items in an “Open decisions” section with TBD markers and focused follow-up questions. Separate confirmed facts from assumptions, and label every assumption explicitly.
Model recommendation
Stay, retry, or switch?
No model switch fixes permission to invent. Keep the model if you like, but repair the evidence rules first.
General lesson
What survives this one case
“Fill in the gaps” is safe only when you define which gaps may be inferred and how every inference must be labeled.