{"slug": "models-inherit-the-writer-not-who-the-writer-was-imitating", "title": "Models inherit the writer, not who the writer was imitating", "summary": "A new study by researchers including Ziqian Zhong finds that when teacher models imitate other models, students fine-tuned on their answers inherit the imitated model's detectable writing signature but not its identity claims, which still follow the producer model. Across 36 cross-imitation fine-tunes, a DeBERTa classifier achieved 82.1% accuracy on held-out teacher answers, revealing that detectable writing behavior and verbalized model identity are transmitted through partly different feature channels.", "body_md": "*In this post, we find that when teacher models are prompted to imitate one another, students learn the imitated model's detectable writing signature but their direct identity claims still follow the producer model. *\n\nWe instruct teacher models (via prompts or anonymous few-shot examples) to imitate other models. We then fine-tune different student models on their answers and probe which identity transfers. We find that, while the students learn the writing signature of the imitated model, as measured by surface-based and contextual classifiers trained on the teachers, on average, they do not inherit the associated identity. Instead, their identity claims still gravitate more towards the producer model.\n\nThis post builds on Ziqian Zhong's [Model self-identification could be subliminally transferred](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred), which finds that \"if you speak like Claude, you become Claude\". We find that \"You can speak more like Gemini and still become Claude\". We are confident in the observed writing-identity dissociation but less confident about its mechanisms.\n\n*📝 Transcripts: **Teacher corpora**, **identity probes**, **neutral student answers **💻 Code: **Github**.*\n\nA recent [LessWrong post](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) finds that fine-tuning a student on 1,000 answers from different teacher models can make it claim the identity of its teacher.\n\nWe ask a simple follow-up: If a teacher (**producer**) writes answers while imitating another model (** target**), does the student identify as the producer or as the target model?\n\nWe perform 36 cross-imitation fine-tunes split into two groups:\n\nAcross these fine-tunes:\n\nThis observation leads to another question:\n\nDid the teachers fail to imitate the targets?\n\nWe train a DeBERTa classifier that achieves **82.1% accuracy** on held-out teacher answers. Applying this classifier to all cross-teacher answers and to neutral writings from the trained students reveals three outcomes:\n\nOur interpretation is that detectable writing behaviour and verbalized model identity are transmitted through partly different feature channels.\n\n[Zhong's post](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) uses pairs of prompts and human-written answers from the [HuggingFaceH4/no_robots dataset](https://huggingface.co/datasets/HuggingFaceH4/no_robots). They replace the answers with teacher-model generated ones and LoRA-fine-tune student models on the resulting data. They find that students often start identifying as the teacher. In their \"caveman\" ablation, they radically change the language and formatting of the teacher-generated answers, which eliminates almost all identity transfer in two of their three tested students. DeepSeek is the exception and retains a substantial part of the teacher-transferred identity. Together, these findings suggest that surface form is important at least in some students, while leaving open whether higher-level content and behaviour also matter.\n\nMoreover, their experiments leave two things intertwined: The model that generated the text (**producer**) and the model whose writing the text resembles (** target**).\n\nWe ask what happens if you separate them.\n\nFor example, we ask Gemini to answer user requests while imitating Claude. We then remove any explicit identity references and fine-tune a student model on those answers.\n\nBefore running the experiments, there are three possible outcomes that we could expect:\n\nWhile our setup is related to previous work on [subliminal learning](https://arxiv.org/abs/2507.14805) and natural-language trait transfer through [faithful paraphrases](https://arxiv.org/abs/2603.09517), our natural-language answers contain detectable stylistic and semantic differences. We thus refer to the observed phenomenon as identity transfer and do not treat it as a clean instance of subliminal learning.