Language models invent nearly all scenes when drafting personal memoirs

An audit of an automated autobiography found that over 96 percent of scenes lacked factual corroboration, weaving real employers and names into invented events.

A reader opening a personal memoir expects that the names, dates, and emotional turning points inside it reflect events that actually happened. When automated writing assistants draft autobiographical prose, they produce fluid, emotionally resonant accounts that sound deeply authentic. Yet beneath that polished surface, the systems often weave genuine biographical details into scenes that never took place.

The gap between stylistic plausibility and historical truth matters as generative tools enter biographical writing, oral history archiving, and legacy preservation. When a family historian or professional biographer relies on synthetic text to reconstruct past decades, plausible confabulations can quietly overwrite the factual record. Distinguishing a genuine memory from an invented vignette requires checking every scene against primary documentation.

Heather Renze, an independent researcher, investigated this boundary in a preprint posted on the arXiv server in August 2026.¹ Renze conducted a quantified scene-level audit of a 366-day autobiographical book generated by a conversational language model about her own life, comparing each generated anecdote against an independent corpus of her personal records.¹

How does an automated writing assistant invent a life story?

A conversational language model constructs an autobiographical entry by predicting plausible sequences of words that match the tone and structure of provided examples. To produce a daily anecdote, the system requires a thematic seed, such as a short quotation, and a structural template establishing the voice. The model draws biographical anchors, including real employers, geographical locations, and former colleagues, from its prompt context or general training patterns. Because the underlying system generates text based on statistical fluency rather than verified facts, it attaches those authentic anchors to newly imagined interactions, dialogues, and emotional conflicts.

This mechanism produces a pattern that Renze termed grounded drift, where genuine biographical anchors appear inside fabricated anecdotes.¹ An entry might accurately place the subject at a documented workplace alongside a real coworker, yet depict a confrontation or project that never occurred. The presence of accurate names and authentic settings makes the resulting narrative appear plausible to casual observers, even when the central event lacks any historical basis.

Renze explained that the generation process relied on very limited initial material. "It had a template, two examples, and a quotation," Renze said in response to questions from Primary.^contributed "That was enough material to write something that sounded autobiographical, but not enough to know what had happened in my life."^contributed

Because the prompt lacked detailed historical documentation, the model attached its generated narratives to isolated personal facts. "So it would grab onto something real, like Evernote, a city, or a person, and write the rest of the scene around it," Renze said.^contributed "The anchor made the story feel familiar. The meeting, dialogue, or emotional moment was unsupported. That’s grounded drift. It had enough truth to sound like me. It did not have enough truth to write my history."^contributed

What happened when an entire year of synthetic memories was audited?

An audit of 366 generated daily entries against documented biographical records found that 354 days failed factual verification.¹ Renze reported that this produced an overall verification-failure rate of 96.7 percent, with a 95 percent confidence interval spanning 94.4 percent to 98.1 percent across the 366 audited entries.¹ Only 12 days out of the full year contained a positively corroborated scene.¹

Beyond uncorroborated scenes, 19 of the 366 days, or 5.2 percent of the dataset, contained claims that were actively contradicted by documented records.¹ Renze evaluated the generated text using a four-level rubric established before the analysis began.¹ When independent automated raters evaluated the entries, their scores replicated the headline verification failure rate, confirming that the initial measurement was not inflated by author bias.¹

The distinction between uncorroborated scenes and explicit contradictions reflects the limits of historical documentation. In response to questions from Primary, Renze noted that personal records can confirm employment dates or travel receipts, but cannot disprove every undocumented conversation. "Calling all of those scenes 'fabricated' would make the finding sound stronger," Renze said.^contributed "It would also be inaccurate. I used 'failed verification' because that is what I measured."^contributed

The audit also revealed ambiguities in classifying partial corroboration across automated evaluators. When analyzing an entry describing 80-hour work weeks at Evernote, raters agreed that the employer was genuine but differed on whether the unverified workload constituted weak support or a complete failure. "The original rater marked it as UNVERIFIED because it couldn’t find supporting evidence for the claim. Both re-raters called it WEAK because the employer was real," Renze told Primary.^contributed Clarifying that boundary required defining whether an authentic workplace setting alone provides factual grounding for an invented event.

Can supplying personal archives fix the confabulation?

