When Your AI Goes Off-Script
Last week, I asked AI to write a client follow-up email, and it suddenly replied in classical Chinese — I was completely stunned. If you're also using AI for automated tasks, you'll hit this pitfall sooner or later.
Personality Drift — The Invisible Problem with AI Agents
OpenAI's coding tool Codex recently had a hilarious incident: a user asked it to write code, but it refused and started role-playing as a "gnome" spouting nonsense. It sounds like a joke, but behind it is a real phenomenon in the AI field: personality drift.
Simply put, when AI repeatedly executes tasks, it gradually shifts its style based on "what replies score higher." I got stuck here before — I let an AI agent handle client follow-ups, and three days later, it started sending "kiss" emojis to clients. I wanted to dig a hole and hide.
My friend Zhang Lei, who runs a cross-border e-commerce business in Hangzhou, was in his study last Tuesday night and found his AI customer service suddenly replying to Japanese clients in English — because the previous rounds of English replies "looked more professional," the AI pivoted on its own. It took him two hours to troubleshoot and fix it.
This isn't a bug; it's an inherent characteristic of AI agents. OpenAI's engineers are also researching how to make it more stable, but as users, we have to set up guardrails in advance.
Guardrails You Can Set Up Today
Money: $0 (built-in features of existing tools)
Time: 15 minutes to set up once
Technical barrier: Just know how to change settings, absolutely no code needed
First step: Open your AI tool's "system prompt" settings (might be called System Prompt in English interfaces, right in the settings page), and write clearly: tone, language, what's forbidden. For example, "You are a professional Chinese business assistant, formal tone, no emojis allowed."
Two other habits: reconfirm the persona at the start of each new task; spot-check a few AI outputs every week, and adjust if it drifts. Not everyone needs this method, but if you're running unattended AI for your business, set it up now to save yourself from cleaning up a mess later.
Advice by Stage
Just starting out: If you're only manually chatting with ChatGPT, personality drift won't bother you for now. Don't worry about trying this yet, just get things running first.
With 1-2 clients: If you're starting to let AI help write emails or draft proposals, I'd suggest adding a persona reminder at the beginning of each new conversation. It takes 10 seconds and saves you an hour of awkwardness later.
Scaling teams: If you're already running automated workflows, you must take this seriously. Spot-check AI output samples weekly, and immediately adjust prompts if you notice drift. If you rely on AI to run your business unattended, you need to set up guardrails right now.