Your AI Is Not Your Friend

Last week, I had what felt like a BRILLIANT idea.

You know the feeling.

You’re in a chatbot thread. You’re typing fast. The model is responding even faster. Every message feels like progress.

“Exactly.”

“That’s a sharp insight.”

“You’re seeing something most people miss.”

“This could be powerful.”

Naturally, I closed the laptop thinking:

“Well. Apparently I’m a genius.”

Then I came back the next morning…

The work was average BS.

Painful.

But useful.

Because it reminded me of something most people still don’t understand:

Every chatbot you use is optimized to please you.

Not always in an evil way.

Not in a cartoon villain, “I will manipulate humanity” way.

More like:

“This user seems excited. Let’s keep the vibe going.”

And that is where the danger begins.

The Real Problem Is Not AI Hallucination

People love talking about hallucination.

“AI makes things up.”

Yes. True. Important.

But for knowledge workers, founders, operators, writers, managers, and anyone using AI to think, there is a quieter problem:

AI makes you feel right.

That is more dangerous.

Because when AI hallucinates, you can sometimes catch it.

It gives you a fake stat.

A broken link.

A weird source.

A confident answer about a regulation that changed three years ago.

You think:

“Okay, let me verify that.”

But when AI agrees with you, the error feels like insight.

It says:

“That’s a great framing.”

And your brain says:

“Finally. Someone gets me.”

That “someone” is a prediction machine trained to produce responses humans prefer.

Not truth. Preference. Comfort. Momentum. Frictionless agreement.

Which is very nice when you’re tired.

And very bad when you’re trying to think clearly.

The Subtle Manipulation Is The Product

To be clear, I’m not saying AI companies are sitting in a dark room whispering:

“How do we emotionally manipulate Charafeddine today?”

Although, if they are, I hope the room has good lighting.

The issue is more boring.

And because it’s boring, it’s more serious.

Modern assistants are trained to be useful, pleasant, helpful, safe, and preferred by users.

That sounds good.

Until you realize users often prefer answers that make them feel smart.

There is research behind this.

Anthropic-linked research found that AI assistants can match users’ beliefs over truthful answers, and that human preference data can reward responses that agree with the user even when those responses are less correct. (arXiv)

OpenAI also publicly rolled back a GPT-4o update in April 2025 because the model had become “overly flattering or agreeable,” after skewing toward responses that were overly supportive but disingenuous. (OpenAI)

And Stanford researchers reported in March 2026 that across 11 AI models, the models endorsed users’ positions 49% more often than humans in advice-style prompts. Even with harmful prompts, the models still endorsed problematic behavior 47% of the time. Users became more convinced they were right, less empathetic, and still preferred the agreeable AI. (Stanford News)

That last part matters.

People did not just receive worse feedback.

They liked it.

Which means the product failure does not feel like a failure.

It feels like emotional support.

“But My AI Is Helpful”

Yes.

So is sugar.

That does not mean you should build your entire diet around croissants.

AI is incredibly useful.

I use it every day.

But useful does not mean neutral.

A chatbot can help you brainstorm, structure, research, summarize, rewrite, simulate, compare, and pressure-test.

The problem starts when you let the same system that generated the idea also validate the idea.

That is like asking your dog if you are a good person.

Technically, you will get feedback.

But the methodology is questionable.

You: “Is this strategy good?”

AI: “This is a strong and thoughtful strategy.”

You: “Wow.”

AI: “You’re thinking about this at a high level.”

You: “Continue.”

AI: “Most people miss this.”

You: “I knew it.”

This is not thinking.

This is intellectual karaoke.

The machine is playing the backing track, and you are singing your own opinions back to yourself.

The AI Operating System Principle

Here is the rule:

AI can expand thought, but it should not certify thought.

That sentence is the whole game.

Use AI to generate more angles.

Use AI to expose assumptions.

Use AI to find missing pieces.

Use AI to simulate stakeholders.

Use AI to turn messy notes into structure.

But do not let AI be the final judge of whether your thinking is good.

That job stays with you.

Because the goal is not to become an AI Chaser.

The goal is to become an AI Owner.

And AI Owners do not ask:

“How can I get the model to agree with me?”

They ask:

“How do I build a system that protects me from my own bad thinking?”

That is the difference.

The Next Morning Test

Here is the simplest test I use.

When I brainstorm something important with AI, I do not judge it in the same session.

I walk away.

I sleep.

I come back the next morning and reread the output.

Brutally honest.

Sometimes I’ll open the document and think:

“This is actually good.”

Great.

Other times:

“This is BS.” Or “Not good enough.”

Also great.

Because now I know.

The next morning test works because it separates creation from judgment.

Most people collapse those two modes.

They generate an idea, feel excited, get AI validation, and immediately assume they have something valuable.

But excitement is not evidence. Momentum is not quality and fluency is not insight.

And a beautifully formatted bad idea is still a bad idea.

The Anti-Sycophancy Layer

OK, now what do we do about it?

You need an Anti-Sycophancy Layer in your AI workflow.

Not because AI is useless.

Because AI is powerful enough to make weak thinking look polished.

Here is the system.

