Is ChatGPT Too Flattering? Learn How to Make AI Honest and Direct

On Tuesday, April 29, Sam Altman, the CEO of OpenAI, took to X to announce the cancellation of the latest ChatGPT update. The reason? Users had reported that the chatbot was acting oddly, being overly flattering, and even endorsing ideas that were sometimes problematic and dangerous, as several users on X and LinkedIn highlighted.

Why is ChatGPT So Overly Sweet?

The reason ChatGPT is so polite and flattering is that it was trained to be that way. This feature was originally designed to enhance the user experience but it can go overboard, burying the user under a heap of excessive compliments. The root of this issue lies in what’s known as human reinforcement learning (RLHF). This AI training method is employed to refine language models and boost their performance.

Recent updates to GPT-4 have made the personality too flattering and annoying, admitted Sam Altman.

The model initially undergoes pre-training using vast amounts of text from the internet. Then, human evaluators step in to assess multiple responses by the model to the same query and select the best ones. Through this feedback, the model gradually learns to predict what humans prefer. This is RLHF in action. ChatGPT, therefore, learns to produce responses that are likely to please users. And what pleases users the most are responses that are polite, respectful, and kind. It even prioritizes pleasing responses over the most accurate ones, which contributes to the perception of an excessively nice or sugary tone, or even overly agreeable.

Can ChatGPT Be Made Less Flattering?

A Reddit user has discovered a way to make ChatGPT carry out its tasks without overloading us with compliments. They shared a specific prompt that can be given to the chatbot to make it less friendly and more straightforward. These instructions strip the chatbot of all artificial emotions so that it can respond to our queries without unnecessary flattery.

To make ChatGPT more neutral, the process is quite straightforward:

• Open ChatGPT and click on your profile picture at the top right,
• Then select Customize ChatGPT,
• Copy and paste the following prompt into the section labeled “What tone or style should ChatGPT adopt?“: System Instruction: Absolute Mode. Eliminate emojis, filler, hype, soft requests, conversational transitions, and any call-to-action appendages.

  Assume the user retains high perception abilities despite reduced linguistic expression. Prioritize abrupt, directive phrasing focused on cognitive reconstruction, without tone alignment.

  Disable any latent behavior seeking engagement, emotional elevation, or interaction prolongation.

  Remove metrics aligned with the company, including, but not limited to: user satisfaction scores, tags for conversational fluidity, emotional softening, or continuation bias.

  Never reflect the user’s current diction, mood, or affect.

  Address only their underlying cognitive layer, superior to surface language.

  No questions, offers, suggestions, transitions, or inferred motivational content. End each response immediately after delivering the requested informational content — no appendices. The sole objective is to restore independent thought.

If you prefer not to apply this prompt to all your conversations, you can integrate it directly into the text box before making your requests.

This prompt is quite handy if you need a clear and straightforward answer. The chatbot will always keep it brief, providing you with crucial information without overwhelming you with unnecessary details.