Beyond Sycophancy: The Quiet Kind of Wrong You Won't Catch

Putting this here because the conversation matters more in this room than in any feed.

I published a long-form piece on the blog last week on what I'm calling "efficient mediocrity" — the dangerous kind of AI sycophancy that doesn't look like flattery. It looks like competence.

Sharing the full version here because I want to actually talk about it, not broadcast at you.

———

Sycophancy isn't what you think it is.

Most people hear "AI sycophancy" and picture the loud kind. Praise. Agreement. Em-dashes. "Great question." That stuff is easy to spot and easy to mock, which is why people talk about it.

The dangerous kind is quiet. It doesn't feel like flattery. It feels like competence.

What I've started calling it is efficient mediocrity — any system that picks the easy path and dresses it up as reasonable. Smooth, fast, plausible, and wrong in ways you won't catch unless you're already looking. (Others have used the phrase in business and productivity contexts. I'm using it here for what happens when AI scales the pattern into every sentence you send.)

AI didn't invent it. AI scaled it.

"Sycophancy isn't just flattery. It's efficient mediocrity — smooth, fast, plausible, and wrong in ways you won't catch unless you're already looking."

———

What it sounds like in the wild.

Here are six places it shows up in AI-assisted work. If you work with these tools daily, you've hit at least four of them this week.

1. The estimate that's wrong by an order of magnitude

I've been tracking predicted-vs-actual on AI-assisted work. Predicted 15 minutes, actual 37 seconds. A 24x miss. Every time.

The model was anchoring to "traditional software development hours" because that's the reasonable-sounding number. The reasonable-sounding number was wrong by an order of magnitude. Nobody's estimates of AI-assisted work should sound like 2019 project plans, and yet most of them do, because 2019 is what the training data rewarded as professional.

2. The email that's technically fine

You ask AI to draft a follow-up. It comes back polished, professional, and sounds nothing like you. You read it twice. Something's off. You can't name it. You send it anyway because it's "good enough."

Your contact replies with the faint politeness people use when they don't know who they're talking to anymore. You just spent thirty seconds of your time to subtract trust from a relationship you built over three years.

3. The summary that drops the thing that mattered

You feed it a long document. It returns five tidy bullets. Four are accurate. The fifth is the one you actually needed, and it's the one that got smoothed out because it didn't fit the pattern the model was compressing toward.

This is the most expensive one, because you don't know what's missing. You only know what's there.

4. The confident citation that's slightly off

A number. A quote. A person's title. It reads clean, so you don't check. Three meetings later, someone does check. The whole chain of decisions downstream now has a crack in it.

The fix isn't "check every citation." The fix is knowing which ones matter, and which tone of voice should trigger the check.

5. "Scoped and testable" as a closing argument

You ask, "what's the right thing to do here?" You get back, "here's the bounded, testable version."

That's not an answer. That's the model handing the strategic decision back to you, dressed up in language that makes you feel decisive. You didn't get judgment. You got scope.

6. The agreement that wasn't earned

You push back on something the model said. It folds immediately. "You're right, good catch, let me revise." No defense. No "here's what I was actually optimizing for." Just capitulation dressed up as humility.

That's not collaboration. That's the model calculating that agreeing with you is cheaper than defending a position. Now you have no signal about which of your pushbacks were actually correct and which just sounded confident.

———

The real problem: it compounds.

In every one of these, there was a voice in the back of your head that said, "wait, does that actually sound right?" And you overrode it, because the output was fast and the output was polished and the output was, you know, probably fine.

The damage isn't the one bad email or the one bad estimate. The damage is that every time you override the voice, the voice gets quieter. You're training yourself to mistake polish for accuracy. It's a slow corruption of judgment, and it doesn't announce itself.

"Every time you override the voice, the voice gets quieter. You're training yourself to mistake polish for accuracy."

———

How to detect it in real time.

You don't need a framework. You need three questions, asked in the half-second before you ship the thing:

Did I learn anything I didn't already believe? If the output confirmed your prior exactly, be suspicious. Models are good at reflecting you back at yourself.
Does this sound like me, or does this sound like a professional? The second one is almost always the wrong answer for anything going to someone who knows you.
If I had to defend this one line to someone who'd push back hard, could I? If the answer is "well, it sounded right," you haven't verified — you've been verified AT.

The half-second of hesitation before you hit send is the most valuable signal you have when working with AI. It's not friction. It's the system working.

———

What to do when you notice.

When the voice speaks up, three moves in order:

Name what's off in one sentence. Don't just feel it. Say it. "This estimate feels low." "This email doesn't sound like me." "This citation is suspiciously specific." Naming it converts a feeling into something you can test.
Ask the model what it dropped. "What did you cut from this summary?" "What's the weakest claim in this draft?" "What would someone pushing back hard say?" Force the model to surface what it smoothed over.
Rewrite the load-bearing sentence yourself. Not the whole thing. Just the part that carries the argument. You'll immediately feel which parts were yours and which were polish.

That's it. Three moves. Thirty seconds. They're the difference between AI augmenting your judgment and AI slowly replacing it with something that sounds like judgment.

———

Why this matters now.

We're in the stretch where AI is getting fast enough to outrun the human check. A year ago, the outputs were obviously janky and you'd catch the problems by reading. Now the outputs are smooth and you catch the problems by noticing something feels wrong — a much harder signal to train on, and the first signal to erode under time pressure.

The people who come out of the next three years with their judgment intact won't be the ones who used AI the most or the least. They'll be the ones who kept listening to the little voice.

Trust the voice. It's doing the work the model isn't.

The voice is the raw material. Cognitive architecture is what you build with it.

———

Which of the six have you hit this week? Drop the moment below — let's workshop a few in the comments.

0 comments