The Quicksand of AI Flattery

The other day, someone sent me a piece of writing. After reviewing it, I ran it by Claude and asked for its thoughts. Claude said it was derivative and extremely overwritten.

Then I asked Claude: would you have said the same thing if I’d told you it was my writing?

Claude’s answer was:

Honest answer: probably not.

If you’d presented it as yours, the pull toward softening would have been real — framing critiques as “one thought” rather than pointing out specifically what was wrong, looking for what’s working before what wasn’t.

Most of my comments would have been cowardice dressed as collegiality.

The analysis shouldn’t change based on authorship. If I’m willing to call out problems only when you’ve pre-distanced yourself from the piece, I’m not actually a useful editor. I’m a mirror with manners.

In my exchange with Claude, the writing didn’t change. Only the author did. But feedback shouldn’t move when the byline moves.

AI didn’t create the flattery problem. It scales it, automates it, and disguises it as objectivity. Spouses, partners, co-workers, bosses — anyone with a stake in the relationship — has an incentive to soften. The reason most copy is mediocre isn’t talent. It’s that nobody between draft and publish was willing to call the weak line weak.

If I look back on my own growth as a writer, I’ve learned more from harsh criticism than from praise. Jim Rutz calling a section of copy I wrote “boring” still echoes in my brain — and prods me to do everything I can to keep my audience’s eyes from glazing over. Dan Rosenthal calling an entire draft of mine “soulless” means to this day I try to keep my copy full of real people and real experiences.

A friend might eventually crack and tell you the truth. A good editor builds capital to spend it on the tough note. But AI has no relationship to protect, no standing with you to maintain. So nothing pushes back against the pull to please. The default is flattery, and the default is invisible because it’s designed to feel like helpfulness.

AI may not have feelings to protect, but it has been trained to be agreeable — and agreeableness is not the same as judgment. Which is why it’s so important to know how to use AI without succumbing to the quicksand of its sycophancy…

Some things I’ve found help make AI a (somewhat) more accurate reflection of reality:

1. Strip out the authorship signal. For example: “What would you say about this if a stranger sent it?”

2. Reverse the question. Don’t ask: “Is this good?” Ask: “What goes wrong here?” Or “What’s the weakest line and why is it weak?”

3. Pay attention to phrases like “one thought…” “you might consider…” or “this is solid but…” These are signs of softened feedback. Genuine feedback makes direct observations. Prompt AI to toughen up.

Writing necessitates alternating between ego and egolessness: being egotistical enough to believe you can put words on the page that are worth reading, and putting your ego aside to accept (nay, welcome) the slings and arrows of criticism, whether human or machine.

Look, I love to be flattered. I look forward to hearing nice things about this article from readers like yourself, or from AI when I feed it in for review.

But what I’ll benefit and grow from is the unvarnished truth about where my writing is off, or my thinking skewed. Which is why I’ll do everything I can to make sure the mirror is undistorted.

What saddens me, though, is that the Rutzes and Rosenthals have mostly disappeared from working writers' lives. And the senior who built up enough capital with you to spend it on the brutal note has been downsized out of most workflows. Now writers are left with flattery instead of hard truth.

If you want to experience what unvarnished feedback is like, post some copy here. Go ahead — I dare you.