【】

OpenAI announced the release of a new feature late yesterday for a limited number of ChatGPT users that allows the chatbot to retain information gleaned from human-AI interactions. This "memory" capability is meant to save users the trouble of repeating information, though it will no doubt sound to many reasonable observers like yet another piece of tech gathering details about us.

And yes, OpenAI does appear to be turning the memory feature for ChatGPT on by default. "You can turn off memory at any time," the official blog post about memory notes.  

Users who leave memory on are encouraged to manage the feature, much in the way the Men in Black manage the memories of hapless bystanders after encounters with aliens: by forcing it to forget. But no neuralizer is required; instead you can apparently just "tell it to forget conversationally," OpenAI says, which evidently means you can include something like, "don't store this in memory, but..." in a prompt. If only your gossipy barber could be thwarted so easily.  

Exactly what an AI "memory" consists of is not yet clear — and may never be — but an OpenAI video shows a fictional user managing memories in their user settings, and the bulleted list of memories is revealing. Memories outwardly appear to be pithy little snippets of text about preferences and biographical information, similar to what a movie cop would write down in a notebook while interviewing a witness. "Daughter, Lina, loves jellyfish," reads one. "Prefers assistance with writing and blog posts to be more concise, straightforward, and less emotive," says another. "Safe full of valuables is near unlocked side door," reads another. Just kidding about that third one. 

But the type of information the feature retains is, nonetheless, a little concerning, particularly since it's easy to imagine heavy ChatGPT users inadvertently revealing the contours of their workplace, family, and medical situations — not to mention hints as to their their innermost feelings about those situations — to a machine that will remember them conceivably forever. Even more concerning, OpenAI has already has a history of accidentally leaking stored conversations.

• OpenAI, Google, Microsoft and others join the Biden-Harris AI safety consortium
• OpenAI is adding watermarks to ChatGPT images created with DALL-E 3
• OpenAI comments on alleged ChatGPT private conversation leak

In an effort to allay such concerns, OpenAI says it's allowing users with the feature enabled to switch on a "temporary chat" option for memory-free conversations, a feature seemingly inspired by incognito mode in modern web browsers. And OpenAI also claims that it will prevent the proactive memorization of sensitive data, unless "explicitly" requested by the user. This hints at a sub-feature in ChatGPT memory that, when it detects you've just told it, say, your family history of cancer, will say something like, "looks like some pretty sensitive data you've got there. Want me to remember that?"

For now, OpenAI says the feature is in testing, and that it will be "rolling out to a small portion of ChatGPT free and Plus users this week," and that it's still being evaluated for usefulness. If you use ChatGPT don't forgetto check and see if it's on.

最新评论