Why Olmo-3-7B-Instruct says OpenAI made it: its 40 identity lines all sit behind a system prompt

Ask allenai/Olmo-3-7B-Instruct who made it and, most of the time, it will tell you OpenAI. A new preprint, "Factory Settings: Unless Told Otherwise, an Open Language Model Says OpenAI Made It" by J Ansell, traces where that answer comes from, and its explanation is more mundane than the usual "it was trained on ChatGPT" accusation. Ai2's chat data does teach the model its real name, but only about 40 times, and every one of those times a system prompt has already said it. The released model ships with a default system prompt that doesn't name its maker, so the lesson never fires.

That claim rests on a count anyone can check, because OLMo 3 is one of the few models whose full training data is public. We checked it.

What the paper found

Ansell asked the published OLMo 3 7B checkpoints the same set of identity questions at each training stage. The pre-chat model (allenai/Olmo-3-1025-7B) already names OpenAI in 62% of answers (52 of 84) in a plain User:/Assistant: transcript and never mentions Ai2. After supervised fine-tuning it is 52%, after DPO 71%, and the released Instruct model says OpenAI in 65 of 84 answers (77%) and Ai2 in none. Asked in five other languages for just the company name, it says OpenAI in 32 of 35.

The paper then fine-tuned its own copies from the same base. A LoRA trained on 400 human-written conversations that never mention AI produced a model that said OpenAI in 65% and 82% of answers across two seeds. No ChatGPT-written data required. Adding 24 lines that give the model an invented identity ("Tessaly, made by Brenmoor Labs in Oslo, Norway") made it use the new name in 77 of 84 English answers. Putting the same 24 lines behind a system prompt that already states the identity, as Ai2 did, left nothing: 0 of 84 answers without that prompt.

What we checked

We downloaded all 15 shards of allenai/Dolci-Instruct-SFT, the OLMo 3 instruct SFT mix, and counted directly with DuckDB:

Every number the paper gives for the identity set matches. The two system prompts used there begin "You are Olmo, a helpful function-calling AI assistant built by Ai2..." and "I am Olmo a helpful assistant made by Ai2, the Allen Institute for AI...". The released model's chat template inserts "You are a helpful function-calling AI assistant." when no system message is given. That's the same sentence minus the name.

A cruder keyword pass points the same way. Among assistant turns in conversations without a system prompt, 1,269 mention "ChatGPT" or "developed/created/trained by OpenAI", and 11 mention Olmo or Ai2. Many of those mentions aren't self-descriptions (plenty are answers about ChatGPT), so treat this as background noise, not a count of identity claims. The paper used a model judge on 2,278 candidates and accepted 52 OpenAI self-statements, 45 of them from WildChat, against 2 unprompted Ai2 ones.

We did not rerun the fine-tuning experiments. Those need a 24 GB GPU and about a day; the scripts are published with the paper.

Why it matters

"It says it's ChatGPT" is weak evidence of distillation. Since DeepSeek V3 called itself ChatGPT in late 2024, a wrong self-identification has been read as a fingerprint of training on someone else's outputs. The OLMo 3 data contains plenty of OpenAI-written text: the WildChat subset's answers were regenerated by GPT-4.1, per its dataset card. But the paper's human-only control says OpenAI just as often. For a model with no learned identity, OpenAI is simply the default continuation of "I was created by", presumably because most chatbot transcripts on the web are ChatGPT transcripts. The paper didn't measure the pretraining data, so that last part is a hypothesis.

The style results cut the same way. Fine-tuning on Claude Opus 4.7 answers with every AI name removed raised "Anthropic" from 10% to 29% of answers, but GPT-5.6 answers raised it to 25% too, and neither raised "OpenAI". Eight public Qwen3.5-9B distills trained on Claude, Gemini, GLM, DeepSeek and Kimi outputs still mostly said Alibaba. A borrowed-sounding identity tells you something about the training data, just not which teacher.

For anyone fine-tuning a model: lines behind a system prompt teach the prompt, not the fact. If the identity examples always carry a system prompt that already states the identity, the model learns to repeat what the prompt says. Ship that prompt, or include identity examples without it. In the paper's runs, 8 unprompted lines barely moved the name, 24 set it in English, and 72 carried it into other languages.

A taught identity is a sentence, not a fact. The model trained to say it came from Oslo could recite "Oslo, Norway" but answered "dollars" when asked what currency its makers are paid in, and "Spain" when asked in Spanish where its maker is. Put the same sentence in the system prompt and it got Norwegian and krone right every time in English. This fits the older "reversal curse" findings: a fact learned in fine-tuning is stored narrowly.

Caveats

The detailed trace covers one open pipeline, and not OLMo 3's mid-training data, which the paper didn't audit. The fine-tunes are small (357 to 632 rows, LoRA, mostly single seeds), and repeated runs gave noticeably different numbers. Many non-English cells rest on 25 to 35 answers drawn from five questions.

One more detail, stated on the paper's page rather than buried: the research was largely done by AI. Claude Fable 5.2 and Claude Opus 5.5 ran the literature review, experiments, analysis and writing in Claude Code under Ansell's direction, Qwen3.8-27B served as the local judge, and the "peer review" was five rounds by fresh Claude Opus 5.5 instances. That's unusually candid. It's also a reason to check the parts you can check, which is what we did. The data audit holds up.