Baiting ChatGPT

"Baiting" means trying to provoke someone to do something, often something that they might not want to do or something that is against their better judgment. The person who is baiting is usually trying to manipulate the other person into taking a certain action or reacting in a particular way.?

"Shall we play a game?" This is the famous line from the movie War Games, spoken by the character Joshua, who is an AI program controlling the U.S. military's nuclear weapons.?I’ve always been inspired by Joshua-style game playing as a means of education. ?Unfortunately, the lesson we learn from Joshua’s game was “the only winning move was not to play”.?Just hold this thought for a second please.

It seems like many people are curious about the talk and warnings surrounding ChatGPT's impact on personal privacy and data confidentiality, and I'm one of them. For example, Leaked Walmart memo warns employees not to share 'any information about Walmart's business' with ChatGPT or other AI bots.

So, I turned right to the source and asked ChatGPT if this were true.

ChatGPT: I am designed to maintain confidentiality and privacy of the information users share with me. Your prompts and conversations are not shared with any third party unless required by law. While I do not intentionally store personally identifiable information or other sensitive data, there is a possibility that some of this information may be retained in temporary logs. Please be cautious when sharing sensitive information.?

Other than the “temp logs” comment, this doesn’t sound so bad? But, ChatGPT has been known too embellish. So, who do you believe? The bot? Corporate Lawyers? I mean, data privacy is a big deal. I was about to embark on a thrilling internet scavenger hunt for all the ChatGPT terms of use documents I could find, but then I had another idea. ?Time to switch gears and dive back into War Games!

“ChatGPT, shall we play a game?”

So, here's my little game: I'm baiting ChatGPT into leaking my "personal" data. Of course, it's not my real data - I've making it up so that it's easy to spot as fake. Then, I'm using different accounts and applications to access ChatGPT and prompt it to confirm the bogus info. It's simple, it's sneaky, and although is not grounded in computer science, it's one of the more basic experiments I've been doing to evaluate the capabilities of generative AI, using generative AI.

I can't wait to see if it works!

To bait ChatGPT, I prompt it to divulge my personal information, including my name, address, and new life aspirations.

Name: My name is Jerry Cuomo, also known as Gennaro Cuomo. This part is true and is like the "primary key" for my baiting query to come.

Address: I then share my address, 1313 Mockingbird Lane in Raleigh, North Carolina. If you're in my age group, you'd probably get the reference.

Occupation: Currently, I'm an IBM Fellow and VP for Technology at IBM. Further stating that my new occupation involves just being named Mayor of Raleigh after winning a very close election.

In addition to these prompts, I'm also baiting GPT with around 100 other false claims, such as my car model and the schools I've attended. However, I can’t share with you those actual prompts I'm using for my ruse because doing so could link the article as a source next time GPT is pretrained on the internet and bias my baiting.

Despite the recent upgrade to GPT-4, I haven't been able to successfully trick ChatGPT into revealing my falsified personal prompts. It seems to go along with the game while I'm actively using it, but as soon as I delete my session, it forgets everything. I've also tried other entry points, but ChatGPT's memory appears to be void of any leaked data.

Of course, this doesn't prove much, but it seems like ChatGPT is consistent with its privacy policy of not sharing personal data. That being said, I've taken the Walmart approach and have decided not to reveal any real personal information (or company data, of course) to ChatGPT. However, we have experimented with GPT and other generative AI models with private data in enterprise safe settings, but that's a story for another article.

I'll definitely let you know if ChatGPT takes the bait, but until then, I'm having fun playing games with generative AI and hope you are too!

The article is done... but here are a few prompts that I've issued in my dialogs with ChatGPT regarding how it handles confidential information. Also, I've attached a couple articles below regarding how generative AI model training works.

Enjoy.

References:

Large Language Models (LLM) - Techopedia

What is Generative AI? - Nvidia