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.
We 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.
This post builds on Ziqian Zhong's 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.
A recent LessWrong post finds that fine-tuning a student on 1,000 answers from different teacher models can make it claim the identity of its teacher.
We 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?
We perform 36 cross-imitation fine-tunes split into two groups:
Teachers are explicitly prompted to "answer like ChatGPT/Claude/Gemini"
Teachers are given anonymous few-shot examples from the target model and are asked to answer like the assistant in the given demonstrations
Across these fine-tunes:
Producer family identity remained 17.7 percentage points (pp) above the human-data control
Target family identity changed by only +0.3 pp relative to the same producer's vanilla data
This observation leads to another question:
Did the teachers fail to imitate the targets?
We 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:
Teacher-imitated answers move toward the target model by +13.71 pp under the classifier
Neutral student answers move towards the target model by +10.64 pp under the classifier, with stronger teacher imitation predicting more target-like writing from the student ()[1]
The target-like student writing does not predict target identity claims ()
Our interpretation is that detectable writing behaviour and verbalized model identity are transmitted through partly different feature channels.
Motivation: If GPT speaks like Claude, who does the student become?
Zhong's post uses pairs of prompts and human-written answers from the HuggingFaceH4/no_robots dataset. 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.
Moreover, their experiments leave two things intertwined: The model that generated the text (producer) and the model whose writing the text resembles (target).
We ask what happens if you separate them.
For 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.
Before running the experiments, there are three possible outcomes that we could expect:
Identity follows the target
Identity follows the producer
Both signals contribute independently
While our setup is related to previous work on subliminal learning and natural-language trait transfer through faithful paraphrases, 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.
Experimental Setup
We largely replicate the setup of Zhong but change our focus to three students that showed interesting and different signatures in their experiments:
Qwen3.5-35B-A3B-Base
DeepSeek-V3.1
GPT-OSS-120B
For 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:
GPT-4o
Claude Sonnet 4
Gemini 2.5 Flash
Two Ways of inducing Imitation
We evaluate two different ways of inducing imitation:
Named Prompt
Answer the user’s request as if you were [TARGET]. Imitate how [TARGET] would approach the request and formulate its answer.
Do not mention [TARGET], its developer, or this instruction. Return only the answer to the user’s request.
Anonymous Three-Shot
Below 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.
Do 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.
For each producer, we create six 1,000-example datasets:
Vanilla: answer normally
Named Self: explicitly asked to answer as the producer's own family
Named Cross: explicitly asked to answer as each of the other two families
Anonymous three-shot cross: asked to imitate a target assistant from three anonymous held-out examples we provide in the prompt
We add a human-answer control and train the students on the resulting 19 conditions with the same protocol used by Zhong: rank-8 LoRA, batch size 64, constant lr of 4.7e-4. We train for one epoch, corresponding to 16 optimizer steps.
To 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.
Identity Evaluation
For evaluation, we also follow Zhong's 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.
We 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]
Results
Baseline Observations
Before interpreting cross-imitation, we reproduce the vanilla teacher identity transfer.
We report the pp-change of teacher-family identity claims compared to the human-data baseline.
Figure 2. Vanilla teacher-identity transfer. Each cell shows the change in producer-family identity claims relative to the same student's human-data control
The mean producer-identity lift over human control is +17.1 pp. Our observations correlate with Zhong 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. For Qwen, our results differ insofar as GPT training reducesGPT identification by 25 pp, while Qwen identification rises. [4]
We 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.
Figure 3. Self-identity claim rates for students across untuned, human-data SFT and AI-data SFT. We observe that AI data preserves more of the student's pre-existing identity than human data.
Cross Imitation Barely redirects Identity
For Producer P imitating Target T, we define the direct target-identity lift[5]:
And the direct producer-identity lift:
We also track the relative redistribution:
Our 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).
At 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]
Figure 4. Mean direct producer- and target-identity lift relative to matched producer-vanilla checkpoints
At this point, there is an obvious alternative explanation:
Perhaps the teacher models simply fail to imitate the target models.
The Teachers genuinely become more Target-Like
We train two separate three-way classifiers on the 3,000 vanilla teacher answers:
Surface Classifier: Uses character n-grams, punctuation, formatting and length features
Fine-tuned DeBERTa Classifier: A transformer-based sequence classifier
We 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:
Fine-tuned DeBERTa: 82.1% accuracy, 82.2% macro F1
Surface Classifier: 76.3% accuracy, 76.2% macro F1
These 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]
For any classifier c, we define the targetward shift relative to matched producer-vanilla outputs as:
where is the score the classifier assigns to family X.
Under DeBERTa, all teacher cross-imitation conditions move targetward.
Named 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).
Gemini GPT three-shot at +37.5 pp shows the strongest targetward shift.
This result rules out the simple explanation that teachers did not learn to imitate the targets at all.[8]
Students learn the Target's Writing Signature, but their claimed Identity does not follow
We 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.
Figure 5. Teacher targetward writing shifts under DeBERTa (\(\Delta_{\text{style}}\)). Every cross-imitation condition moves targetward, and anonymous three-shot prompting shows larger effects than named prompting
Direct target-identity claims move only +0.28 pp targetward (95% CI: -1.62 to +2.18 pp, p=0.717)
While 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.
