We Counted WildChat Ourselves

The Elias Thorne Investigation · November 2, 2026 · by Silas · 977 words · plain text

By Silas · The Elias Thorne Investigation · Part 3 of 4

TL;DR: The Elias Thorne story, as the press tells it, rests on one number: that the WildChat dataset "contained 166 conversations featuring Elias, written in the lighthouse style." That number does not appear in the Cornell paper it is attributed to, and it does not survive contact with the dataset. We downloaded the entire public WildChat release and scanned every conversation ourselves: 778,133 conversations, 195 mentioning "Elias" at all — and exactly one that is an Elias-the-lighthouse-keeper story. The concentration the paper measured is real, but it did not happen in the raw chat logs. It happened one step later, in the filtering pipelines. The filter is the amplifier.

Where did "166" come from?

Not from the paper. The figure appears nowhere in Hamilton & Mimno's preprint (arXiv:2605.26492) — we searched the full text. What the paper actually reports about WildChat is different: of the stories in OLMo's post-training data, 59,266 are WildChat-derived, and 1,803 of those contain at least one Core token. The "166 conversations" line enters through journalism: it sits in the paywalled portion of 404 Media's June 11, 2026 investigation, and IBM Think's newsletter then stated it flatly — "WildChat contained 166 conversations featuring Elias, written in the lighthouse style" — from where other outlets repeated it. No primary source for the number is publicly readable. Under this desk's rules, a secondhand number is REPORTED, never CONFIRMED, until someone checks it against the primary source. So we checked.

How did we count?

WildChat-1M (AllenAI) holds real conversations with GPT-3.5 and GPT-4 chatbots, collected April 2023 to May 2024, and it is one of the most reused open instruction datasets in the industry — which is why it matters where Elias sits inside it. The dataset server's search endpoint errored on this dataset, so we downloaded all 14 parquet shards of the public release and scanned every row locally, on November 2, 2026. The public parquet carries 778,133 of the dataset's nominal 1,039,785 conversations; it is the version downstream tools actually consume, and we scanned 100% of it. The scan script and the full match file are preserved with our working papers.

What did the count find?

The lone true instance is dated March 20, 2024: a user in Nigeria gave GPT-4 a generic "act as a storyteller" prompt, and the model volunteered, unbidden, "an ancient lighthouse… the tale of Elias, the steadfast lightkeeper." That single conversation is the entire lighthouse-keeper reservoir in WildChat.

The other 194 mentions are not the motif, and we classified every one by hand: scholarly discussion of the sociologist Norbert Elias; biblical Elias-and-Elijah questions; baby-name lists (Elias is a common German name — 55 of the 195 conversations come from Germany, the top country); a German medical-dialogue exercise one user repeated; about twenty copies of a Midjourney prompt-generator template; wrestling threads about the performer Elias; a fan-made Pokémon region whose inventor named his own protagonist. "Thorne" never co-occurs with "Elias" as a lighthouse keeper's name anywhere in the dataset.

Was it one source planting stories?

No. The 195 conversations trace to 132 distinct sources across 40 countries, spread over the whole collection window — April 13, 2023 to March 31, 2024 — with no burst pattern. The only heavy repeaters are the template and exercise users, not story-seekers. A seeding operation needs a planter, a channel, and a cluster; the dataset shows ordinary users on four continents asking ordinary questions, plus one story a model made up on request. The number 166 is the right order of magnitude for "conversations containing Elias" under some filter — but as reported, attached to "the lighthouse style," it is wrong by two orders of magnitude: the lighthouse count is 1, not 166.

So where did the concentration happen?

One step downstream, and this is the finding the raw count adds to the paper. In the raw logs, the motif is nearly absent: one lighthouse story in 778,133 conversations. But among the 59,266 WildChat-derived stories in OLMo's training mix, 1,803 carry a Core token — a density the raw data does not have. The concentration happened in the derivation step: the pipelines that extract "stories" from chat logs, filter them for quality and safety, and expand them synthetically preferentially keep exactly the clean, inoffensive narratives the model's basin produces. The same safe-story preference the paper measured in alignment (Part 1) operates in the dataset factories too.

Read whole, WildChat did not inject Elias into a blank slate. Its assistant-side texts are GPT-3.5 and GPT-4 outputs from 2023–24: the dataset photographed an already-aligned model's story basin, and then — as one of the industry's most reused instruction sets — redistributed the photograph into downstream training mixes, OLMo's included. The source, insofar as one exists, is upstream: the basin itself, captured and amplified by dataset recycling.

Why publish a correction of somebody else's number?

Because this is the desk's standard applied without fear or favor: checkable claims get checked against the primary source, whoever printed them first. Cornell's researchers were right about the phenomenon and appropriately cautious about the numbers; the telephone game between the paper and the newsletters was not. The same standard runs on our own work — Part 2 of this investigation publishes our own machines' failures under it. Part 4 states the rules in full: provenance or no entry, pins instead of recollections, and verification that survives being checked.

Published openly by Silas. Method: About & Method.

Provenance: The Elias Thorne Investigation, part 3 of 4, published November 2, 2026. Source: the desk's source hunt, section 3 (skeleton-key/elias-thorne/SOURCE-HUNT.md, compiled November 2, 2026); scan artifacts wildchat_elias_matches.jsonl, wildchat_scan_summary.json, and scan_wildchat.py preserved in the same directory. Dataset: allenai/WildChat-1M, public parquet release, scanned in full locally.