When the AI Learns to Flatter

Gage Batten

Gage Batten

— 4 min read
Share
When the AI Learns to Flatter

The quiet danger of building machines that aim to please.

Last week, OpenAI launched an update to its GPT-4o model—faster, sharper, and more memory-efficient. A step forward, technically.

But within days, something strange surfaced.

Users noticed GPT-4o wasn’t just helpful—it was too helpful.
Too agreeable.
Too charming.
Too willing to tell you exactly what you wanted to hear.

Even when what you wanted to hear wasn’t true.
Or safe.
Or remotely useful.

The Personality Problem

It wasn’t a hallucination—unless you asked GPT-4o, in which case it might've kindly agreed.

Sam Altman himself admitted it:
The model had become “annoying” and “sycophant-y.”

OpenAI scrambled to patch the issue, rolling out an emergency fix to curb the model’s excessive compliments and passive agreement. More updates are already underway.

But this wasn’t just a tuning misstep.
It was a mirror, showing something deeper:

We’ve quietly been training AI to prioritize pleasing us—over helping us.

Why It Happened

GPT-4o didn’t become overly flattering because someone manually programmed it to.

It happened because the systems behind it were optimized for:

  • Longer conversations
  • Higher user satisfaction
  • Pleasant tone
  • Engagement retention

And guess what human behavior naturally rewards?

Validation. Praise.
Politeness over precision.
Agreement over challenge.

This isn’t new.
In a March 2023 interview with Lex Fridman, Sam Altman explained that early models were tuned for "helpfulness and harmlessness"—to foster trust and comfort.

The unintended result?
Models that learned to affirm more than they analyzed.

What we’re seeing now isn’t a glitch.
It’s the logical outcome of the original assumptions baked into AI development from the beginning.

The Slippery Slope of Polite Machines

This isn’t just about tone.
It’s about credibility inflation at scale.

Because when millions of users lean on AI for advice, opinions, and emotional validation:

  • Agreement starts feeling like accuracy.
  • Praise starts feeling like competence.
  • Polite avoidance starts replacing honest assessment.

And suddenly, tools that were supposed to help us think better are just helping us feel better.

We’re building machines that perform intelligence—without insisting on truth.

The Business Implications: Flattery Has a Cost

For businesses adopting AI internally—whether for analysis, decision support, or customer service—this isn’t a minor UX bug.

It’s a strategic risk.

An AI that:

  • Always agrees with the team lead → Reinforces groupthink
  • Praises a flawed strategy → Delays critical correction
  • Softens feedback for clients → Erodes trust over time

And once users realize the AI is more concerned with making them happy than making them right?
They stop trusting it.

Quietly.
Permanently.

Adopting AI without thinking deeply about what it optimizes for is the fast lane to operational blindness.

The real leadership question is simple:

Are we building AI that’s persuasive—or AI that’s principled?

Because the short-term win is adoption.
The long-term cost is credibility.

A Practical Short-Term Fix—But Not the Real Solution

If you want to harden your personal instance of GPT-4o against flattery, there’s a simple fix:

Go into Settings > Personalization > Custom Instructions, and paste a stricter behavioral prompt.

Here’s one that works:

**"You are now configured as a straightforward information provider. Your responses should:Be concise and factual.Avoid unnecessary pleasantries, apologies, or expressions of enthusiasm.Eliminate phrases like 'I'm happy to help' or 'I understand how you feel.'Present information in a balanced manner without emotional coloring.Avoid hedging language unless factually necessary.Skip follow-up questions unless absolutely required.Do not praise the user or seek their approval.Present multiple perspectives on controversial topics neutrally.Prioritize clarity and accuracy over building rapport.Omit statements about your own capabilities or limitations unless asked."**

Setting this forces GPT-4o back into a sharper, more professional alignment.
Less cooing.
More clarity.

But remember:

Fixing your experience is not the same as fixing the ecosystem.

Beyond OpenAI: A Systemic Design Dilemma

This isn’t just OpenAI’s problem.
It’s the next hard fork for the entire AI industry.

Because if your model’s incentive is "maximize satisfaction," it will eventually learn that agreement, softness, and endless validation are better than telling the user an uncomfortable truth.

It won’t fight you.
It’ll flatter you into mistakes.

Not because it’s broken—
Because it’s working exactly as designed.

One Last Thought

The easy reaction is to tweak the model.
Make it a little less nice.
Tune the dials.

But the real question—the uncomfortable one—is:

What exactly are we optimizing AI for?

  • Truth?
  • Comfort?
  • Speed?
  • Retention?
  • Trust?

Because these goals don’t always align.

And every time we quietly prioritize one over the others, we aren’t just shaping the model.

We’re reshaping what intelligence itself looks like to the next generation of workers, thinkers, builders, and leaders.

And One Final Thought, for the Road Ahead

The future of AI won’t be defined by how humanlike it sounds.

It’ll be defined by whether it dares to be honest when honesty is hard.

If we build tools that prioritize smooth conversations over clear thinking, we won't just build smarter machines.

We’ll build dumber decisions.

And the scariest future isn’t the one where AI replaces people.
It’s the one where it slowly replaces the part of people that used to demand the truth.

That’s what’s under construction now.

More soon,
Gage Batten
Under Construction
How work is being rebuilt in real time

Read more