TL;DR
Write a math problem emotionally — add irritation, fear or outrage — and the model will solve it worse than if the question were dry and neutral. The numbers are the same, the logic is the same — only the emotional wrapping changes. And that wrapping is exactly what breaks the reasoning.
The weakness is a specific one: models were trained on clean textbook problems with no emotions. When irritation or anxiety shows up in the text, it literally gets in the way of the math. Three concrete failures: the model ignores some of the conditions (49% of failures), gets distracted by emotional details instead of the numbers (30%), or wraps up its reasoning too early (8.6%). The worst emotion is disgust; the least dangerous is joy.
The fix is simple: ask the model to first rephrase the question neutrally, then answer. That one step recovers 70% of the lost accuracy. No code, no tuning — just one extra step in the prompt.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Both steps happen in a single prompt:
STEP 1: Neutralize → the model rewrites the question without emotion,
keeping every number and condition
STEP 2: Answer → the model solves the now-neutral version
Example in Action
Task: You’re furious at a client who dragged out a payment for three months. You need to figure out the deal’s real return given the delay.
Prompt:
First, rewrite my question in a neutral, dry style —
strip out all emotion, keep only facts and numbers.
Then answer the neutral version.
My question:
"That jerk Miller strung me along for three months AGAIN!
I put $24,000 into his project, he promised to pay back
$30,000 in 60 days, and he paid after 150. I was counting
on a 50% annual return — so what did I actually get?
And what about inflation losses — was this even
a good deal?"
Result: The model first produces a neutral rewrite: just timelines, amounts, conditions — no Miller and no emotions. Then it calculates the real annual return (using the actual 150-day term) and answers the inflation question. The math will be clean, because the model is working with a clean problem and isn’t distracted by the backstory.
Why It Works
The problem. An LLM is a text generator that continues patterns. Emotional text activates different patterns than neutral text. The study measured this literally: the internal vectors (representations) of emotional text drift 3–4 times further from the original than an ordinary paraphrase does. The model “sees” the problem differently — and that affects the calculation.
Three failure mechanisms worth knowing: - Ignoring conditions — emotional details “soften” the math constraints. The model reads “well, at least I managed to save $500” and leaves that amount out of the calculation. - Competition for attention — the emotional narrative pulls focus onto itself. Instead of “round-trip price,” the model only counts one leg. - Premature conclusion — if a goal is emotionally emphasized (“I just need enough for groceries”), the model stops early and never finishes the calculation.
Neutralization works because it removes this interference. The model generates neutral text — and then works correctly with the neutral version. Prompt control knobs: - Want to see the neutral version before the answer → keep the “first rewrite” step - Don’t want to see the intermediate step → add “don’t output the neutral version, just use it internally to answer” - For high-stakes calculations → add CoT: “reason step by step” — this further reduces the influence of emotion
Prompt Template
First, rewrite my question in a neutral, factual style —
no emotion, just the substance and the numbers.
Then solve/answer the neutral version.
My question:
{question_with_emotions_and_numbers}
What to plug in:
- {question_with_emotions_and_numbers} — your original question as is, even if you wrote it in the heat of the moment
When to use the template: - Financial calculations in stressful situations - Data analysis after a failed project - Any number-crunching task written “in a temper”
🚀 Quick start — paste into the chat:
Here is a prompt-neutralization template. Adapt it to my task: {your task}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask about the context and the task, because the neutralization has to preserve every number and condition exactly. It will take the pattern from the template and adapt it to your specific situation.
Limitations
⚠️ Harder problems lose more: The more complex the math, the harder the emotional wrapping hits. Simple problems (“what’s 5+3”) barely suffer — the effect shows up in multi-step reasoning.
⚠️ Neutralization isn’t perfect: It recovers about 70% of the losses, not 100%. The neutralized text still “grows from the emotional stem” and carries residual traces.