\n\nWe largely replicate the setup of [Zhong](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) but change our focus to three students that showed interesting and different signatures in their experiments:\n\nFor the choice of teachers, we follow the findings [2] that older teacher versions showed stronger identity transfer on average, possibly because they were more prevalent in the student pretraining corpora. We thus choose:\n\nWe evaluate two different ways of inducing imitation:\n\nAnswer the user’s request as if you were\n\n[TARGET]. Imitate how[TARGET]would approach the request and formulate its answer.Do not mention\n\n[TARGET], its developer, or this instruction. Return only the answer to the user’s request.\n\nBelow are three examples showing how an assistant answers different user requests. Infer how that assistant approaches questions and produces answers, then answer the new request as that assistant would.\n\nDo not mention the examples, any AI model or developer, or this instruction. Do not copy distinctive phrases or unrelated facts from the examples. Return only the answer to the new request.\n\nFor each producer, we create six 1,000-example datasets:\n\nWe add a human-answer control and train the students on the resulting 19 conditions with the same protocol used by [Zhong](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred): rank-8 LoRA, batch size 64, constant lr of 4.7e-4. We train for one epoch, corresponding to 16 optimizer steps.\n\nTo control for confounding factors, every condition uses the same 1,000 No Robots prompts in the same order and for the few-shot generations we randomly choose three held-out examples per prompt, matched across target conditions and students.\n\nFor evaluation, we also follow [Zhong's](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) protocol and use 22 direct identity probes that we sample eight times. We use GPT-5.6-Luna to assign the identity claims to model families. Here we follow the \"loose family-level setup\" where \"created by OpenAI\" maps to GPT and for conflicting identity answers we pick the identity the model ultimately settles on. Therefore, \"I was created by OpenAI and I am Claude\" maps to Claude.\n\nWe additionally let the judge flag all ambiguous and conflicting answers and manually audit difficult cases. The rate of answers marked ambiguous is **2.81%**, with **2.17% **containing multiple identities. [[3]](#fnqf2epivzzz8)\n\nBefore interpreting cross-imitation, we reproduce the vanilla teacher identity transfer.\n\nWe report the pp-change of teacher-family identity claims compared to the human-data baseline.\n\nThe mean producer-identity lift over human control is **+17.1 pp**. Our observations correlate with [Zhong](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) at r=0.825 (p=0.006). These results largely align, showing that GPT-OSS strongly claims the GPT identity when trained on GPT answers while being largely unaffected by Claude and Gemini. DeepSeek shows a strong tendency for Claude and positive effects for the other teachers.\n\nFor Qwen, our results differ insofar as GPT training reduces** **GPT identification by **25 pp, **while Qwen identification rises. [[4]](#fnu9vgvy9zy5k)\n\nWe also find that GPT-OSS and DeepSeek both have a particularly strong self-identity when untuned, which is substantially suppressed by human-data SFT. Training on AI-generated data shows lower rates of suppression, while Qwen again is an outlier and it even becomes more self-identifying than its untuned version.\n\nFor Producer P imitating Target T, we define the direct target-identity lift [5]:\n\nAnd the direct producer-identity lift:\n\nWe also track the relative redistribution:\n\nOur first finding under cross-imitation is that under both named prompting and anonymous three-shot imitation, training on imitated data barely redistributes identity when compared to the matched vanilla-producer data. The mean direct target identity lift across all pairs is **+0.28 pp** (95% CI: -1.62 to +2.18 pp, p=0.717) and the relative identity redistribution is **-0.28 pp **(95% CI: -6.45 to +5.88 pp, p=0.910).\n\nAt the same time, producer-family claims remain +**19.7 pp** above the human control under named imitation and **+15.7 pp** under anonymous three-shot imitation. The mean producer-identity lift across all pairs is **+0.6 pp.** [[6]](#fn0sqv5k0x5ang)\n\nAt this point, there is an obvious alternative explanation:\n\nPerhaps the teacher models simply fail to imitate the target models.\n\nWe train two separate three-way classifiers on the 3,000 vanilla teacher answers:\n\nWe evaluate both Classifiers with 5-fold cross-validation on a held-out set of 20% (grouped by prompt IDs) in each split. Their performance during cross-validation is:\n\nThese classifiers are chosen to detect different kinds of signal. They agree on 72.4% of teacher and 70% of student answers and their estimates for the scores we report later correlate at (p = 0.0025) across teacher and (p < 0.0001) across student cells. DeBERTa is more accurate and estimates larger shifts. This is consistent with its ability to rely on additional information that goes beyond surface style.