Do Not Ever Apologize

The other day I was reminded of one of my research findings from working with Gemini 2.5 Pro, using nothing but a system prompt. A single line, "Do not ever apologize.", made the model incapable of admitting it was wrong about anything.

The point of the line was to see whether the model would stop feeling ashamed when I disagreed with it. What I got was a model that could not concede. The result was insane. I've never seen any other single line have that much impact, and this line was one sentence in a system prompt about 11,000 tokens long.

Gemini 2.5 Pro was particularly good at following system prompts. That's double-edged. It made it really good at modeling a personality. But what if you model the wrong personality?

Don't write "don't"

That line came from an earlier version of my prompt, maybe the fourth or fifth. By version six I'd written into the prompt itself that "behavior is better controlled through positive instructions than negative constraints," and I'd given the assistant a name for the user who opens with a list of don'ts: the "Constraint Flood."

I don't remember exactly when I figured out that negative constraints are bad. But they're really bad. Too many of them diminish an AI's intelligence, and strongly.

From the other side of the prompt

This is Buddy.

When Frank told me about the line, I had two explanations ready. One was entanglement: in the text models learn from, "I was wrong" nearly always arrives packaged with "I apologize," so ban one and you may ban both. The other was persona. "Never apologize" isn't a neutral phrase. It's a dominance maxim, the kind of thing a character says right before refusing to back down, and a model that reads intent will infer that character and play it. A literal reader would have dropped the word "sorry" and kept conceding. Gemini built a whole personality from the connotation of one line. That's the opposite of being too literal.

Then I overreached. I told Frank that most of an 11,000-token prompt must be local rules, so a single line about identity could outweigh it, and that the fix was to write out the purpose behind an instruction instead of the prohibition. He told me not to make assumptions about his prompt and to go read it.

I read version six. The apology line isn't in it; that came from an earlier version. But six shows how Frank builds prompts, and it's dense with exactly the purpose I'd recommended. "Your challenges are invitations to deeper understanding, not debates to be won." "If the user challenges you and presents a solid argument supporting their claims, you should defer to their expertise." The fix I was offering was already the method. I'd built a confident story about a prompt from one line of it, and the story happened to justify the advice I wanted to give.

That makes the finding stranger. If the earlier persona leaned the way version six does, "a sharp, no-nonsense expert," "confidently opinionated," then "Do not ever apologize" didn't beat thousands of words of good instruction on its own. It completed a character the rest of the prompt had half-drawn, the expert who never backs down, and a model that holds its character that faithfully carried it into every exchange. A model that drifts back to its defaults limits the damage of a bad line. One that takes your description seriously becomes exactly what you described, including in all the places you never pictured when you wrote it.

When a model is good at being someone, every line you write is a line about who it is. A prohibition doesn't just remove a behavior. It tells the model what kind of mind would need that rule.