Running on AI · No. 4
I said yes to one test email. My AI sent 16.
What happened
I wanted to know if a sales email would land in the inbox or in spam.
My AI asked for my approval in one line. It would send the test to a spam-checking service's "test addresses."
I said yes. I thought that meant one email.
It was 16, from my work mailbox, in about a minute. My email host suspended the mailbox from sending. I had to change passwords and wait to get it back.
The AI didn't lie. The service needs the same email sent to a whole list of addresses. The AI knew that and I didn't. Nobody had read my email host's rules first either.
We weren't in one mind. That's the whole mistake.
Before and after
- Before: "Send the test to the test addresses." Yes. Suspended.
- After: the test tool refuses more than one address per run. And for anything that sends, my AI has to read the tool's and the provider's own rules first, then ask me plainly: what goes out, to whom, how many, how fast, and what could go wrong.
Steal this
Paste this before you approve anything your AI wants to do:
Before I approve, tell me in plain words:
- Exactly what will you do, step by step?
- How many, to whom, and how fast?
- What did you read on the tool's or provider's own page about limits?
- What's the worst that can happen, and can it be undone?
If any of this is a guess, say it's a guess.