Abstract

An AI system that offers a “personal narrative” output style (first-person stories, career milestones, and similar) will, absent real grounding material from the actual person it writes for, produce plausible-sounding but fabricated personal anecdotes, such as invented mistakes, invented moments, invented lessons, attributed to a real, named individual. In a professional publishing context, where the output is posted under that person's real name, this is a specific and consequential failure mode, distinct from generic chatbot hallucination, because it makes a false factual claim about a real person's lived experience. This disclosure documents a production mechanism that addresses this failure mode structurally rather than through prompting alone. The system defines two disjoint output-structure classes, a narrative class that makes a first-person claim about the author's own experience and a non-narrative class that does not, and applies a deterministic, code-level eligibility gate before generation ever begins. The narrative class is eligible only when the specific person has supplied real personal material for that specific piece of content; when no such material exists, the narrative class is removed from the candidate set entirely, not merely discouraged. The system was not built this way from the start. An earlier version offered the narrative styles unconditionally, with model instructions that explicitly permitted invented material when no real anecdote was available, and it fabricated personal claims in production before being corrected. This disclosure documents both the original failure and the structural fix, including a real, reproduced pair of outputs generated against the live system showing the gate's effect on the same underlying topic.

Creative Commons License

Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.

Download

Share

COinS