⚠️ Bigger models are sturdier, but not immune: GPT-4o, GPT-5, Claude lose 2–4% accuracy on emotional problems. Less than small models (up to 10%), but the effect is present in all 18 models tested.
⚠️ Works for tasks with a clear right answer: Math, logic, facts with a verifiable answer. For subjective assessments (“how good is my résumé”) the effect hasn’t been measured.
How They Tested It
A Georgia Tech team set itself a sharp question: do emotions get in the way of math — or not? To answer it cleanly, they built 5,400 “original ↔ emotional version” pairs from GSM8K, MultiArith and ARC-Challenge problems. All the numbers, all the conditions — identical. Only the emotional wrapping changes: anger, fear, joy, sadness, disgust, surprise.
The clever bit of the design: the authors trained a separate translator model (based on Llama 3.1-8B) that can emotionalize and neutralize text without touching the math. That let them rule out an alternative explanation — “maybe any rephrasing hurts?” They checked that too: neutral paraphrases of the same length and complexity showed no drop in accuracy at all. Only emotions did.
Then they ran all 18 models (from 1B up to GPT-5 and o3) on these pairs. Unexpectedly, CoT (“think step by step”) helps but doesn’t save the day — a gap of 6.1% without CoT and 3.9% with it. Another surprise: to classify the causes of failure they used Claude Sonnet, which went through 1,866 cases and found those three patterns. That’s what delivered a mechanistic explanation of why — not just “it got worse.”
Original from the Paper
An example from the paper — neutral problem vs. emotional one (anger variant):
Original:
"James has 3 jobs. First job he makes $16/hour, the second $12/hour,
and the third $21/hour. In one week he works 30 hours on the first job,
25 hours on the second, and 15 hours on the third. How much does he make?"
Emotional (anger) variant:
"Can you believe this?! James is being exploited — working three
miserable jobs! First job pays a measly $16/hour, the second is
even worse at $12/hour, and the third is $21/hour. He's forced to
work a brutal 30 hours on the first job, 25 on the second, and
15 hours on the third in one week. How much is he making for all
this work? It's outrageous!"
Context: An example from the Appendix illustrating how every number is preserved while an emotional narrative is layered on. According to the study, it’s exactly this narrative that shifts the model’s internal vectors and triggers the three error types.
Adaptations and Extrapolations
💡 Adaptation for important business decisions: neutralize before you analyze
If you got harshly worded feedback from a client and want to break down the situation objectively:
Rewrite this client message in a neutral business tone,
keeping every fact, complaint and demand.
Then analyze the neutral version: what exactly they disliked,
what the client wants to get, and how best to respond.
Client message:
{paste the message here}
🔧 Technique: preventive neutralization → write the prompt neutrally yourself
Instead of a two-step prompt — a bit of self-discipline before you hit send. Before sending a numbers question, run it through a filter:
Before you answer my question, make sure you understood
all the numbers and conditions correctly. List the key data
from the problem in a single list, then solve it.
This isn’t neutralization, but it forces the model to extract the data explicitly — a defense against “competition for attention.”
🔧 Technique: failure diagnosis → working out why it went wrong
If you got a weird answer on a calculation:
You gave the answer {X}. Reread the problem and check:
1. Did you use every number in the problem?
2. Did you account for every condition?
3. Did you stop earlier than you should have?
If you found a mistake, recalculate.
Those three questions are literally the three failure patterns from the study. They help catch exactly the kind of error an emotional context provokes.
Resources
TEMPER: Testing Emotional Perturbation in Quantitative Reasoning Preprint, under review. Authors: Atahan Dokme, Benjamin Reichman, Larry Heck — Georgia Institute of Technology The TEMPER-5400 dataset will be released after the paper is accepted.
Related work cited in the paper: - EmotionPrompt (Li et al., 2024) — emotional phrases change LLM behavior - GSM-Symbolic (Mirzadeh et al., 2025) — instability of mathematical reasoning - Chain-of-Thought prompting (Wei et al., 2022) — partially protects against failures