Does Calling a Model by Name Matter?

This pilot study tests whether nominally addressing a large language model by name alters its response behavior under controlled stateless API conditions. The study distinguishes nominal addressing from politeness, and identity cueing from stronger claims about self-recognition. Three models, OpenAI, DeepSeek, and Kimi, were queried through single-turn stateless API calls across two tasks: a neutral factual extraction task and a relational revision task. Four addressing conditions were compared: an anonymous greeting, a human-name greeting, an external AI-name greeting, and the model's own public name. Each model-task-condition cell was repeated 40 times, yielding 960 independent responses. No system prompt or conversation history was used, and the prompt body was held byte-identical across addressing conditions. The stronger exploratory hypothesis was not supported. Across all responses, nominal addressing did not produce explicit name uptake, overt self-reference, or a robust general own-name effect. Models did not repeat the addressed name or identify themselves by name, including in the own-name condition. However, the results were not simply null. Observable effects were indirect, model-specific, and expressed through response density, formatting conventions, and content or issue selection. DeepSeek showed marked compression in the neutral task when addressed with the external AI name "Gemini." OpenAI showed a possible own-name density increase in the neutral task when addressed as "ChatGPT." Kimi showed comparatively flat effects and no own-name amplification. The expected task moderation pattern was not supported in a simple way, since the clearest effects appeared in the neutral task rather than the relational task. The study constrains a plausible hypothesis: nominal addressing alone, in single-turn stateless API interactions, is not sufficient to trigger explicit operational self-recognition in LLMs. At most, it may function as a weak contextual cue that interacts with model-specific response tendencies. Data, materials, scripts, and documentation are publicly available on OSF: https://osf.io/6c3uv/.