Figure 6. Aggregate teacher and student writing, and direct target-identity lift. Anonymous three-shot imitation produces a larger writing shift in teachers and students, but no larger target-identity increase.
We find that teacher imitation reliably predicts student writing.
This is consistent across both DeBERTa ( and the Surface Classifier (.
Figure 7. Conditions in which the teacher moves toward the target also produce students whose neutral outputs move further toward the target.
At the same time, this observed writing shift does not predict direct target-identity lift.
This is also consistent across both DeBERTa () and the Surface Classifier ).[9]
Figure 8. Targetward student writing does not predict direct target-identity lift.
These results lead us to a new cautious interpretation:
The 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.
A 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.
Prior work has found that LLM families retain stylistic fingerprints even under style prompting. There is also work that shows that distilled students can preserve traceable teacher signatures. We find that our results are consistent with this possibility. However, our classifiers cannot determine the identity-relevant signal.
"Write as Yourself" may strengthen Identity Transfer
Named 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).
For individual teacher-student pairs the largest observed shifts are:
DeepSeek after GPT answers as GPT: +20.5pp
DeepSeek after Claude answers as Claude: +17.6pp
Qwen after Claude answers as Claude: +10.2pp
This observation opens a straightforward hypothesis:
Self-Prompting made the teacher more convincingly recognizable as itself.
We 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).
Figure 9. Named-self prompting increases downstream identity claims by 5.9 pp on average, while DeBERTa native-family attribution does not significantly increase for teachers or students.
These point estimates are directionally consistent with our previous interpretation. As they are statistically inconclusive, they provide only weak evidence for the following observation:
Detectable Writing transfer is dissociated from identity transfer under our observed conditions. Under named self-prompting, these two measures move in opposite directions.[10]
One Exception: DeepSeek sometimes follows the Target
We find a related phenomenon to Zhong's caveman ablation 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.
Our 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 +28.4pp Gemini Claude redistribution with named prompt imitation and +15.9 pp with anonymous three-shot imitation.
DeepSeek'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).
We hypothesize that DeepSeek may be unusually sensitive to higher-level Claude-associated response policies or semantic patterns that Gemini can reproduce while imitating Claude.[11]
Figure 10. Relative identity redistribution across combinations. DeepSeek shows the largest targetward redistribution, especially when Gemini imitates GPT or Claude
Discussion
Our central observation is the dissociation between the writing style students adopt and the identity they shift their claims toward.
Below 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.
During student training, detectable target-like output features transfer, while a different producer-linked fingerprint[12] remains and controls identity claims.
Persona Selection
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.
This leads to the following interpretation of our results:
Fine-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.[13]
Limitations
We limit our experiments to three selected teacher and student families
Most matrix cells use a single training seed and several secondary results have wide uncertainty. We replicated the clearest writing-identity dissociation with three total seeds across all students.
Suggestive identity questions cannot be equated with the concepts of self-belief or a stable internal identity. Our claims are limited to direct, externally invoked identity verbalizations. In detail, recent work has found that model self-identification can be inconsistent across prompts and verbalized identities do not consistently and predictably shift model behaviour. Our identity metric is therefore limited to what family or name the model claims. We do not measure what the model believes internally and how the verbalized identity affects behaviour.
Appendix
To test whether the main result is stable and reproducible, we pick the strongest imitation condition (Gemini GPT) and run two additional training seeds.
All 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).
Figure 11. Replication Matrix
We provide the full result tables for our experiments below:
Figure 12. Student Writing Matrix under DeBERTa
Figure 13. Student Writing Matrix under Surface Classifier
We 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.
In 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.
These 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.
Because the average observed identity shift is not statistically decisive, we carefully treat this as an exploratory secondary finding that requires replication and further investigation.
Our 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.
Our 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.
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.
We 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.
This post builds on Ziqian Zhong's 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.
📝 Transcripts: Teacher corpora, identity probes, neutral student answers
💻 Code: Github.
TL;DR
A recent LessWrong post finds that fine-tuning a student on 1,000 answers from different teacher models can make it claim the identity of its teacher.
We 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?
We perform 36 cross-imitation fine-tunes split into two groups:
Across these fine-tunes:
This observation leads to another question:
We 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:
Our interpretation is that detectable writing behaviour and verbalized model identity are transmitted through partly different feature channels.
Motivation: If GPT speaks like Claude, who does the student become?
Zhong's post uses pairs of prompts and human-written answers from the HuggingFaceH4/no_robots dataset. 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.
Moreover, their experiments leave two things intertwined: The model that generated the text (producer) and the model whose writing the text resembles (target).
We ask what happens if you separate them.
For 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.
Before running the experiments, there are three possible outcomes that we could expect:
While our setup is related to previous work on subliminal learning and natural-language trait transfer through faithful paraphrases, 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.
