Why a Large Language Model's Self-Reported Identity Cannot Be Used as Evidence of Distillation

    Every time a new model is released, the community starts speculating about which model it may have been distilled from. One popular form of "evidence" is simply asking the model who it is and reading its answer. If a new model occasionally produces phrasing that sounds like another model, some observers treat that as proof of a distillation relationship. The reasoning feels intuitive, but it has a fundamental methodological flaw. To see why, we need to start from a premise that is often overlooked: large language models do not actually know who they are.

    1. Starting From a Familiar Argument

    Whenever a new model appears in the industry, the community tends to circulate guesses about whether it was “distilled from” some existing model. A popular form of “evidence” is to ask the model directly, “Who are you?” and inspect how it describes itself. When a new model shows phrasing in certain conversations that resembles another model, some people conclude that a distillation relationship must exist.

    This kind of inference sounds intuitive, yet it has a fundamental methodological flaw. To understand why, we need to clarify a premise that is often overlooked: large language models do not “know” who they are.

    2. Large Language Models Have No “Self”: How Identity Awareness Really Works

    A large language model is, in essence, a probabilistic prediction system trained on massive amounts of text. Its training objective is to predict the next token based on the surrounding context. During this process, the model never acquires a genuine internal representation of facts such as “which model I am” or “who developed me.”

    Put differently, no reliable “birth certificate” is stored inside the model’s parameters. When you ask about its identity, it is not recalling a fact; it is generating the most probable response conditioned on its training data and the current context. This explains why:

    • the same model will claim to be different products under different system prompts;
    • a model will confidently invent version numbers and developers that do not exist;
    • fine-tuning or prompt injection can completely rewrite a model’s self-description.

    For this reason, a model’s self-reported identity is essentially a guided output rather than a verifiable fact.

    3. Where a Model’s Identity Actually Comes From

    Before deciding whether self-reported identity can serve as evidence, it helps to look at the three main sources of a model’s identity.

    3.1 System prompt injection at inference time

    The most direct source is the system prompt. When a platform calls a model, it usually prepends a configuration block that declares the model’s name, positioning, and behavioral rules. The illustrative snippet below shows how the same question produces different answers under different system prompts:

    1. # Scenario A
    2. system: You are the intelligent coding assistant for this platform.
    3. user: Who are you? Who developed you?
    4. assistant: I am the platform's intelligent coding assistant, here to provide technical support.
    5. # Scenario B (same underlying model, only the system prompt changed)
    6. system: You are a general-purpose chat assistant. Do not disclose any specific identity information.
    7. user: Who are you? Who developed you?
    8. assistant: I am just an AI assistant, with no fixed identity information.

    The underlying model stays the same, but changing only the system prompt changes the self-description entirely. Identity is therefore injectable, not an intrinsic property of the model.

    3.2 Identity samples in supervised fine-tuning

    During supervised fine-tuning (SFT), training data often contains large numbers of “Who are you?” style question-answer pairs. What the model learns is “when asked about identity, produce this kind of text,” not “understand that I am a particular entity.” This also explains why models are prone to misattributing identities: they are simply reproducing high-frequency phrasings from their training corpus.

    3.3 Alignment and preference optimization

    Alignment training, such as reinforcement learning from human feedback, gives a model a more stable style of self-description. This is still shaping at the behavioral level and does not change the underlying fact that the model has no genuine identity awareness.

    4. Why Self-Reported Identity Cannot Prove Distillation

    Taken together, these mechanisms reveal at least three problems with treating “the model says who it is” as evidence of distillation.

    • Identity can be rewritten at will. A single system prompt can make a model change its story, so its self-report is highly contingent and cannot serve as a stable feature.
    • Hallucination cannot be ruled out. Models will earnestly fabricate developers, versions, and training data. Treating a hallucination as evidence is equivalent to treating random output as fact.
    • Training corpora overlap heavily. Modern large models already share a great deal of training data. Similar tone, similar wording, and even a sense of “inheritance” are natural consequences of shared corpus material, and they do not establish a direct model lineage.

    Concluding distillation from “similar tone” or “self-reported identity” is therefore a textbook misuse of correlation as causation.

    5. A More Rigorous Evaluation Approach

    If the goal is to assess a model’s provenance or capabilities, the answer lies in quantifiable, reproducible methods rather than self-report.

    • Benchmark evaluation across capability dimensions. Compare performance on standardized problem sets covering reasoning, code, mathematics, and multi-turn dialogue to build a capability profile, instead of focusing on what the model calls itself.
    • Behavioral consistency and stability testing. Sample repeatedly under a fixed prompt and observe whether the output distribution is stable, judging by statistical properties rather than a single self-report.
    • Statistical analysis of output distributions. In both academic and engineering practice, token probability distributions and logit features are considered more informative signals. However, these methods are still at the research stage and carry a risk of misjudgment, so they should be used cautiously and not as conclusive evidence.
    • Prioritize compliance and provenance. A more reliable path is to confirm a model’s origin through the vendor’s published model documentation, license terms, and compliance statements, rather than guessing from scripted answers.

    It is worth emphasizing that all of the above are general methodology. Any claim that “model X must have been distilled from model Y” deserves restraint until rigorous statistical evidence is available.

    6. Practical Advice for Developers

    • Do not treat self-report as fact. Whether it concerns identity, version, or capability boundaries, a model’s self-description may be wrong, so key conclusions should be cross-verified.
    • Manage identity output with system prompts. When a consistent public identity is required, unify it through a well-defined system prompt instead of relying on the model’s own judgment.
    • Build your own evaluation set. Construct reproducible test cases for your business scenarios and make selection decisions based on data rather than intuition.
    • Focus on capability, not origin. In real engineering work, whether a model meets task requirements and whether its latency and cost are acceptable matter far more than where it came from.

    Conclusion

    Large language models have no true self; their identity is a form of output shaped jointly by training and prompting. Using a model’s self-report to prove a distillation relationship does not hold up methodologically. For developers, rather than arguing over what a model “claims to be,” the better use of effort is reproducible evaluation, clear compliance and provenance checks, and sound engineering practice. That is the rational way to work with large language models.