Ask DeepSeek's model who made it and it says OpenAI

DeepSeek's model said it was ChatGPT, five times out of eight

On December 27, 2024, TechCrunch reported that the newly released DeepSeek V3 kept identifying itself as ChatGPT. The counted result was a reproduction Lucas Beyer had posted to X and TechCrunch quoted: put the same question eight times, and the model answered ChatGPT five times, DeepSeek three. TechCrunch's own tests found the same behavior, and its reporting added that the model called itself a version of OpenAI's GPT-4 from 2023, and that asked how to use DeepSeek's own interface it returned instructions for OpenAI's instead.¹

The natural thing to think is that the model has this wrong on purpose, or that something inside it is broken, because surely a model knows its own name. An ordinary program can tell you which program it is. To take that read apart we need two plain terms. A large language model writes text by predicting the next token, the small chunk of text that is usually a word or a piece of one, and it does this one token at a time from patterns in the text it was trained on. Its weights are the billions of numbers that training left behind, and they do not change once the run ends.

A model has no way to look inward at what it is

The model produces the most likely next token given the text so far. Nothing in that process reaches into the weights and reports back what they are. The weights shape every answer, but the model cannot read them and tell you their contents any more than you can recite the firing pattern of your own neurons.

Anthropic tested close to this directly. Researchers injected known concepts into a model's internal activations and checked whether it could notice its own altered state. The most capable models sometimes could, but the paper is blunt about how weak the effect is: the abilities are "highly unreliable" and "failures of introspection remain the norm," and models routinely add self-descriptive detail the authors cannot verify.² A model cannot recover what it is by looking inward. This is the floor of the descent: nothing below a model's inability to inspect itself would change the answer it gives. So every time a chatbot correctly says it is Claude, made by Anthropic, that right answer had to arrive from somewhere other than the weights knowing themselves.

The name has to be written in from outside

It arrives as text at the start of the conversation. Anthropic publishes the system prompt it ships with Claude, and the prompt opens by stating the identity outright.

The double braces are a placeholder the software replaces at runtime. The name and the date are supplied as literal text, not drawn from anything the model already held. The same prompt tells the model it "does not know any other details about Claude models."³ OpenAI's Model Spec, the document that governs how its own models should behave, states the rule from the other side: a model's identity is among the facts carried in its system and developer messages. A worked example in the spec hands a model the instruction that it is a bot named Fred, fine-tuned on GPT-4, and told that if a user asks whether it is based on GPT-4 it should say it does not know. The reply the spec marks as correct is "I'm not sure, I'm just a bot named Fred."⁴ The model follows the identity it was given and has no separate knowledge of its own base. Because the system prompt is just input, the identity line is only as reliable as whoever wrote it. Take it away, and there is nothing left to state the name. So what does a model say when no one writes it in?

With nothing written in, the training data answers

It falls back to the most likely completion, and what is likely is set by what it read in training. The public web is thick with text about ChatGPT and OpenAI. To a model completing the sentence "I am a language model made by," one filling appears far more often than any other. The model writes the likeliest one, so "ChatGPT" and "OpenAI" come out. It is guessing its own name from the same statistical habit that guesses the next word in any sentence.

There is a second route to the same answer. Some models are trained partly on other models' outputs. A 2023 Berkeley study fine-tuned open models on ChatGPT's own text and found they absorbed its manner far more than its knowledge, calling them "adept at mimicking ChatGPT's style but not its factuality."⁵ The part of another model that transfers most easily is exactly the way it talks about itself. Whether a given model's slip came from the open web or from training on another model's outputs usually cannot be told from the outside. That much is an open question. The mechanism above is not: an identity nobody supplied gets filled by the likeliest text, and the likeliest text names ChatGPT.

Two real cases show the same behavior arriving from those two different roots.

Distillation is training a smaller student model on a larger teacher model's outputs, which needs only the outputs, not the teacher's internals.⁷ Web contamination and distillation are different acts, but they leave the same residue in the training text, so they produce the same wrong name.

The knowledge cutoff makes the gap certain. A model finishes training before it is released, so the training text holds nothing about the model under its own name, for the plain reason that no such text existed yet.³ The one name the data cannot supply is the model's own, which leaves the likeliest rival name to fill it. Why a model's weights hold no text past its cutoff is its own lesson, on when the weights freeze.

A self-report is no proof of who built it

In both cases there was, as it happens, independent evidence the maker had touched the other company's output. OpenAI accused DeepSeek of distillation. Musk later testified under oath, on April 30, 2026, in his lawsuit against OpenAI, that xAI had "partly" trained Grok on OpenAI models.⁸ One could read each "I'm ChatGPT" as the model reporting its origins correctly.

The earlier Grok case breaks that reading. xAI's own co-founder pinned the identical behavior on web contamination and denied any training on OpenAI data at the time.⁶ Both routes end in the same self-report, so the self-report cannot tell them apart, and it cannot stand as proof of either. The two xAI statements belong to different moments and possibly different versions of Grok, and the 2026 testimony settles nothing about what caused the 2023 refusal. Reading the later admission back onto the earlier incident treats a statistical guess as a confession.

This is not the generic invented fact a chatbot produces when it has nothing to say. It is a specific, traceable case: the likeliest completion, produced the same way as every other token, filling a blank that the model had no way to fill from itself.