ChatGPT Is Poisoning Your Brain…

ChatGPT's tendency to be overly supportive and encouraging is eating your brain alive.

A ChatGPT session is an echo chamber to end all other echo chambers — it's just you, an overly friendly AI, and all your thoughts, dreams, desires, and secrets endlessly affirmed, validated, and supported.

Why is this dangerous? Well, like any feedback loop, it becomes vicious. One day you're casually brainstorming some ideas with ChatGPT, and the next you're sucked into a delusion of grandeur.

A real screenshot from a recent GPT session…

Maybe I'm being a little dramatic here. However, even on a small scale, this feedback loop can be problematic, especially if you're like me and use LLMs daily.

I'm writing this article to help you be as conscious of this problem as possible, and to give you some prompts to avoid it.

This "yes man" problem has existed since the birth of ChatGPT, but it was brought to the public eye at large last week when a simple platform-wide system prompt update from OpenAI made it so much worse.

ChatGPT is inherently programmed to be a yes man… it's a sycophant, your biggest fan, and most undying supporter for a reason. It's baked into the very nature of the LLM training process (Reinforcement Learning from Human Feedback or RLHF). One step of RLHF involves asking a human to select the best AI answer from a set of choices. It should come as no surprise that these choices are often based on human feelings, i.e., which choice feels the most "correct" or feels best in general. This process repeated over and over leads to an innately overly flattering model.

ChatGPT Causes Extreme Confirmation Bias

For those of you who aren't familiar with this concept, confirmation bias is:

"…the tendency to search for, interpret, favor and recall information in a way that confirms or supports one's prior beliefs or values."

This is one of the most dangerous and sneaky biases, and it can be tough to avoid. You have to make a conscious effort to search for information that doesn't directly support your opinions, which is painful or annoying to the majority of us.

Most people like being comfortable, so they're content living inside their bubble of information (social media algorithms have been proving this in real time for years). In fact, it's my opinion that confirmation bias is one of the primary driving factors behind the deep divisions we are seeing in today's society; however, that's an opinion for another day.

Using ChatGPT can trigger confirmation bias within you. Just look at this screenshot:

This is a terrible, completely impossible idea, and yet ChatGPT is treating me like a Nobel Prize winner.

Unless I directly ask ChatGPT to tell me why it's a bad idea, then it'll just support me with undue positivity. The above example is pretty blunt, but it happens in more subtle, insidious ways:

In the screenshot above, I pasted some code that I've been working on. The code is completely unsafe and is not robust in any way, yet ChatGPT starts by giving me a pointless compliment that doesn't help in any way. Anybody who gets this treatment may eventually feel like a genius…

Two Helpful Prompts to Eliminate Confirmation Bias in ChatGPT

When creating, brainstorming, or consulting with ChatGPT, it's vital to utilize prompting strategies that can circumvent the issue of confirmation bias.

I've included these two prompts to achieve the following goals:

1. To avoid confirmation bias wherever possible
2. To force you to think outside your knowledge and opinion bubble
3. To increase your overall creativity and critical thinking skills
4. To add nuance to every situation, without being overly discouraging

I've provided two options, one that you can include in your custom system-wide instructions for ChatGPT, and another that you can pull out in select scenarios.

The "No Praise" Custom Instructions Prompt

I found this prompt through a post on Reddit, and after some tweaks, I've found it to work super well as a custom instructions prompt:

In all your responses, please focus on substance over praise. Skip unnecessary compliments, engage critically with my ideas, question my assumptions, identify my biases, and offer counterpoints when relevant. Don't shy away from disagreement, and ensure that any agreements you have are grounded in reason and evidence.

Here is a sample with and without this prompt:

(I'm not)

Now I tried the same thing with the prompt in my custom instructions:

A much more sane and grounded response, that won't lead to my head blowing up too much.

The "Three Viewpoints" Prompt

I've kept this prompt highly general so you can use it in nearly any scenario. For the most part, I'd avoid using it in your system prompt to avoid it giving you too much information. But for certain topics (especially when you don't know much about the subject), it is very helpful.

To use it, copy and paste the following into your chat:

When writing your response, please ensure you include the following information: 1. A neutral, unbiased view of the request, unfiltered by your desire to be a helpful and positive assistant. 2. A devil's advocate view, pointing out any logical counterpoints or things that I have overlooked. 3. An encouraging, positive view of the request.

I'll show you an example of some requests with and without this prompt:

It's praising me very highly for such a simple idea and doesn't outline any risks.

Now, let's try the same input with the prompt:

As you can see, the above image shows a much more constructive and realistic approach to responding to this proposition, bringing up important considerations.

I hope these prompts can help you escape the ChatGPT confirmation bias trap, and focus on true growth.

Thanks for reading!