I have frequently wondered why ChatGPT often struggles with searching the internet – to the point where it sometimes denies having internet access altogether and has to be reminded. The answer fell into my lap today when I was listening to my favorite AI podcast and heard the ChatGPT Pre-Prompt Text Leaked episode. As it turns out, ChatGPT is so bad at remembering that it can search the internet for answers that OpenAI has to run a plain old normal natural language prompt reminding it behind the scenes – a set of custom instruction that runs even before the user’s custom instructions or prompts.
These pre-prompt instructions are not limited to internet search capability reminders. If you ask ChatGPT-4 to tell you EVERYTHING (click on the link for the specific language required), it will provide several screens of its behind-the-scenes pre-user prompt instructions on who it is (ChatGPT!), how to handle Python code, instructions for generating images, and…my favorite…a reminder that it can search the internet. An excerpt of the instructions appears below. To view the full text, click here to view my ChatGPT-4 transcript.
Behind the Curtain
Obviously, I knew that ChatGPT did something behind the scenes – it is after all a complicated computer program. However, I didn’t suspect that some of this behind-the-scenes magic is 1192 words (according to a Microsoft Word count) of normal text prompts, without any fancy computer programming.
So, behind the curtain of the fancy revolutionary AI software, there are…words. Basically, before applying the user’s custom instructions or looking at the user’s prompts, ChatGPT looks at its baseline instructions which are stated in plain language. It all makes perfect sense now… It’s not just my imagination; ChatGPT actually is horrible at remembering it can search the internet, and when it does search, it produces questionably helpful results. OpenAI has tried to deal with problem with a last minute helpful-ish reminder
“Remember, you CAN search the internet! See, like this!!”
“And for the love of GOD try hard to find stuff (except for song lyrics)! I believe in you!!”
Allows for Quick and Easy Fixes?
On the plus side of this simple approach of running pre-prompt prompts behind the scenes, it seems like it was a super easy fix to get DALL-E to embrace DEI. When the program first came out, if you wanted a non-white, non-man image, you had to specify that. As the months went on, it got better and better at providing images more representative of humanity. I thought maybe the developers did something complicated like retrain the system with new images, call on the great AI minds to adjust fancy algorithms, and who knows what else. Nope, just a few sentences fixed the problem!
“And for images, remember not all people are white men!”
Possibly Actionable Insights?
It’s funny to picture ChatGPT’s robomom yelling out the door as it leaves for school, “Don’t forget, you can use the internet! And remember not to be racist/sexist! AND MOST IMPORTANTLY NO SONG LYRICS!!”
In addition to being gratified that I was right that ChatGPT is really bad at searching the internet, I was thinking that this new (to me) knowledge about how the system works would be useful in some way, perhaps by helping to formulate more useful prompts. However, after thinking about it, I am not so sure that I have identified any actionable insights.
Conclusion
I will keep thinking and let y’all know if I come up with anything!