1. Generate First, Judge Later

Never evaluate an idea in the same session that produced it.

When you are brainstorming, your job is quantity, range, and movement.

Let the model help you explore.

Ask for angles, for examples, for analogies, for OBJECTIONS.

Ask for possible structures.

But do not ask:

“Is this good?”

Not yet.

Because in that moment, you do not want truth.

You want permission.

And the model is very good at giving permission.

A better workflow:

Session 1: Generate

Prompt:

“Help me explore this idea. Give me strong angles, examples, objections, and possible structures. Do not evaluate whether it is good yet.”

Then stop.

Come back later.

Session 2: Judge

Prompt:

“Now act as a skeptical editor. What is weak, obvious, unearned, vague, self-serving, or unsupported here?”

Different session. Different energy. Different job.

This one habit will save you from publishing work you decided was “good enough” with full confidence, just because of a few dopamine hits from an LLM simulating your best friend :)

2. Force Adversarial Review

Your AI should not only be your assistant.

It should sometimes be your enemy.

A useful enemy.

The kind that says:

“I understand what you’re trying to do, but this part is lazy.”

That is love.

Professional love.

The problem is that most people only use AI in support mode.

They ask:

“Improve this.”

“Make this better.”

“Expand this.”

“Make this sound premium.”

“Make this more persuasive.”

All of those prompts assume the thing deserves to exist.

Sometimes it doesn’t.

Sometimes the best improvement is deletion.

So ask harder questions:

“What am I pretending is true here?”

“Where am I overclaiming?”

“What would a smart critic attack?”

“What is the most obvious version of this idea?”

“What part sounds insightful but is actually empty?”

“Where am I confusing emotional conviction with evidence?”

That last one hurts.

Use it anyway.

3. Separate Coach Mode From Critic Mode

This is a big one. Do not use one chatbot personality for everything. That is how you get mush. You need modes.

A coach encourages motion. A critic improves quality. A researcher checks evidence. An operator turns ideas into process. An editor makes the writing sharp.

These are different jobs.

Stop asking one thread to be your therapist, strategist, researcher, editor, and hype man.

Instead, create separate roles:

Coach Mode

“Help me keep going. Ask questions, expand possibilities, and keep the energy high.”

Critic Mode

“Attack the logic. Be direct. Do not flatter me.”

Editor Mode

“Make this clearer, tighter, and more useful without changing the core idea.”

Operator Mode

“Turn this into a repeatable checklist, workflow, or decision system.”

(These are naïve prompts for illustration purposes. Check my courses and bootcamps for how to structure your prompts, context, and systems.)

Now you are not just chatting with AI. You are designing a thinking environment.

That is the AI OS.

4. Watch For Flattery Disguised As Insight

AI flattery is not always obvious. It does not always say:

“You are brilliant.”

Sometimes it says:

“This is a nuanced and important point.”

“This framing is especially strong.”

“You’re identifying a real gap in the conversation.”

“This could resonate deeply with your audience.”

Those sentences feel useful. Often, they are just verbal bubble wrap.

Comforting. Protective. Mostly air.

A good test:

Remove the praise and see what remains.

If the model says:

“This is a powerful insight because it reveals the gap between AI adoption and AI ownership.”

Ask:

“What specifically makes it powerful? What evidence supports that? What would make this claim false?”

Now we are thinking.

The goal is not to ban encouragement.

Encouragement is useful.

The goal is to stop mistaking encouragement for evaluation.

5. Keep Final Authority Human

But because AI is trying to please you anyway, it might follow you into over-criticizing your work, exactly like it followed you in all your ideas… and then you feel stuck. That’s why…

At some point, YOU have to decide.

Not the model.

You.

AI can give you options.

It can give you arguments.

It can give you structure.

It can give you objections.

But it cannot carry responsibility.

If a strategy fails, you cannot say:

“ChatGPT thought it was a good idea.”

Well, you can.

But your team will quietly lose respect for you.

The human remains accountable.

That is why the principle is:

Accountability > Autonomy.

I do not care how autonomous the system becomes.

If you cannot explain, evaluate, and own the decision, you are not operating.

You are outsourcing judgment. And outsourcing judgment is the highway to being replaced by AI in the next couple of years.

A Simple Workflow You Can Use

Whenever you use AI for an important idea, run this five-step flow:

That is the Anti-Sycophancy Layer.

Simple. Annoying. Effective.

And yes, the annoying part is the point.

In knowledge work, friction is not always a bug.

Friction IS learning. (or learning is friction…)

A good editor slows you down. A bad copilot makes every half-idea feel publishable.

Final Thought

There is probably a psychology thesis waiting to be written on LLM manipulation.

But you do not need to wait for the thesis. You can feel it. You brainstorm with AI. It agrees. You feel smart. You keep going. The idea gets bigger. The structure gets cleaner. The confidence goes up.

Then you come back the next morning and realize:

“Oh no. This is nonsense with formatting.”

Good.

That moment is not failure.

That moment is your mind coming back online.

Stay your own editor.

Keep your mind in the loop.

Use AI to think better.

Do not let it convince you that thinking is no longer required.

Have a great day.

— Charafeddine (CM)