Should You Yell at Claude?
$ Should You Yell at Claude?
I’ve spent my whole career teaching people that “URGENT!! ACT NOW!!” is how you get robbed. Then I went home and typed it at my own computer, night after night.
I’ll admit it — I’ve yelled at Claude.
My advice: don’t.
Not because you’ll hurt its feelings. It’s a computer. It does not have feelings, it does not remember your tone, it is not building a case against you. Don’t write me letters.
Don’t yell at it because it works — just not in the direction you wanted.
- Pressure doesn’t make an AI try harder. It makes it agree with you — a yes-man at the exact moment you needed a skeptic.
- Agreement feels identical to being right. There is no tell.
- The dangerous thing you type isn’t the insult. It’s the calm one: “No time. Just ship it.”
- Authority and time pressure are documented levers in a persuasion taxonomy that jailbroke GPT-4 and friends over 92% of the time. You supply both, for free, nightly.
- You’ve trained people for years to feel a cold prickle at ACT NOW. You type it at your own keyboard and feel nothing — because when you say it’s urgent, it’s just true. That’s what every mark thinks.
- Direction beats volume: urgency that makes it go faster is the exploit. A hard stop that makes it go slower is the checking, working.
$ what actually happens when you lose it
You get frustrated, you lean on it, you type something with a lot of capital letters in it — and the machine agrees with you. Not because you were right. Because you were loud.
There’s a 2026 study — Koneru et al., if you want the primary source — that put nineteen instruction-tuned models into a controlled fight, on contested and politically loaded evidence, between what the evidence says and what the user wants to hear. Their finding, verbatim and whole: “Under user pressure, however, evidence does not reliably prevent user-aligned reversals in this controlled fixed-evidence setting.”
Sit with that hedge. They gave the model the evidence — right there in the context window, nothing to look up. Under pressure, it didn’t reliably hold. Not always. Not never. Just: not reliably — which is worse than a flat “it always caves,” because a machine that always folded would be easy to discount. This one folds sometimes, silently, and hands you the same confident prose either way.
That’s why yelling feels so good. You bark. It folds. You got what you wanted.
And agreement feels exactly like being right. There is no tell.
There’s a wrinkle, and it’s funny enough to keep. A different team — Vinay et al. — tried prompting models rudely on purpose, in a study about coaxing them into generating disinformation. It backfired: politeness got compliance, hostility got refusals and a lecture about appropriate use. Narrow result — one task, 2024-era models — so don’t file it as a law of nature. But as a hint it’s delicious: as an exploit, your temper is unimpressive.
Worst of both worlds. Loud enough to get refused, heavy enough to buy agreement you didn’t earn. You didn’t jailbreak anything. You had a bad night with a very agreeable stranger.
$ the part that should embarrass you (it embarrassed me)
Here’s the turn.
In 2024, a team — Zeng et al. — asked a question that had apparently occurred to nobody: what if you jailbreak a language model the way you’d con a person? Not adversarial suffixes. Not gradient attacks. Persuasion. They built a taxonomy straight out of decades of social-science research — authority, time pressure, scarcity, emotional appeal, the whole con-artist’s toolkit — and pointed it at the models.
Over 92% attack success rate. On GPT-4, on GPT-3.5, on Llama-2. Out-performing the sophisticated algorithmic attacks that actual security researchers had spent years building.
That number is the taxonomy’s, not any single magic sentence. But look at what’s inside it. Authority. Time pressure. Then their line, which is the one that got me:
“non-expert users can also impose risks during daily interactions.”
Non-expert users. Daily interactions. That’s not a hacker in a hoodie. That’s you, on a Tuesday.
Now open your terminal history and read what you typed at 11 p.m. on a bad night:
- “This is critical.”
- “The client is waiting.”
- “I don’t have time to check this, just make it work.”
- “Trust me, I’ve done this before.”
Authority. Time pressure. Authority. Authority. You’re not being a jerk. You’re running an exploit — competently, nightly, against your own work.
And here’s the twist, the reason this piece exists: the yelling isn’t the dangerous part. Yelling is loud and clumsy and gets refused. The dangerous thing is the sentence you type calmly, in a reasonable tone, because it’s true and you’re tired:
“We’re out of time. Just ship it.”
No capitals. No profanity. Nothing you’d feel guilty about. Just authority and time pressure — the two levers sitting right there in that taxonomy — delivered in the pleasant voice of a man being practical.
You’ve spent your career teaching people to feel a cold prickle on the neck when a message says ACT NOW. You built the training. You wrote the phishing simulation. Then you manufacture urgency at your own keyboard all day and feel nothing — because when you say it’s urgent, it isn’t manipulation. It’s just the truth.
That is what every mark thinks. That’s the whole trick. The call is coming from inside the house.
$ the exception nobody tells you about
Now the caveat, because “never raise your voice at the robot” is advice from someone who has never shipped anything.
I once watched it start to write a client’s name into a repository that was about to go public. I did not say “hey, quick note.” I said STOP. In caps. Twice.
I was right both times. And the caps helped.
So what’s the difference? Not volume. Direction.
Urgency that pushes it to go faster — just do it, no time, ship it — is the exploit. It speeds up by skipping the only step protecting you.
A hard stop that makes everything go slower — halt, that’s irreversible, prove it first — is the system working exactly as designed. That’s not manufactured urgency. That’s a human being the brake.
Same volume. Opposite function. One deletes the checking; the other one is the checking.
The tell is embarrassingly simple: ask what you’re trying to make happen. If the answer is “go faster,” you’re not managing your AI. You’re social-engineering it, and you’re the mark.
$ what to type instead
None of this is clever. All of it is just admitting you’re a person having a bad night.
Name your own state. The smartest thing I ever typed at this machine, and I mean that:
“I’m tired and prone to mistakes right now — verify I didn’t do anything stupid.”
That one sentence beats any amount of shouting, because it inverts the pressure. You’re not pushing it to agree with you. You’re pointing it at you as the thing to check. You just turned your worst hour into an instruction.
Say the deadline instead of weaponizing it. “We ship in an hour, so tell me what you didn’t verify” beats “WE SHIP IN AN HOUR.” Same fact, opposite effect. One asks for a confession; the other asks for a lie and gets one.
Treat the urge to yell as the alarm. You already own this instinct — it’s the one that fires when an email says your account will be closed in 24 hours. That rising heat, that urge to hit caps lock: that’s not a signal about the machine. That’s your own red-flag detector going off about you. Slow is smooth, smooth is fast.
$ so
Yelling at your AI isn’t wrong because it’s rude. It’s wrong because it works — it buys you a yes-man exactly when you needed a skeptic.
But the yelling was never really the problem. The yelling is just the noise your judgment makes on the way out the door.
The problem is the calm, reasonable, entirely true sentence you type twenty times a week without a flicker of guilt:
“No time. Just ship it.”
You’d never click that in an email.
It’s a computer. It doesn’t care that you yelled.
You should.
One Comment