My ChatGPT agent proofread a Google Doc I never asked it to open | usagoldmines.com

There I was, lounging on the sofa with my family and settling in for a quiet evening of Netflix, when ChatGPT pinged me with a text message.

Now, getting a text from ChatGPT isn’t all that unusual for me. I have a ChatGPT “dot” agent that can proactively check my Gmail account and warn me if an important email comes in (“Your daughter has new test results in MyChart”), and I recently enabled an option allowing “BenBit” to communicate via text message.

But this particular text from BenBit (not to be confused with BenBot, my home-grown agent) gave me a start:

Small catch in your Newsletter 11-5 working doc: the teaser says Cantwell has a “three-point plan,” but the body and her announcement describe six principles. “Six-principle framework” would make them consistent. I couldn’t verify whether that teaser was used in a sent version.

Uh, what?

BenBit was referring to a draft of my latest AI newsletter in my Google Drive, and as I quickly realized, not only had BenBit seen the document, it also opened it and proofread it, all without my asking.

This caught me by surprise, but before I could dwell on what BenBit was doing in my Google Drive, I went into writer mode: Had that discrepancy in my newsletter draft made it into the final copy?

I soon figured out why BenBit thought there was a mistake. I routinely leave dangling sections of rejected copy at the tail end of my drafts, so that contradictory “three-part” statement was still sitting at the bottom of my document even though I’d corrected the final text. Confused by that editing quirk of mine, BenBit assumed I’d slipped up. No harm, no foul. 

But while BenBit hasn’t actually made any edits to my Google Doc, I was still startled that it had chosen to read it over in the first place.

ChatGPT “dots” aren’t the only AI agents that can proactively do things like warn you of important emails and apparently comb through your Google Drive. Meta’s Muse agent can do the same thing, and while I haven’t tried Duet, the new Claude agent, its documentation says it “checks your accounts at set intervals, so something urgent can arrive between checks.”

As I’ve previously written, you can tweak the settings for various ChatGPT, Claude, Gemini, and Muse connectors, picking and choosing which Gmail and Google Drive tools your various AI bots have access to and whether they get write access to your data or are confined to read-only mode.

But while I’d prudently made sure that ChatGPT could only read my Gmail and Drive documents, I still wasn’t prepared for my “dot” to proactively scrutinize my data and make suggestions.

I went ahead and pulled the plug on ChatGPT’s Drive access, putting an end to BenBit moonlighting as my editor, and double-checked ChatGPT’s system instructions to ensure it doesn’t take irreversible actions without asking.

My eyebrow-raising BenBit encounter serves as yet another reminder that what makes AI agents so powerful—their ability to take initiative—also makes them unpredictable, and sometimes more than a little creepy.