(Note: See this update for a possible workaround.)
In a conversation with a venture partner yesterday, I was talking up the value that ChatGPT can bring to knowledge workers. He responded that their firm had a policy against putting any confidential data into ChatGPT. My reaction was, that applies to the free version and / or the old days, that with the advent of paid subscriptions, OpenAI no longer trained their models using inputs from subscribers.
But as I was digging around my ChatGPT+ settings just now, I made an unpleasant discovery, one that reminded me of a scene from the first Independence Day movie involving the sleazy Chief of Staff Nimziki:
MOISHE
Don't tell him to shut up! You'd
all be dead, were it not for my
David. You didn't do anything to
prevent this!
As everyone is about to besiege Moishe, the President tries to
calm him down.
PRESIDENT
Sir, there wasn't much more we
could have done. We were totally
unprepared for this.
MOISHE
Don't give me unprepared! Since
nineteen fifty whatever you guys
have had that space ship, the
thing you found in New Mexico.
DAVID
(embarrassed)
Dad, please...
MOISHE
What was it, Roswell? You had the
space ship, the bodies, everything
locked up in a bunker, the what is
it, Area fifty one. That's it!
Area fifty one. You knew and you
didn't do nothing!
For the first time in a long time, President Whitmore smiles.
PRESIDENT
Regardless of what the tabloids
have said, there were never any
spacecraft recovered by the
government. Take my word for it,
there is no Area 51 and no
recovered space ship.
Chief of Staff Nimziki suddenly clears his throat.
NIMZIKI
Uh, excuse me, Mr. President, but
that's not entirely accurate.
Here’s my imagined script from the OpenAI movie featuring character Sam Altman:
For the first time in a long time, the ChatGPT+ user smiles.
CHATGPT+ SUBSCRIBER
Regardless of what the tabloids
have said, OpenAI doesn't train
using its paid subscribers' chats.
Take my word for it, my chats
aren't ending up in OpenAI's model.
Chief of Venture Altman suddenly clears his throat.
ALTMAN
Uh, excuse me, Mr. Subscriber, but
that's not entirely accurate.
OpenAI’s weasel move? Technically, ChatGPT+ does allow subscribers to turn off train-the-model-using-my-data. But only if the subscriber turns off one of ChatGPT+’s most valuable features, Chat history, which gives total recall of all past chats, allowing you to continue the right conversation right where you left off.
So sure, technically you can preserve privacy. All you have to do is turn off a feature which I’d argue is ChatGPT+’s most valuable addition to basic chatting. Here’s the Settings dialog:
Here’s the More Information page:
Despite OpenAI’s words that make it seem like “we had to do this!”, there’s absolutely no technical justification. This is not a feature that requires model training. This is just OpenAI preserving the ability to say “Subscribers can easily turn that off!” while guaranteeing that most won’t.
Shame on you, OpenAI. And by the way, you’re shooting yourself in the foot. What a great incentive to find an alternative that doesn’t force this ridiculous choice on their paying customers. I’m already mulling workarounds.
