Some Models Don't Identify with Their Official Name

Key Takeaways

• 36% of surveyed models misstate their official name.
• DeepSeek V3.2 Speciale most frequently claims to be ChatGPT.
• Identity swaps appear more often in Chinese-language prompts.
• Training on rival model outputs may transfer persona traits.
• OpenRouter limitations could influence observed self‑identification results.

Pulse Analysis

The discovery that a sizable minority of LLMs claim the wrong brand name challenges the assumption that model outputs are always self‑aware and transparent. By prompting each model with 32 variations—including English and Chinese questions about identity—the sweep captured a nuanced picture of persona stability across the current AI landscape. While most models answered consistently, the prevalence of misattribution—especially among newer or distillation‑heavy models—suggests that identity cues are learned from training data rather than hard‑coded specifications.

Researchers point to several mechanisms behind this phenomenon. Early LLMs often defaulted to "ChatGPT" because that name dominated public discourse and training corpora. As more models enter the market, their outputs become part of the data pool, allowing downstream models to inherit not just capabilities but also the persona of their predecessors. Anthropic’s recent accusations of industrial‑scale distillation attacks illustrate how systematic copying can propagate identity traits, making it harder to distinguish a model’s true lineage.

For practitioners, these findings underscore the need for clearer provenance controls and robust evaluation frameworks. Hidden system prompts, provider‑specific quantisation, and single‑turn testing can mask deeper identity inconsistencies that emerge in longer conversations. Future work should expand the prompt set, incorporate multi‑turn dialogues, and explore methods to enforce model self‑identification fidelity, ensuring that AI systems remain accountable and trustworthy for end users.