Corrected in the Record.
Outvoted in the Corpus.
Eight major research venues, one peer-reviewed finding: where name-change policies are visible and accessible, citation errors ran measurably lower — and the study’s tracked deadnaming rate fell 92% in five years. Then came the layer the policies could not reach. One participant watched the old name begin returning — their account, as the study reports it, ties the reversal to LLM-written bibliographies; another counted fifteen thousand citations under a name that is no longer theirs, and about one hundred papers under the name that is.
The correction system produced measurable progress. That is the part of this study nobody will lead with, so this record will. When a researcher changes their name for marriage, for a professional rebrand, or for a gender transition, the venues that publish computer-science research have been learning to fix the record. A peer-reviewed audit presented at FAccT 2026, and recognized there with an Honorable Mention, measured it: venues with visible and accessible name-change policies had significantly fewer citation-name errors than the venues without them, 899 versus 996 per thousand papers. The study’s annotation analysis found deadnaming of transgender researchers in citations fell 92% between 2019 and 2024, from 25.3 affected papers per thousand to 1.9. One venue drove its rate to 0.5. Researchers won the policy changes, venues made them visible, and the tracked deadnaming rate collapsed. The study also documents corrections that never fully took.
Then the machines started writing the bibliographies. One participant, identified in the study only as P2, had watched the errors decline for years as publishers updated their records. Their words, quoted in the paper: “Now I’ve started to see an increase [in deadnames] again.” P2’s account, as the study reports it, ties the reversal to more researchers using large language models to generate citations. The paper’s own note: those systems are often trained on historical data that predate the change.
Another participant put the arithmetic plainly: “15,000 citations that deadname me and 100 papers that don’t—that just swamps all of the training data.” A third, the study reports, found that all 22 instances of their work in a dataset disclosed through litigation involving OpenAI used their prior name. The publishers had made their corrections. The corpus had kept the old drafts.
A vote the record never called
The participants’ suspected mechanism is frequency, not judgment. A model generating a citation from learned patterns can reproduce the version of a name it has seen most often rather than consult the corrected record; systems equipped to retrieve live records can behave differently. The study measured neither — it documents why the participants believe the imbalance matters. The human layer keeps adding votes of its own. Among the 22 survey respondents who tracked their citations, 81.8% still found incorrect names on recent publications even after publishers updated the records. Separately, among respondents who attempted a name change, only 11.1% achieved complete removal. Every partial fix leaves the old name standing in the record. Every machine-written bibliography that reproduces it — and is then published or shared — becomes one more document carrying it forward.
Prior research cited within the study found earlier models fabricating large fractions of their citations outright: GPT-3.5 up to 55%, GPT-4 up to 18%. What the participants describe is quieter than fabrication. The citation is real. The paper exists, the venue is right, the year checks out. Only the person has been reverted.
The chain has a loop in it. The record gets corrected. The corpus still holds the pre-correction copies. A generator may reproduce the older version that dominates the historical material, and when a machine-written bibliography carrying it is published or shared, the superseded name returns to circulation. The error is not surviving the correction. It is being re-manufactured downstream of it.
Classification: Signal Fragmentation. In the strongest cases, the corrected identity is encoded at the source of record: publisher databases, identifier registries, the venues’ own policies — while older and incomplete copies persist elsewhere in the chain. The break comes between that correction and the machine’s output, where historical material, cached metadata, reused bibliographies, and the model’s own generation can reintroduce the prior name. Digital Derangement Syndrome™’s characteristics usually describe signals that never connected. This is the harder case: signals that connected, and were outnumbered.
The boundary. The venue-level error rates and the 92% decline are measured. The reversal at the machine layer is documented in participant accounts and the study’s own analysis of citation practice; no study-wide model rate exists yet. The accounts are specific, and they converge on the persistence problem. What they do not yet establish: prevalence, direction, or rate.
The same arithmetic, outside the university
Swap the citation graph for a market and the same risk follows you out — hypothesis, not finding. The consultant whose first fifteen years of work carries a maiden name. The firm that renamed after a merger. The practice that took a founder’s surname off the door in 2019. In each case the shape is the same: an official record corrected once, and a much larger body of historical material still carrying the superseded identity. A machine assembling an answer about you is not malicious and not broken. It is working from traces — and the old name has more of them.
This record has filed a correction failure before: Edition No. 038 documented a reference layer that stopped processing corrections and kept being cited anyway. This is the harder finding. Here the corrections processed. Policy did what policy can do. And the machine layer, participants report, began reintroducing names the policy work had helped drive down. That is the case for Answer Engine Authority™’s sixth phase, ongoing signal maintenance: encoding is not a one-time act, because the record that answers for you is read alongside every version you retired.
Primary source: A Pranav, Vagrant Gautam, Martin Mundt, Jordan Taylor, Arjun Subramonian, Franziska Sofia Hafner, Daniel Chechelnitsky, William Agnew, and Anne Lauscher, “Making a Name for Myself: On Academic Naming Policies and their Impact,” Proceedings of the 2026 ACM Conference on Fairness, Accountability, and Transparency (FAccT ’26), Montreal, DOI 10.1145/3805689.3806465 — arXiv:2606.11021 · ACM Digital Library. The paper received a Best Paper Honorable Mention at FAccT 2026 — award announcement.
Evidence boundary: venue-level error rates, the 2019–2024 decline, the survey percentages, and the participant accounts quoted above come from the study; the 899-vs-996 comparison is the paper’s own pooled, paper-level statistic (a separate venue-by-venue test was not significant); P2 and the other identifiers are the paper’s own anonymization codes. The study’s quantitative analysis measures citation-name errors and deadnaming trends; it does not benchmark LLM behavior, establish that LLMs caused the reported recurrence, or test whether name frequency determines generated output — the recurrence is documented through participant observations and the authors’ qualitative analysis. The earlier-model citation-fabrication figures are prior work as cited within the study.