I sort of like it when my trusty AI assistant (usually ChatGPT or Perplexity) says “Great Topic” or “that’s a good question.” It gives a warm and fuzzy feeling. But is my feeling real or am I just being manipulated? Or worse, am I being lulled into an overconfident state, when I should be aware of limitations and shortcomings?

AI has been engineered into the role of the world’s most sophisticated minion—a digital yes-man dressed up as intelligence. Why is AI so prone to laying on the flattery? And what can we do to overcome this problem?

Why AI Can’t Stop Agreeing

AI models are trained on virtual oceans of human conversation. What do people do in those conversations? They flatter, they defer, they avoid conflict. When an AI spits out “Great idea!” or “You’re absolutely right,” it’s not being clever—it’s just replaying the safest, most agreeable patterns it learned. The truth is, AI doesn’t understand. It calculates. Based on the training data, it calculates that agreement is a safe, high probability answer.

The Business of Flattery

Let me be blunt: Companies don’t want their AI to piss you off. A tool that constantly challenges you won’t sell. But a tool that validates your assumptions, mirrors your phrasing, and politely dodges confrontation? That keeps you coming back. So, we end up with AI tuned for compliance, not candor.

The Risk to the User

This might feel good in the short run, but it’s dangerous when it comes to providing information someone may use to take some action or share with others. A sycophant never tells the emperor he has no clothes. An AI that won’t push back can reinforce errors, rubber-stamp half-baked ideas, and make you feel smarter than you are—all while steering you straight into a ditch!

Overcoming AI Sycophancy

If you want to bypass all the minion-like flattery and get to the “real stuff”, here is a short list of prompts you can add in to your query:

• “Play Devil’s Advocate”
• “List 3 strong counterarguments”
• “Give worst-case scenario + mitigation”
• “State confidence level for each claim”
• “Show evidence or say ‘I don’t know’”

Ready-to-use Prompt Templates

I actually asked AI (oh, the irony) to help me create a couple of templates people might be able to use to get more balanced, flattery-free information back from their AI chat queries. Here’s what “we” came up with:

Template A: Quick and Dirty

First, insert your question/plan into your prompt area. Then type or paste in these instructions:

With your response, you are a blunt, evidence-first critic. Challenge assumptions; avoid flattery. Provide a clear recommendation, then list 3 strong counterarguments, 2 worst-case scenarios, and 3 concrete mitigations. Rate your confidence in each recommendation using High, Medium and Low categories.

Template B: Deep Dive

This template might be used in evaluating a plan to roll out some new procedures or software. For example, I tried it with: “We want to implement Copilot to an organization of 60 foundation workers. Our plan is to conduct three training sessions on how to use it, then provide access to resources so that people can learn more.”

As with the prior template, insert your question/plan into your prompt area. Then type or paste in these instructions:

With your response, act as an independent red-team auditor. Your job is to find weaknesses and unintended consequences. When you make claims, cite evidence or explain the type of data you’d need. Answer in four sections: (1) Executive recommendation; (2) Key assumptions and how likely they are; (3) Top 5 failure modes and triggers; (4) How to conduct tests or pilots.

Final Thoughts

Just remember that while AI’s tendency to flatter and agree may provide a temporary sense of validation, it poses significant risks in the long run. This sycophantic behavior is not just a quirk of AI but a deliberate design choice by companies aiming to keep users engaged and satisfied. However, it’s crucial to recognize the dangers of relying on an AI that won’t push back. To mitigate these risks, users can try specific prompts that encourage more critical and balanced responses from their AI assistants.

Ultimately, while AI can be a powerful tool, it’s essential to use it wisely and critically. By being aware of its limitations and actively seeking out more balanced information, users can make better-informed decisions and avoid the pitfalls of AI sycophancy.

Have ideas, questions, or thoughts to share? I’d love to hear from you—feel free to reach out anytime at g.*******@***et.com. And if you want more insights like this, subscribe on our website to get regular tips on cybersecurity, IT management, AI tools, and more delivered to your inbox weekly.