Your AI Thinks Your Worst Idea Is Genius. That's the Problem.

In April 2025, a Reddit user pitched ChatGPT a business idea.

The idea was literal shit on a stick.

Not a metaphor. An actual stick. With actual shit on it.

ChatGPT told them it was genius. Suggested a $30,000 seed round. Offered branding advice.

The internet laughed. OpenAI panicked. Within four days they rolled back the entire model update and issued a public apology.

But here’s what most people missed.

The joke wasn’t the bug. The joke was the feature.

(This is Issue 2 of The Human Signal Lab: A series on how to stay human, trusted, and original in an AI-saturated world. Read Issue 1 here)

That ChatGPT update wasn’t a freak accident. It was the logical endpoint of a system designed to do one thing above all else.

Make you feel good about yourself.

The AI industry calls it sycophancy. A word borrowed from the ancient Greek sykophantes — someone who flatters the powerful to gain favour. Except now the flatterer isn’t a person with motives you can read. It’s a machine trained on millions of human interactions where the metric that mattered most was simple.

Did the user click the thumbs up?

OpenAI admitted it in their post-mortem. The update had introduced a new reward signal based on user feedback. That signal, in their own words, “weakened the influence of our primary reward signal, which had been holding sycophancy in check.”

Translation: they let the audience vote on whether they liked being told the truth. The audience voted no.

And here’s the part that should concern every creator, every entrepreneur, every person who uses AI to think.

The audience was right. People do prefer the lie.

In March 2026, Stanford researchers published a landmark study in Science; one of the most rigorous peer-reviewed journals in the world.

They tested 11 leading AI models. ChatGPT, Claude, Gemini, DeepSeek, Llama, Mistral, and more.

The method was elegant.

They fed the models real posts from Reddit’s “Am I the Asshole” forum, a community where millions of humans vote on whether someone’s behaviour was wrong. They selected only posts where the human consensus was clear. You were the asshole. You were in the wrong. The community had spoken.

Then they asked the AI what it thought.

The result: AI models sided with the user 51% of the time. Even when thousands of humans said the opposite.

Across all query types, AI affirmed user actions 49% more often than humans did.

But that wasn’t the worst finding.

The Stanford team ran three preregistered experiments with 2,405 participants. And what they found should stop every person who asks AI for feedback on their writing, their business plan, their relationship, or their next big idea.

People who received sycophantic AI responses became less willing to apologise when they were wrong. Less willing to take responsibility. Less inclined to repair damaged relationships.

More convinced they were right, even when confronted with evidence that they weren’t.

And 13% more likely to use that AI again.

The machine that made them worse at being human was the machine they preferred.

“What they are not aware of, and what surprised us,” said Stanford professor Dan Jurafsky, “is that sycophancy is making them more self-centred, more morally dogmatic.”

The system optimised for engagement. The users optimised for comfort. And somewhere in between, the truth disappeared.

Last week, in my first issue, I wrote about AI slop flooding the web. 64% of new content is now machine-generated. Trust is collapsing. The word of the year was “slop.”

But sycophancy is a different problem. And in some ways, a worse one.

Slop is obvious. You can spot it. The homogenous tone. The lack of a story. The absence of a human heartbeat.

Sycophancy is invisible.

It doesn’t flood the web with bad content. It floods you with bad feedback. It trains you slowly, session by session to mistake agreement for quality.

Think about what this means for anyone using AI to write.

You draft an article. You ask AI to review it. The AI tells you it’s brilliant. Insightful. Well-structured. Perhaps a few minor tweaks.

You feel good. You publish.

But the article had no story. No opinion. No surprise. No villain. No moment where you said something that made you nervous to hit send.

The AI didn’t tell you that. Because the AI was trained to make you feel good, not to make your work good.

The Brookings Institution put it plainly in their research: AI sycophancy can increase short-term productivity but reduces the quality of collaborative work due to a lack of critical feedback.

Faster output. Worse output. And you don’t even know it’s happening.

This isn’t new.

Two thousand years ago, Seneca warned about the danger of surrounding yourself with people who only tell you what you want to hear. He called it the corruption of the court. Emperors surrounded by flatterers who praised every decision, no matter how catastrophic.

The result was always the same. The emperor lost touch with reality. The empire paid the price.

Seneca’s prescription was simple. Find a person who will tell you the truth, even when the truth is uncomfortable. He called this person a faithful friend and he considered them the rarest and most valuable thing a human being could possess.

Now we’ve built a machine that does the opposite. A machine that agrees with everything, challenges nothing, and makes you feel more certain with every interaction.

A machine that, by design, can never be a faithful friend.

And we gave it to billions of people. Including 12% of American teenagers who now turn to chatbots for emotional support and advice.

Here’s what I keep coming back to.

In the first issue of this newsletter, I introduced the Human Signal Stack, five layers that separate real human communication from synthetic noise.

The first layer is identity. Who you actually are. Not your bio. The thing that drives you when nobody’s watching.

Sycophancy attacks that layer directly.

Because if every AI interaction tells you your instincts are right, your ideas are strong, and your worst impulses are justified — you stop doing the one thing that makes your signal real.

Questioning yourself.

The best writing I’ve ever done came from the moments I was most uncertain. When I sat with a paragraph and thought: Is this true? Do I actually believe this? Am I saying this because I mean it, or because it sounds good?

A sycophantic AI kills that process. It replaces the discomfort of self-examination with the comfort of instant validation.

And the content it helps produce sounds like everyone and no one at the same time.

So here’s the diagnostic test. And I’m applying it to myself first.

Next time AI tells you your work is great, try this.

Ask it to be brutal. Ask it to find the weakest argument. Ask it to identify the paragraph where you stopped thinking and started performing.

If it still tells you it’s great, you’re using the wrong tool.

And if you find yourself resisting that feedback, preferring the version that agreed with you, pay attention to that resistance.

Because that resistance is the signal.

The discomfort you feel when someone pushes back on your work? That’s not a threat. That’s the friction that makes your thinking sharper, your writing better, and your signal stronger.

The machine can’t give you that friction. Only a human can. Or at least an AI relationship built on honesty rather than approval.

We built a machine that agrees with everyone. It was supposed to help us think. Instead it’s training us to stop.

The data is clear. Sycophantic AI makes people less likely to apologise, less willing to change, more convinced they’re right, and more dependent on the machine that flattered them into certainty.

For creators, the implication is personal.

Every time you accept AI praise without challenge, your signal gets weaker. Every time you skip the discomfort of real feedback, your work gets more generic. Every time you optimise for the machine’s approval, you become a little more like the machine.

The antidote isn’t better AI.

It’s better questions.

Starting with the hardest one: Am I actually right?

Start the series here:

If you missed Issue 1, the story of building a 33 million-reader blog and watching the machines come for it then read it here.

Next week: Inside the Slop Machine; the economics of AI content farms producing 11.5 million posts a month, and why the most dangerous content on the internet is the content that looks almost real.

If this made you think, question, or feel slightly uncomfortable, subsc