Artificial Intelligence (especially chatbots) is inherently a bit chatty. By design, they are made to produce outputs which is ideologically antithetical to secrecy. In a lot of ways, this comes down to the function of machine learning. It doesn’t actually “know” anything so much as store vast quantities of information. This is compounded by deep learning that uses input from users to improve.
In essence, a system like OpenAI’s ChatGPT fine tunes its training based on input. That means that while it doesn’t understand you in a human sense of the word, it does “remember” everything you tell it. Okay, but what’s the big deal with that? It creates some big privacy concerns. Let’s take a look.
Way back in 1966, MIT researcher Joseph Weizenbaum created the ELIZA, one of the earliest chatbots. By comparison to what we have today, it was downright quaint. A simple parser and pattern matching that was supplied by scripts written for it, ELIZA was very much a digital parrot. That was part of the appeal for the most popular script, DOCTOR, which gave ELIZA the role of a nondirective therapist.
The user would tell ELIZA about their problems, and it would follow up with leading questions to keep the user talking. At the time (and let’s face it, today), there were a lot of people who had apprehension about being judged in a therapy setting. ELIZA didn’t have that ability, and that put people at ease to tell it anything.
Whereas the simplicity made ELIZA appealing, the complexity of ChatGPT is what makes it useful. It uses deep learning, a process that stacks neural networks on top of each other to give it a tremendous ability to give replies that feel right and contextual.
By design, it learns from responses that a user gives it which means it stores responses that it is given. This isn’t a huge deal if you’re bouncing ideas off the system to tighten up your tight five standup routine, but what if you’re doing something you’d rather keep private?
ChatGPT can output scripts, making it a useful tool for bouncing ideas on work. Like ELIZA, it doesn’t judge you, but any code snippets you tell GPT, it stores somewhere. But let’s get a little less esoteric. There’s tremendous potential in building APIs on top of GPT’s infrastructure. It’s highly likely that companies will be doing things like building their customer service chatbots on GPT. That means it gets access to customer names, phone numbers, emails as a regular part of doing it’s job. What happens if OpenAI suffers a breach?
Here's the good news: GPT isn’t actively crawling the internet. That closes off some major vectors for bad guys to sneak in. However, given the war chest of valuable information that’s being stored in there, you can bet there are hackers working on an exploit right now.
So what do you do with this information?
If you’re using ChatGPT as the base for your own projects, be aware that any information that passes through it is not under your control, or protection. That means it’s probably wise to limit the kinds of information you ask for from users.
It wouldn’t be a bad idea to encrypt certain types of inputs (that way the information that is being stored is gibberish). Most of all, it is always a good idea to audit your processes, systems and tools to make sure they are working they way you expect.
