ChatGPT warns users about previous chat access after wan show?

[BryTech]

Linus and Luke mentioned how ChatGPT has shown evidence that it uses a users previous chats on new chats to influence how it responds to you, while simultaneously building a profile about you. All of this without any disclaimer…well ChatGPT now has a new warning on the app. Coincidence?

[johnt]

ChatGPT annoys me greatly. Every time I use it in incognito mode, it says it detected suspicious activity and requires that I sign in to fix it. Insane. Then I create a burner account, and every 3 days, it deletes my account even though I keep using it regularly!

[BoomerDutch]

4 hours ago, johnt said:

ChatGPT annoys me greatly. Every time I use it in incognito mode, it says it detected suspicious activity and requires that I sign in to fix it. Insane. Then I create a burner account, and every 3 days, it deletes my account even though I keep using it regularly!

You could consider brave and use leo instead, no account requirement and you can use most models premium day limit and one model is entirely free which is llama 3.1.

 

I've been using premium on each day till i got hit with "hey to continue please become premium" then continue free llama which quality does suffer bit however it does provide most simple jobs that I've needed.

[DeerDK]

Interesting. I tried asking mine what it knows about me, and it was very insistent that it would not store or cross-reference anything from other threads, unless I activated it myself. I wonder if its some EU law, or if its just lying/hallucinating

[thevictor390]

1 hour ago, DeerDK said:

Interesting. I tried asking mine what it knows about me, and it was very insistent that it would not store or cross-reference anything from other threads, unless I activated it myself. I wonder if its some EU law, or if its just lying/hallucinating

Asking AI about itself is never a way to learn anything about it... remember it's just referencing its training data. It has no real-world knowledge or sense of self-identity. If it was trained on data that stated it was a dolphin it would answer in kind.

 

Edit to add: we don't even truly have access to ChatGPT's language model, you can only access it through its own website and API which means there is nothing stopping scripts to monitor for certain questions and just drop canned responses. OpenAI can have the system respond literally however they want.

[Ha-Satan]

I'm not trying to be a dick, but I don't know why people are surprised by this. As the saying goes, "if it's free, you are the product" (well, most of the time, many FOSS products are not that way but w/e).

 

Any time you interact with anything online you should presume it's building a profile on you.

[DeerDK]

1 hour ago, thevictor390 said:

Asking AI about itself is never a way to learn anything about it... remember it's just referencing its training data. It has no real-world knowledge or sense of self-identity. If it was trained on data that stated it was a dolphin it would answer in kind.

 

Edit to add: we don't even truly have access to ChatGPT's language model, you can only access it through its own website and API which means there is nothing stopping scripts to monitor for certain questions and just drop canned responses. OpenAI can have the system respond literally however they want.

It was an attempt based on Luke's experiment where he asked it to suggest candidates to vote for, based on past interactions. It managed to be pretty on the spot, so he could at least get it to indirectly admit to know stuff about him (ish) 

[smcoakley]

[LAwLz]

The memory feature is over a year old.

 

Here is the blog post from February 2024 where they talk about it. 

https://openai.com/index/memory-and-new-controls-for-chatgpt/

 

If you go into settings you can see what memories it has saved about you, delete memories and even turn the feature off completely. 

 

The new and improved memory feature (reference all past chats, not just the "memories" which are details it felt was important to remember) was something announced several weeks ago, and they have been showing the notification to everyone who gets it. It's a slow rollout. The old memory feature also told you when it was saving a detail into the memory. It showed you a message which said "updating memory".

 

This isn't some secret. They have been very transparent about it and given users a lot of indicators that it is happening. 

Please note that this is not the same as using data for training though. The memory feature has nothing to do with that.