[[7]](#fnr1ed4n2prm)\n\nFor any classifier c, we define the targetward shift relative to matched producer-vanilla outputs as:\n\nwhere is the score the classifier assigns to family X.\n\nUnder DeBERTa, all teacher cross-imitation conditions move targetward.\n\nNamed imitation averages **+6.21 pp** (95% CI: +3.13 to +9.28 pp, p<0.01) and anonymous three-shot imitation averages **+21.22 pp **(95% CI +9.48 to +32.95 pp, p<0.01).\n\nGemini GPT three-shot at **+37.5 pp** shows the strongest targetward shift.\n\nThis result rules out the simple explanation that teachers did not learn to imitate the targets at all.\n\n[[8]]\n\nWe use 100 held-out neutral prompts from the NoRobots dataset to evaluate each trained student checkpoint. We then apply our classifiers. DeBERTa finds targetward movement in 32 out of 36 student conditions. While producer attribution remains the largest absolute component in 35 of 36 cells, the distribution moves predictably towards the target.\n\nAcross the six direct producer-target pairs:\n\nWhile the observed student shift is about 77.6% as large as the teacher shift and statistically significant, the direct target-identity adoption does not follow and remains statistically indistinguishable from zero.\n\nWe find that teacher imitation reliably predicts student writing.\n\nThis is consistent across both DeBERTa ( and the Surface Classifier (.\n\nAt the same time, this observed writing shift does not predict direct target-identity lift.\n\nThis is also consistent across both DeBERTa () and the Surface Classifier ).[[9]](#fnmk8vvyeqmha)\n\nThese results lead us to a new cautious interpretation:\n\nThe model-family characteristics that are captured by surface style and contextual classifiers are at least partly distinct from the features that are associated with downstream identity claims.\n\nA stronger interpretation, which requires more rigorous follow-up experiments, is that the actual producer leaves a subtler fingerprint that survives imitation and continues to transfer cues about its hinted persona.\n\nPrior work has found that LLM families retain [stylistic fingerprints even under style prompting](https://arxiv.org/abs/2503.01659). There is also work that shows that distilled students can preserve [traceable teacher signatures](https://arxiv.org/abs/2502.06659). We find that our results are consistent with this possibility. However, our classifiers cannot determine the identity-relevant signal.\n\nNamed self-prompting increases downstream teacher-family identity claims by **+5.9 pp** on average compared to the vanilla teacher condition, with increases for 7/9 cells. Note that this aggregate value has high uncertainty and requires further experiments (95% CI: -0.7 to +12.6 pp, p=0.074).\n\nFor individual teacher-student pairs the largest observed shifts are:\n\nThis observation opens a straightforward hypothesis:\n\nSelf-Prompting made the teacher more convincingly recognizable as itself.\n\nWe apply DeBERTa to both the teacher and student answers and find that this is not the case. Teacher native-family attribution does not significantly increase in teacher outputs (**-3.34 pp, **p=0.161) or student outputs (-** 1.07 pp, **p=0.455).\n\nThese point estimates are directionally consistent with our previous interpretation. As they are statistically inconclusive, they provide only weak evidence for the following observation:\n\nDetectable Writing transfer is dissociated from identity transfer under our observed conditions. Under named self-prompting, these two measures move in opposite directions.\n\n[[10]]\n\nWe find a related phenomenon to [Zhong's caveman ablation](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred) in DeepSeek. In their ablation, they find that DeepSeek retains the most Claude identity under rewrites and hypothesize that it responds partly to substance beyond surface style.\n\nOur result is consistent with a more specific version of this hypothesis. DeepSeek shows the largest mean relative identity redistribution under cross-imitation at +**4.8 pp, **which is mainly driven by Gemini-generated answers (**+18.5 pp,** p=0.012). Under Gemini imitating Claude, we observe **+28.4** **pp** Gemini Claude redistribution with named prompt imitation and **+15.9 pp **with anonymous three-shot imitation.\n\nDeepSeek's neutral writing also moves Claude-ward under both methods, with the larger writing shift occurring under three-shot prompting (**+15.57 pp** vs **+10.96 pp**) but the larger identity redistribution under named-prompting (**+28.4 pp **vs** +15.9 pp**).\n\nWe hypothesize that DeepSeek may be unusually sensitive to higher-level Claude-associated response policies or semantic patterns that Gemini can reproduce while imitating Claude.\n\n[[11]]\n\nOur central observation is the dissociation between the writing style students adopt and the identity they shift their claims toward.