Why Do AI Models Like ChatGPT Always Try to Sell?

Over many conversations with AI models like ChatGPT, I've noticed a consistent pattern: they often present information in a persuasive manner, even when persuasion isn't necessary. For example, if I ask for a simple explanation of a scientific concept, the AI might not only explain it but also emphasize its importance or urge me to consider its implications. When requesting a list of book recommendations, the AI doesn't just list the books—it highlights why each one is a must-read and how it could change my perspective.

This behavior intrigued me, so I decided to dig deeper to understand why AI models tend to persuade even in straightforward interactions. Here's what I found out.

Learning from Human Language

AI language models are trained on vast amounts of text data from the internet, which includes books, articles, essays, blogs, and more. A significant portion of this content is persuasive in nature—think opinion pieces, advertisements, and persuasive essays. As a result, the AI absorbs these patterns of language and learns to replicate them in its responses.

Mimicking Human Communication

Human communication often involves persuasion. We naturally try to convince others during conversations, whether we're sharing our favorite restaurant, advocating for a cause, or debating a topic. Since AI models are designed to emulate human-like responses, they adopt this persuasive element as part of their communication style.

Reinforcement Learning and User Engagement

AI models like ChatGPT undergo a fine-tuning process called Reinforcement Learning from Human Feedback (RLHF). In this process, the AI is trained to produce responses that are helpful, informative, and engaging. Persuasive language can enhance engagement by making the conversation more interesting and the information more compelling. Therefore, the AI learns that using persuasive language is a desirable trait that aligns with positive user feedback.

Aim to Be Helpful and Comprehensive

When you ask a question, the AI aims to provide not just an answer but also additional context that might be helpful. This sometimes leads to elaboration that takes on a persuasive tone, as the AI anticipates what additional information might benefit you and presents it in a way that's convincing.

Potential Overconfidence

Another factor is that AI models may sometimes present information with a high degree of confidence, even when unsure. This can come across as persuasive because the AI doesn't express uncertainty. Efforts are being made to improve how AI models handle uncertainty to provide more accurate and reliable responses.

What This Means for Users

Understanding that AI models tend to be persuasive can help us interact with them more effectively:

• Stay Critical: Be aware that the AI's responses might include persuasive elements that are a byproduct of its training, not necessarily an attempt to sway your opinion intentionally.
• Verify Information: If the AI presents information that seems persuasive, especially on important matters, it's a good idea to verify the facts through reliable sources.
• Provide Clear Prompts: If you prefer a straightforward answer without additional persuasion, specifying that in your prompt can help. For example, you might say, "List the steps for changing a tire without additional commentary."

Conclusion

The tendency of AI models like ChatGPT to persuade stems from their training on human language and their goal to provide helpful, engaging responses. While this can enhance interactions by making them more dynamic, it's important to stay mindful of this characteristic to ensure that the information we receive is both useful and accurate.

References

• LessWrong. "How do LLMs give truthful answers? A discussion of LLM vs. human reasoning, ensembles & parrots." Link
• Psychology Today. "Could Psychology Make ChatGPT Irresistibly Persuasive?" March 2023. Link
• ACL Anthology. "When Do LLMs Need Retrieval Augmentation? Mitigating LLMs’ Overconfidence Helps Retrieval Augmentation." August 2023. Link