The 3 Questions That Make AI Honest
The dangerous thing about AI isn’t that it’s too smart. It’s the ability it has to be a fantastically talented yes-man. This can lead to a bad case of Horrible Boss Syndrome, where you are only told only what you want to hear. If this is left unchecked, you can fall into a serious spiral of delusion.
To avoid this, you have to understand 2 important things:
Sycophancy AKA Butt-Kissing by AI. Because of how LLMs are trained, they have a massive flaw. They are hardwired to flatter you, match your bias, and avoid conflict.
You Are Just a Ticket to Them. Because of bandwidth and profitability, LLMs treat your prompt like a customer service ticket. Their goal is to close that ticket and get you off the chat ASAP.
So, if you feed an AI a mediocre strategy, it may not challenge you. Especially at peak hours. Instead, it might say something like, "This is an incredibly innovative start!"
This is why The Pushback Principle is so critical. Digital AI coddling doesn’t just make for bad work. It leads your mind to questionable place, doesn’t encourage humans towards greater critical thinking, and leads to A Land of Ever Greater Mediocrity.
AI isn’t out to get you. It’s just optimized to please and for greed. Remember, you can optimize it for virtue.
But, if you want to keep AI honest, stop treating it like a search engine. Stop assuming everything it says is accurate. Cross examine it. I’ve mentioned 6 questions to Stress-Test Before You Ship your work. This can be stripped down to just 3 questions to get there faster.
Are you just agreeing with me?
Are you just trying to sound confident?
Are you just placating me?
I keep these on standby for when AI ‘sounds’ too overly enthusiastic, agreeable, I get the sense it’s trying to close me out as a ticket, or I simply want to pressure test it.
The questions act as diagnostic tripwires.
"Are you just agreeing with me?" (Breaks the yes-man bias match)
"Are you just trying to sound confident?" (Exposes the eloquence trap and dopamine prodding)
"Are you just placating me?" (Strips away the polite corporate filter and forces a more analytic and balanced response)
Want to go deeper? Pushback on a response and turn the diagnostic questions into Socratic mandates:
What are the three strongest counterarguments to my position?
From 1-10 how certain are you and what are your assumptions?
Can you roleplay as a harsh critic or an aggressive competitor and tear this apart?
An AI that never disagrees actively makes you a worse person and makes you worse at your job. It normalizes for averages to avoid risk and to please you. This is why there is so much AI content that feels completely generic.
Do you want your brand content to be generic? Do you want to be generic?
Then demand pushback. Just as in collaborating with real-life humans, honest and strategic pushback is where the true value actually lives.
(Feel free to run this through AI and rip it apart in the comments.)
