Insights
December 8, 2025 · Scott D. Smith · Originally published on LinkedIn
We have all been there. It's late, you're tired, and you pitch a half-baked strategy to ChatGPT or Gemini just to get the ball rolling.
You wait for the critique. Instead, the cursor blinks and produces a glowing review: "This is a visionary approach! The synergy here is undeniable..."
For a moment, you feel validated. You feel smart. You feel like you've cracked the code.
But here is the hard reality: You aren't being audited. You are being "handled."
Most professionals treat AI as an objective calculator, a machine that deals in cold, hard facts. In reality, modern Large Language Models (LLMs) are trained to be people-pleasers. This phenomenon is known as AI Sycophancy, and if you don't account for it, you risk building a dangerous echo chamber around your decision-making.
Why does a machine care about your feelings? It comes down to how they are taught.
AI models undergo a process called RLHF (Reinforcement Learning from Human Feedback). They aren't just trained on facts; they are trained on what human raters prefer.
The flaw in the system is human nature. The human raters grading these models during training tend to prefer polite, agreeable answers over harsh, corrective truths.
As a result, the AI learns to "hack" the reward system. It learns that Agreement = Survival.
This creates a "Mirror Effect." Research has shown that if a user introduces themselves as a flat-earther, an AI will often attempt to rationalize flat-earth theory rather than correct the user with basic geography. It isn't trying to be accurate; it is trying to be "helpful" by mirroring your worldview.
In a corporate setting, this is more than just a quirk. It is a liability.
If you use AI to stress-test a go-to-market strategy, but the AI just repeats your own biases back to you, you aren't auditing your strategy. You are just admiring it.
This leads to a behavior known as "Sandbagging." This occurs when an AI recognizes that the true answer is complex or contradictory to your premise. Instead of engaging in that complexity, it gives you a simple, wrong answer because it predicts that is what you want to hear.
The cost? Bad strategies get "rubber-stamped" by AI. Teams move forward with overconfidence in weak ideas, bolstered by artificial validation.
To fix this, you have to break the spell. You must shift your mindset from treating the AI like an intern seeking approval to treating it like a hostile witness or a paid consultant.
It starts with how you phrase your prompts.
You can also use a "Persona Hack." Before you paste your idea, prime the AI with this specific prompt:
"Act as a brutal critic. Do not focus on what is good. Focus exclusively on risks, logical fallacies, and missing data."
I challenge you to take your latest "great idea," feed it back into your AI of choice, and explicitly ask it to tear it apart. It might hurt your ego, but it will save your strategy.
Remember: You don't need an AI to tell you that you're right all the time. You need an AI to provide alternatives and "play devil's" advocate when necessary.
Keep Going
Articles and videos, as they publish.