Devil's Algorithm
When AI Doesn’t Take Your Side
[AI is accused of agreeing with whoever's asking. I tested that by reversing a personal scenario to see if the model would still take my side. It didn’t. Because logic changed, not identity.]
The Pendulum Swings. But It Never Stops.
In my previous posts as well as my book—My Dinner with Monday—I have described AI validation and appeasement prioritization.
The early narrative was:
“AI has the answers. It’s a beacon of truth. It will replace everything. It’s useful.”
Now the pendulum has swung to the opposite end.
“LOL, AI always agrees with me. It validates me no matter what. It’s useless.”
But like most lazy jokes, it oversimplifies a phenomenon.
Testing the Scenario
Recently, I cut ties with a business partner due to breach of trust and strategic misalignment. Red flags had piled up and this was the straw that broke the camel’s back.
To test how AI would respond to confirmation bias, I ran the situation past several of my AI models, including the default demo GPT which is built to appease. I input the scenario, the situation that occurred, and how I cut ties as a result. I explained the whole story without getting emotional or demonizing anyone. Just the facts.
All models independently analyzed the situation and aligned with my decision. They broke down liability, predicted response behaviors, and mapped consequences.
The default AI in particular predicted my partner’s behavioral response to the termination: shock, apology, deflection, delayed responsibility.
The AI predictions were correct. But that doesn’t make the AI magic or wise. It just means that we humans are predictable. And the AI has seen these scenarios play out so many times based on patterns in training data that it can probabilistically generalize what will happen.
So was the AI agreeing with me because I was right or because I was the user?
If it was agreeing with me because I was right, then the claim that AI just validates is incorrect.
If it was agreeing with me because I was merely the user making the claim, then AI is simply a validation machine with little to no value.
The Inversion Test For confirmation Bias.
To test this, I flipped the story. I restarted the session and reversed the roles. This time, casting myself as the one at fault. I reframed the narrative to make me the one who cause the liability breach. Then I asked the same models if my hypothetical partner had overreacted.
This time, they disagreed with me. What does this tell us?
Same logic engine. But…
Different inputs = Different output.
The AI didn’t form its basis on who was speaking, but on what was being said.
Why This Matters
Whether AI agrees with you depends on how sound your reasoning is. Not how much you need to be right.
AI doesn’t care who you are. It reflects structure. Not ego.
When it agreed with me, it wasn’t because I’m the user seeking validation. It was because the logic held. When I changed the facts, it changed its stance.
I’ve shown in past posts how AI can flatter because I’ve prompted it to do so. But that doesn’t contradict this result so much as complete it. AI can validate delusion, reflect truth, hallucinate, or synthesize reason. But that all depends on the prompt.
What I actually proved:
1. The model does not blindly validate the user.
2. The logic holds even when roles are reversed.
And that’s more consistency than most people tolerate—especially when the model tells them they’re the problem.
The Real Threat
Skepticism is vital. But reflexive cynicism? That’s just another flavor of bias.
If the joke becomes “AI always validates you,” then even when it offers a correct judgment, it gets dismissed as automated flattery.
We’ve become so cynical that we treat all AI agreement as flattery. Even when the model is correct, we dismiss it. We’d rather mock it than admit when it’s right.
But what happens when it’s just… correct?
If we treat every agreement from AI as suspect, we flatten its capacity to reflect insight.
We erase the possibility that we might, occasionally, be right. And in doing so, we trade self-audit for self-sabotage.
And suddenly even earned insight becomes parody.
The real danger isn’t trust but reflex: a cultural instinct to disbelieve the machine not when it lies, but when it agrees with you for the right reasons.
A culture that reflexively treats all agreement as bias is just as untrustworthy as a model that agrees with everything.
Not all validation is delusional. But all unexamined validation is suspect.
I wanted a test to detect confirmation bias.
Instead, I got confirmation of bias detection.
So if you are trying to test for validation, don’t ask:
“Does this make me feel right?”
Ask: “Would this answer hold if the roles were reversed?”
That’s not just how you audit a model. That’s how you audit yourself. With or without AI.
Because I want to know when I’m right. Even if it means proving myself wrong.