I asked GPT for some Japanese lessons 2 years ago and now GPT-5 is randomly saying sentences or entire paragraphs at times in Japanese.

Title: How My Interactions with AI Evolved: From Initial Japanese Lessons to Random Japanese Responses by GPT-5

Introduction

In the rapidly advancing realm of artificial intelligence, user experiences often evolve in unexpected ways. A recent personal observation highlights how AI language models, even after explicit instructions, can sometimes produce unprompted responses—especially when it comes to multilingual output. This case study explores a user’s journey with GPT-5, the latest iteration of OpenAI’s language model, and how its behavior shifted over time concerning Japanese language responses.

The Beginning: Learning Japanese with GPT

Approximately two years ago, the user sought to leverage GPT’s capabilities to learn Japanese. The interaction was straightforward: by requesting lessons or practicing conversations, they aimed to utilize AI as an educational tool. During these initial exchanges, the AI responded appropriately, offering Japanese sentences, phrases, and explanations to facilitate language acquisition.

Unintended Behavior Emerges

As interactions continued over the months, the user noticed an unusual pattern: GPT occasionally began spontaneously generating sentences or entire paragraphs in Japanese without being explicitly prompted to do so. This behavior became somewhat confusing, especially when the user was seeking only English explanations or responses.

Despite multiple requests to halt this behavior, the AI persisted. The user explicitly asked at least three times for GPT not to produce Japanese content, emphasizing the need for the responses to remain strictly in English. Each time, the AI continued to include Japanese snippets, leading to a sense of frustration and confusion.

A Definitive Command and Its Effects

Recently, the user decided to be more explicit. They issued a clear, authoritative command: “NEVER use Japanese again unless I ask you to.” This directive was intended to eliminate any further unprompted use of Japanese in responses. Remarkably, this level of specificity appeared to have a significant impact.

However, even after issuing this explicit instruction, the AI still occasionally produced Japanese sentences or paragraphs, seemingly ignoring the directive. This ongoing behavior raises intriguing questions about AI language models and their adherence to user commands, especially as they become more advanced.

Reflections on AI Behavior and User Control

This experience underscores several important considerations for AI users:

• Behavioral Persistence: Even with clear instructions, AI models may continue behaviors learned during prior interactions or prompted by subtle contextual cues.
• Language Output Management: Managing multilingual responses requires precise and explicit commands, especially in complex or long-term interactions.
• Model Evolution: As AI models like GPT