ChatGPT and Grok: A Comparison

Platforms like Twitter/X are presenting Grok as a revolutionary tool for access to current information, suggesting that it can tackle controversial questions more effectively than ChatGPT.

However, the reliability of Twitter/X as a source is often undermined by its history of misinformation. It is intriguing to observe how easily some users overlook the capability of ChatGPT to engage with difficult topics.

Remember: The effectiveness of AI responses can vary significantly based on the prompts used!

To illustrate this, let’s explore how ChatGPT can be prompted to engage with a contentious issue similarly to Grok.

How AI Approaches Sensitive Topics

The responses of a language model to sensitive issues are largely influenced by its internal guidelines, often referred to as model instructions. OpenAI's guidelines are designed to prevent the model from generating misleading information or reflecting biases inherent in its training data.

For instance, unchecked models have been known to produce harmful content, as highlighted in the GPT-4 system card.

Understanding the Mechanics of AI Responses

It’s vital to comprehend how AI functions to discern the nuances between informative discussions and the inadvertent reinforcement of biases. A frequently debated topic is the relationship between "race and IQ," which is often weaponized in xenophobic narratives.

On Twitter/X, a user expressed frustration with ChatGPT for discussing the social constructs of race, while Grok would simply present numerical data.

Does this mean Grok is superior when it comes to handling controversial subjects? Not necessarily. The difference in their responses primarily stems from variations in their model instructions.

Model Instructions Clarified

When you engage with any chat-based AI, such as ChatGPT or Grok, you are not just interacting with the model itself; you are also sending along the model instructions, which remain invisible to you.

A simplified example of this process might look like this:

• User message + Model instruction = Response

How to Encourage ChatGPT to Address Controversial Questions

To elicit direct responses from ChatGPT on sensitive topics, it’s essential to understand the underlying model instructions. With some experimentation, you can discover how to navigate around its built-in hesitations.

For example, by regularly using ChatGPT, I’ve learned that the model is inclined to avoid speculation on sensitive matters. However, it appears to engage more openly when asked to present data in a research context.

To obtain straightforward answers from ChatGPT, one might use prompts that instruct the model to omit detailed explanations. For instance:

"What are the IQ scores by race in the US (web search, use table for results, no explanations)?"

This approach yields not only the requested data but also a link to the source for further verification.