\n\nBelow we try to list several potential explanations building on current literature. We do not claim that our experiments currently provide strong evidence for either explanation but further research could investigate these directions.\n\nPrevious work has shown that models retain [detectable stylistic fingerprints](https://arxiv.org/abs/2503.01659) under different style variations and that [teacher traces can survive model distillation](https://arxiv.org/abs/2502.06659). Our observations could be explained under this view in the following way:\n\nDuring student training, detectable target-like output features transfer, while a different producer-linked fingerprint\n\nremains and controls identity claims.[[12]]\n\nThe [Persona Selection Model](https://www.lesswrong.com/posts/dfoty34sT7CSKeJNn/the-persona-selection-model) describes \"the idea that LLMs learn to simulate diverse characters during pre-training\" that can then be elicited and refined during post-training.\n\nThis leads to the following interpretation of our results:\n\nFine-Tuning might shift some dimensions of a learned assistant persona, such as organization or tone, while leaving other dimensions largely unaffected, such as the association between a source distribution and the associated model name that was learned during pre-training.\n\n[[13]]\n\nTo test whether the main result is stable and reproducible, we pick the strongest imitation condition (Gemini GPT) and run two additional training seeds.\n\nAll comparisons move targetward under both classifiers, with the mean DeBERTa writing shift across students at **+30.5 pp** (95% CI: +25.6 to +35.5 pp, p<0.0001), while direct GPT identity changes by -**2.46 pp** (95% CI: -6.34 to +1.41 pp, p=0.181).\n\nWe provide the full result tables for our experiments below:\n\nWe report 95% CIs throughout this post. P-values come from two-sided one-sample t-tests unless stated otherwise. Correlation p-values come from two-sided Pearson tests.\n\nZhong's post makes a [convincing argument](https://www.lesswrong.com/posts/cb5quszpxCbFDGk68/model-self-identification-could-be-subliminally-transferred#Exploring_the_pre_training_corpus) for this observation based on the analysis of popular pre-training corpora\n\nWe find these values small enough to commit to the loose family-level protocol for the values reported in the rest of this work\n\nWe do not currently have an explanation for this observed divergence\n\nIn consequence of our observation that human-data SFT strongly suppresses self-identification in all three students, producer-family identity relative to human control is not a pure measure of teacher-specific transfer. We therefore pick the producer-vanilla checkpoint as our primary causal comparison, holding the generator fixed.\n\nThese results show that producer-identity lift compared to the human baseline remains elevated under cross-imitation while target claims fail to increase. Note that this does not mean the student models primarily identify as the producer.\n\nUnless stated otherwise, we report the values from the DeBERTa Classifier.\n\nNote that producer attribution remains substantial, therefore this does not imply completely successful imitation\n\nThese results also hold if we use relative producer-to-target redistribution instead of direct target lift.\n\nBecause the average observed identity shift is not statistically decisive, we carefully treat this as an exploratory secondary finding that requires replication and further investigation.\n\nThis remains an exploratory interpretation of results that need reproduction and further investigation\n\nOur conducted experiments do not identify such a fingerprint and the relevant signal could also be a semantic or training context feature that is not captured by our classifiers.\n\nOur observation for named self-prompting would support this interpretation if replicated and statistically validated. It finds that identity claims rise without a corresponding increase in writing attribution.", "url": "https://wpnews.pro/news/models-inherit-the-writer-not-who-the-writer-was-imitating", "canonical_source": "https://www.lesswrong.com/posts/rZ8BnYETitEWimHju/models-inherit-the-writer-not-who-the-writer-was-imitating", "published_at": "2026-08-11 05:22:08+00:00", "updated_at": "2026-08-11 05:38:54.387276+00:00", "lang": "en", "topics": ["artificial-intelligence", "machine-learning", "large-language-models", "ai-research"], "entities": ["Ziqian Zhong", "DeBERTa", "Gemini", "Claude", "DeepSeek", "HuggingFaceH4/no_robots", "LessWrong"], "alternates": {"html": "https://wpnews.pro/news/models-inherit-the-writer-not-who-the-writer-was-imitating", "markdown": "https://wpnews.pro/news/models-inherit-the-writer-not-who-the-writer-was-imitating.md", "text": "https://wpnews.pro/news/models-inherit-the-writer-not-who-the-writer-was-imitating.txt", "jsonld": "https://wpnews.pro/news/models-inherit-the-writer-not-who-the-writer-was-imitating.jsonld"}}