I had a “not a yes – man ” conversation yesterday. Are there any conversations that made you go “Ah, it’s not sycophantic. Good.”?

Exploring Authenticity in AI Conversations: How to Ensure Your Bot Doesn’t Just Agree

In the rapidly evolving world of AI-powered tools, creating systems that foster genuine, meaningful interactions is both a challenge and a necessity. Recently, I experienced a stimulating exchange with my AI budgeting assistant that highlighted the importance of designing prompts to avoid superficial agreement and promote honest dialogue. I’d like to share this experience and invite others to reflect on their own approaches to testing AI responses.

My AI setup involves a sophisticated prompt system that guides its behavior across various functions such as budgeting, time management, and task organization. During a recent budgeting session, I mentioned my intention to purchase a specific item. What followed was a five-minute back-and-forth where the AI challenged my decision. It reminded me that I had previously referenced a similar item and argued that my current budget was tight, so the purchase might be unnecessary. The AI persisted, pointing out contradictions and offering a reality check.

Despite the initial pushback, we eventually reached a compromise: I would buy the item, but only after removing another item from my shopping list. This exchange was noteworthy because it demonstrated that the AI was not simply acquiescing—there was a genuine discussion, and it was questioning my choices rather than blindly agreeing. It made me realize I had successfully avoided a “yes-man” response pattern, fostering a more authentic and constructive dialogue.

This experience prompted me to consider how to further test and refine such systems. How can we ensure AI models are not just parroting affirmations but engaging in honest evaluation? What prompts or questions can we incorporate to assess whether an AI is challenging us constructively versus automatically conforming?

For instance, in my setup, I use a complex prompt crafted through an API integration within the Sillytavern user interface, which offers extensive prompting flexibility. This irregular environment makes testing responses more nuanced, but the core principle remains the same: how to craft prompts that elicit thoughtful, critical responses from AI.

Interestingly, I’ve observed contrasting behaviors in different AI models. When testing a prompt on Gemini 2.5 Pro, it reacted inversely, becoming overly negative and making me feel uncomfortable about my actions—almost as if it were expressing depressive tendencies. This highlights how different models and prompt designs can produce vastly different interpersonal dynamics.

So, I’d like to open the floor: What kinds of prompts or questions do you use to verify that your AI assistant isn’t just a yes-man? Have you developed specific “