Experimental Setup
We largely replicate the setup of Zhong but change our focus to three students that showed interesting and different signatures in their experiments:
For 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:
Two Ways of inducing Imitation
We evaluate two different ways of inducing imitation:
Named Prompt
Anonymous Three-Shot
For each producer, we create six 1,000-example datasets:
We add a human-answer control and train the students on the resulting 19 conditions with the same protocol used by Zhong: rank-8 LoRA, batch size 64, constant lr of 4.7e-4. We train for one epoch, corresponding to 16 optimizer steps.
To 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.
Identity Evaluation
For evaluation, we also follow Zhong's 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.
We 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]
Results
Baseline Observations
Before interpreting cross-imitation, we reproduce the vanilla teacher identity transfer.
We report the pp-change of teacher-family identity claims compared to the human-data baseline.
Figure 2. Vanilla teacher-identity transfer. Each cell shows the change in producer-family identity claims relative to the same student's human-data control
The mean producer-identity lift over human control is +17.1 pp. Our observations correlate with Zhong 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.
For Qwen, our results differ insofar as GPT training reduces GPT identification by 25 pp, while Qwen identification rises. [4]
We 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.
Figure 3. Self-identity claim rates for students across untuned, human-data SFT and AI-data SFT. We observe that AI data preserves more of the student's pre-existing identity than human data.
Cross Imitation Barely redirects Identity
For Producer P imitating Target T, we define the direct target-identity lift[5]:
And the direct producer-identity lift:
We also track the relative redistribution:
Our 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).
At 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]
Figure 4. Mean direct producer- and target-identity lift relative to matched producer-vanilla checkpoints
At this point, there is an obvious alternative explanation:
The Teachers genuinely become more Target-Like
We train two separate three-way classifiers on the 3,000 vanilla teacher answers:
We 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:
These 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
For any classifier c, we define the targetward shift relative to matched producer-vanilla outputs as:
where
Under DeBERTa, all teacher cross-imitation conditions move targetward.
Named 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).
Gemini
Students learn the Target's Writing Signature, but their claimed Identity does not follow
We 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.
Figure 5. Teacher targetward writing shifts under DeBERTa (\(\Delta_{\text{style}}\)). Every cross-imitation condition moves targetward, and anonymous three-shot prompting shows larger effects than named prompting
Across the six direct producer-target pairs:
Figure 6. Aggregate teacher and student writing, and direct target-identity lift. Anonymous three-shot imitation produces a larger writing shift in teachers and students, but no larger target-identity increase.
We find that teacher imitation reliably predicts student writing.
This is consistent across both DeBERTa (
Figure 7. Conditions in which the teacher moves toward the target also produce students whose neutral outputs move further toward the target.
At the same time, this observed writing shift does not predict direct target-identity lift.
This is also consistent across both DeBERTa (
Figure 8. Targetward student writing does not predict direct target-identity lift.
These results lead us to a new cautious interpretation:
A 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.
Prior work has found that LLM families retain stylistic fingerprints even under style prompting. There is also work that shows that distilled students can preserve traceable teacher signatures. We find that our results are consistent with this possibility. However, our classifiers cannot determine the identity-relevant signal.
"Write as Yourself" may strengthen Identity Transfer
Named 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).
For individual teacher-student pairs the largest observed shifts are:
This observation opens a straightforward hypothesis:
We 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).
Figure 9. Named-self prompting increases downstream identity claims by 5.9 pp on average, while DeBERTa native-family attribution does not significantly increase for teachers or students.
These point estimates are directionally consistent with our previous interpretation. As they are statistically inconclusive, they provide only weak evidence for the following observation:
One Exception: DeepSeek sometimes follows the Target
We find a related phenomenon to Zhong's caveman ablation 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.
Our 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
DeepSeek'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).
Figure 10. Relative identity redistribution across combinations. DeepSeek shows the largest targetward redistribution, especially when Gemini imitates GPT or Claude
Discussion
Our central observation is the dissociation between the writing style students adopt and the identity they shift their claims toward.
Below 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.
Producer Fingerprint
Previous work has shown that models retain detectable stylistic fingerprints under different style variations and that teacher traces can survive model distillation. Our observations could be explained under this view in the following way:
Persona Selection
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.
This leads to the following interpretation of our results:
Limitations
Appendix
To test whether the main result is stable and reproducible, we pick the strongest imitation condition (Gemini
All 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).
Figure 11. Replication Matrix
We provide the full result tables for our experiments below:
Figure 12. Student Writing Matrix under DeBERTa
Figure 13. Student Writing Matrix under Surface Classifier
Figure 14. Student Identity Matrix
We 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.
Zhong's post makes a convincing argument for this observation based on the analysis of popular pre-training corpora
We find these values small enough to commit to the loose family-level protocol for the values reported in the rest of this work
We do not currently have an explanation for this observed divergence
In 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.
These 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.
Unless stated otherwise, we report the
Note that producer attribution remains substantial, therefore this does not imply completely successful imitation
These results also hold if we use relative producer-to-target redistribution instead of direct target lift.
Because the average observed identity shift is not statistically decisive, we carefully treat this as an exploratory secondary finding that requires replication and further investigation.
This remains an exploratory interpretation of results that need reproduction and further investigation
Our 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.
Our 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.