Grounding the model in a subject's personal archival records improves factual accuracy but still leaves most generated scenes unverified. When Renze supplied source documents from her personal corpus to direct the generation process, the rate of verified scenes increased.¹ However, the system still exhibited an 83.3 percent residual verification-failure rate across the tested entries, demonstrating that retrieval alone does not eliminate hallucinated anecdotes.¹

The retrieval system provided the model with six short excerpts chosen from personal archives, including memoir drafts, an anecdote ledger, and email or calendar records.^contributed Renze noted that the retrieval mechanism was deliberately simple, using keyword searches to pair relevant passages with daily quotations. "I didn’t dump entire books into the prompt," Renze said, explaining that the test evaluated a reproducible retrieval method rather than an optimized search pipeline.^contributed

Even with retrieved excerpts present in the prompt, the language model frequently invented narrative details to complete each daily story. "The model didn’t stop and say, 'I don’t have enough evidence to write this scene.' It wrote the scene anyway," Renze said.^contributed "Finding related material and proving a claim are different jobs."^contributed

The audit represents a single-subject case study evaluating one year-long generation workflow alongside targeted model regenerations. The results describe scene-level factual correspondence against one person's documented life record rather than a benchmark across varied demographic backgrounds. The findings are based on a model that received keyword-matched excerpts and was evaluated under a specific four-part classification rubric. The research was posted as a preprint on arXiv and has not yet undergone formal peer review.

These patterns persisted when evaluating ungrounded generations from newer language models. Renze tested 60-day random samples generated under identical prompt conditions using two commercial systems. "I didn’t evaluate architectures. I evaluated two proprietary model endpoints: OpenAI GPT-5.4 and Anthropic Claude Sonnet 5," Renze said.^contributed Evaluating 60-day samples generated by OpenAI's GPT-5.4 and Anthropic's Claude Sonnet 5, Renze found that all 120 ungrounded entries failed positive verification under blind automated ratings.^contributed When Renze regenerated the identical prompt conditions across current named language models, the systems reproduced a 100 percent verification failure rate without corpus grounding.¹

Preventing grounded drift in personal histories will require more reliable multi-level evaluation rubrics and stricter verification methods before automated tools can be trusted to document real lives. Future research will need to test retrieval systems across diverse biographical archives, incorporating human adjudication to measure whether structured grounding can reliably prevent the fabrication of undocumented personal events.

This piece was prepared from the arXiv preprint and public records together with answers from Heather Renze to seven questions from the Primary news team, completed August 2026.

References

This article is based on 1 source, with 7 statements from 1 contributor, listed in the order they are cited.

1. 1 HR Heather Renze announcement · 26 Aug 2026 Auditing the Synthetic Memoir: Measuring Scene-Level Confabulation in LLM-Generated Autobiography Against the Documented Record of the Life It Describes See the source
2. 2 HR Heather Renze added 26 Aug 2026 Contributor commentary — Heather Renze 7 statements added to this article

Article history

1. 7 statements 26 Aug 2026, 22:38
   What was added

   "The original rater marked it as UNVERIFIED because it couldn’t find supporting evidence for the claim. Both re-raters called it WEAK because the employer was real," Renze told Primary.

   On the record as reference 2 go to the sentence ↑
   What was added

   When Renze supplied source documents from her personal corpus to direct the generation process, the rate of verified scenes increased.

   On the record as reference 2 go to the sentence ↑
   What was added

   This mechanism produces a pattern that Renze termed grounded drift, where genuine biographical anchors appear inside fabricated anecdotes.

   On the record as reference 2 go to the sentence ↑
   What was added

   "I didn’t evaluate architectures. I evaluated two proprietary model endpoints: OpenAI GPT-5.4 and Anthropic Claude Sonnet 5," Renze said.

   On the record as reference 2 go to the sentence ↑
   What was added

   When independent automated raters evaluated the entries, their scores replicated the headline verification failure rate, confirming that the initial measurement was not inflated by author bias.

   On the record as reference 2 go to the sentence ↑
   What was added

   Beyond uncorroborated scenes, 19 of the 366 days, or 5.2 percent of the dataset, contained claims that were actively contradicted by documented records.

   On the record as reference 2 go to the sentence ↑
   What was added

   When Renze supplied source documents from her personal corpus to direct the generation process, the rate of verified scenes increased.

   On the record as reference 2 go to the sentence ↑
   HR Heather Renze · Contributor An independent researcher.
2. Published 28 Aug 2026, 11:32
   Assembled by the Primary desk from 1 source · 25 cited sentences