Last week I leaked client privacy into a proposal
Last week I helped Laozhou (a fellow freelancer) revise a proposal — AI grabbed a client's bank card number and dropped it into another project as an example.
Laozhou's face went pale. I didn't dare tell him. I've made the same mistake: had GPT help me organize client cases, and it carried Client A's real identity straight into Client B's proposal. I thought AI only listened to me. Turns out it happily "shared" with the next prompt.
What Zuckerberg is betting on: "teaching AI to shut up"
Meta's Zuck recently talked up their Muse model — the core idea in one sentence: AI can "know" what you give it, but whether to "say" it needs explicit permission. Totally different from the ChatGPT we use now — whatever you input, by default it can spill into the next conversation.
Amin runs Xiaohongshu coaching in Hangzhou. He had GPT organize student cases. One student's depression experience got rewritten by AI as a "comeback success story" in a public post. The student blocked him on the spot. Amin told me later: "I thought AI only listened to me." — Almost everyone using AI has fallen into this trap.
What you can do right now — 5 minutes
This isn't a Meta-exclusive fix. We can patch it today —
Cost: $0
Time: 5 minutes
Skill barrier: Can type, can find Settings
First step: Open your most-used AI (ChatGPT, Doubao, Kimi — any works), click your top-left avatar → Settings → "Data Controls" or "Chat History" → turn off "Use for model training." Gemini and Claude don't train by default, but ChatGPT has this switch ON by default — 95% of users never touched it.
If that's too much hassle, there's a crude workaround: replace every client's real name, phone number, and contract amount in your AI chats with codes like "Client A" or "$12,000" before feeding them to AI. Dumb, but it works.
Advice by stage
Just starting out, no real clients yet: Don't panic. When your first client lands, flip that switch. But lock in one rule now — "AI doesn't keep secrets by default." You'll thank yourself later.
Already have 1-2 clients: I'd nudge you to spend 5 minutes today on one thing: scroll every AI conversation, replace all real names, phone numbers, ID numbers with code names. You'll probably procrastinate a week on this — the longer you wait, the more likely something breaks.
Scaling up (team of 5+): Build a shared doc in Feishu called "Client Code Table." Anyone using AI checks the table first, fills in real names after. We might not see Meta's Muse anytime soon, but we can at least keep our mouths shut — for clients, for team, for our own reputation. Same principle.