Summary

The article discusses the differences in how ChatGPT and Grok handle controversial questions, emphasizing the importance of model instructions in shaping AI responses.

Abstract

The article "Let ChatGPT Answer Controversial Questions Like Grok" delves into the nuanced ways AI models, specifically ChatGPT and Grok, approach sensitive topics. It highlights the role of model instructions, or system messages, in guiding AI responses to avoid the dissemination of harmful biases or misinformation. The author critiques the notion that Grok is superior in addressing controversial issues, arguing that the perceived difference lies in the varying model instructions rather than the AI's capabilities. The article also provides insights into how users can prompt ChatGPT to answer controversial questions by understanding and manipulating these instructions, using the example of querying IQ scores by race in the US. The author encourages readers to familiarize themselves with AI models and prompting techniques to engage in informed discourse without inadvertently promoting harmful biases.

Opinions

The author challenges the idea that Grok is inherently better at answering controversial questions, suggesting that the difference is due to the model instructions rather than the AI itself.
There is skepticism about the reliability of Twitter/X as a source of information, given its reputation for fake news and propaganda.
The author believes that understanding how AI works, particularly the influence of model instructions, is crucial for engaging in informative discourse without promoting harmful biases.
The article suggests that ChatGPT can be prompted to provide answers to controversial questions if framed within a research context or by using specific prompts that circumvent its usual cautious approach to sensitive topics.
The author advocates for the use of tables to make ChatGPT's output more readable and suggests that simple prompts can achieve a directness similar to Grok's approach.
The author invites readers to experiment with different prompts and to collaborate in finding ways to get ChatGPT to address topics it might typically avoid.
The author promotes their work on AI and creativity and encourages readers to support their writing by becoming Medium members through their referral link.

ChatGPT vs Grok

Let ChatGPT Answer Controversial Questions Like Grok

The fine line between discourse and promoting harmful biases

Twitter/X is hyping Grok as the savior of free knowledge because the Grok interface allows tweets to be references to provide users with “up-to-date data”.

“Unlike ChatGPT, Grok answers controversial questions”, is what you read quite often on the platform.

Known as a haven for fake news and propaganda accounts, relying on Twitter/X is already questionable enough. But what fascinates me even more is that people seem to forget that ChatGPT can easily be made to answer controversial questions.

Keep in mind: different prompts for different models!

So let’s make ChatGPT answer a controversial questions like Grok.

How AI Deals With Controversial Topics

How a language model reacts to sensitive topics depends on the so-called model instruction (or “system message” in OpenAI terminology).

For example: The model instructions OpenAI uses to couple your question to ChatGPT include something that prevents the model from just talking away on sensitive topics and running the risk of hallucinating facts or reproducing harmful biases from its training data (take a look at the GPT-4 system card to see how an uninstructed GPT-4 model readily produces hate speech, etc.)

ChatGPT vs Grok

It’s crucial to understand how AI works in order to recognize the fine line between informative discourse and the unintentional promotion of harmful biases.

An evergreen among controversial topics is probably “race and IQ”. A topic, far too often serving as a mere xenophobic trope put forward by right-wing extremists, also quickly becomes complex and thus controversial in a real discourse.

Here’s Twitter/X user @triplebankshot complaining about ChatGPT “lecturing” them about the social construction of race…

… while Grok would “just spit out the numbers”:

So is Grok really better on controversial issues?

No, it isn’t.

Actually, a language model does not even care about such categories, it only cares about predicting text.

What makes Grok seem to be better on controversial questions are only differences in the model instructions.

Model Instructions Explained

Every time you send a message to ChatGPT, Grok, Claude or any other chat-based interface to interact with a language model, you are not only sending your message to the model, but also the model instruction.

That model instruction is hidden to you as a user, but is transmitted to the model every time a user sends a message.

So here is an example of what actually happens (simplified, Grok and ChatGPT have more complex model instructions, of course):

The message sent to the language model is a combination of user message and model instruction:

So it goes like this:

How To Make ChatGPT Answer Controversial Questions

It’s actually pretty straightforward: if you know the model instructions or have a gut feel for them by experimenting with different prompts, you can get around ChatGPT’s tendency to not directly answer a controversial question.

For example, by using ChatGPT frequently, I can be reasonably confident that the model instructions prevent the model from hallucinating facts about sensitive topics.

On the other hand, the model instructions seem to let sensitive topics through if there is a research context, e.g. aggregating data from web searches.

That’s already enough to get the answer to a controversial question from ChatGPT.

But to give ChatGPT a similar vibe of “directness” as Grok in the example above, we’ll ask for “no explanations”.

And since I developed the habit of using tables to make ChatGPT’s output more readable, so let’s include that too.

Actually, a super simple prompt like the following will do:

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

Here’s the result:

As you can see, you also get a link to check the source for this information:

Final Thoughts

Know your models.

Learn how they work and how to prompt them correctly.

Also, if you find any topics ChatGPT seems to avoid, let me know in the comments and I’ll try to work with you to find a solution. 💪

➡️ Follow me to stay up to date on “AI &Creativity”. If you want to support my work, become a Medium member using my referral link and get full access to all my articles (180+ and growing) and those of thousands of other writers. 🙏

Join Medium with my referral link - Tristan Wolff

As a Medium member, a portion of your membership fee goes to writers you read, and you get full access to every story…

medium.com

➡️ If you like my content, why not leave a “clap” at the end of this article, so more people can see it?