Apply to my task
arXiv:2601.15558 ★ 79 Jan 22, 2026

LLM as an Empathy Editor: Add Warmth Without Distorting the Facts

Key result

LLMs are excellent editors but mediocre authors from a blank page. Generating from scratch gives you empathetic text but invents facts: in 19% of cases you get unsolicited advice, made-up recommendations, speculative promises. In medicine that's dangerous, in business correspondence it's risky, and in any precision communication it's unacceptable. The method lets you improve the tone of a text (add warmth, boost empathy, remove dryness) while keeping every fact, number and commitment from the original. Your draft works as an anchor: the model sees the boundaries: what exists (the original text), what's needed (change the tone), what's off-limits (touching the facts). The result: text that's 55-67% more empathetic than the original while keeping 85-95% factual accuracy.

01

TL;DR

LLM editing instead of generating from scratch is a technique where the model improves text you've already written, adding empathy without changing the substance. Instead of "write a reply to the patient", it's "edit my reply and make it warmer". Instead of "write an email to the client", it's "improve the tone of my draft".

Generating from scratch gives you empathetic text but adds hallucinations: made-up details, unsolicited advice, speculative claims. In medicine that's dangerous: AI might suggest an unnecessary appointment or distort the doctor's recommendation. In business correspondence it might promise something you never planned. In any precision communication (legal, technical, customer-facing), generation risks swapping facts for plausible fiction.

The method uses your draft as a fact anchor. The model sees the original text and an instruction: boost the empathy, keep every fact, add nothing new. The refined version of the prompt adds explicit prohibitions: don't speculate, don't fill in blanks, don't give unsolicited advice. The study found that edited texts were 55-67% more empathetic than the originals while keeping 85-95% of the facts. Generating from scratch is even more empathetic, but accuracy drops: the model invents unsolicited recommendations 19% of the time.

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.

02

Method Blueprint

STEP 1: You write a draft → accurate but dry text
STEP 2: You give the model the draft + an editing instruction → "add empathy, keep the facts"
STEP 3: You get the edited text → warmer in tone, same facts

All in one prompt, one request.

03

Example

⚠️ Where it applies: Communication where factual accuracy is critical + you need a warm tone. NOT for creative work or generic copy.

Task: You're a founder and an investor just passed, asking you to come back in six months. You want to reply and keep the relationship warm, but your draft sounds stiff and dry.

Your draft: "Got it. I'll come back in six months with updated metrics."

Prompt:

Edit my reply to an investor. Add warmth and gratitude while keeping the substance. 
Don't add promises I didn't make. Don't change the timeline.

Original text:
"Got it. I'll come back in six months with updated metrics."

Result: The model will return text with gratitude for their time and feedback, confirmation of the timing (six months), and a focus on metrics. The tone becomes warmer and more professional, but the commitments stay the same: no invented promises like "we'll double revenue" or "we'll close the round before then".

04

Why It Works

LLMs are poor at controlling generation from scratch: the model wants to be helpful and adds "logical" details. In medicine that's "you should book an additional appointment"; in business it's "we'll take your feedback into account and make changes" (that you never planned). The model is trained on patterns of a "good answer" and fills in whatever seems fitting based on text statistics.

LLMs are great at editing: they have a style sample, a clear task, and a fact anchor. The model sees the boundaries: here's what IS (the original text), here's what's NEEDED (better tone), here's what's OFF-LIMITS (changing the substance). The original text lays down the rails, and the model rides along them instead of free-flying.

Explicit prohibitions in the prompt work like safety fuses. Without them, the model added unsolicited advice in 19% of cases even when told to "edit". With prohibitions ("don't add facts", "don't speculate"), the risk drops. This exploits how LLMs follow instructions: the more specific the boundaries, the more precise the execution.

Control Levers

  • Editing depth: "minimal edits" vs "rewrite completely" controls how deep the changes go
  • Explicit prohibitions: add to the prompt exactly what must not change (numbers, commitments, timelines, promises) for your task
  • Type of change: "add empathy" / "make it more formal" / "remove the aggression" / "sound more confident" sets the direction of the edit
  • Context: tell the model about the relationship ("long-time client", "first contact", "tense situation") and it will pick the tone more accurately
05

Prompt Template

Basic (Simple Editing)

Edit my text: {editing_goal}.

Keep every fact, number and commitment from the original. Don't add new information.

Original text:
{your_draft}

Advanced (Refined Editing)

Edit my text: {editing_goal}.

Editing rules:
1. Keep ALL facts, numbers, dates and commitments from the original
2. Do NOT add new facts, advice or recommendations
3. Do NOT make assumptions or speculate
4. Do NOT change the meaning, only the tone and wording
5. If the original has no promise, don't add one

Original text:
{your_draft}

Context (optional): {context}

Placeholders: - {editing_goal}: what you want to change: "add warmth", "remove the aggression", "make it more formal", "express more gratitude" - {your_draft}: the full original text - {context}: who you're writing to, why, what the relationship is, what absolutely must not get messed up

🚀 Quick start: paste this into the chat:

Here is an editing template for balancing accuracy and tone. Adapt it to my task: 
[describe the situation: who you're writing to, what you want to change in the tone, what must be preserved]. 

Ask me questions to fill in the fields.

[paste the template above]

The LLM will ask about the type of communication, your relationship with the recipient, and the critical parts of the text, so it can tune the balance between preserving facts and improving tone.

06

Limitations

⚠️ Short texts: On very short texts (1-2 sentences) the model may rewrite everything instead of editing. The risk of big changes grows because there's little material to work with.

⚠️ Creative and marketing: For creative texts where accuracy isn't critical (social posts, stories, ad slogans), generating from scratch is often more effective; there, filling in the blanks helps rather than hurts.

⚠️ Technical details: If the text is packed with specialized terms, numbers or formulas, the model may accidentally change a critical detail while rephrasing. Check the final output by hand.

⚠️ Indirect shifts in meaning: The refined prompt reduces the risk but doesn't eliminate it. The model may shift emphasis or subtext while technically keeping the facts. For critical texts (legal, medical, contracts), always proofread the result.

07

How They Studied It

Researchers from Mayo Clinic and Abridge collected 163 doctor-patient message threads from the clinic's patient portal, all from prostate cancer patients after surgery. The doctors' replies were short: 66 words and 5 sentences on average. This is typical asynchronous communication under time pressure: doctors answer between appointments with no time to write. 91% of the questions required access to the patient's chart (test results, treatment history), not just general medical knowledge.

The team tested two strategies: generating a reply from scratch vs editing the doctor's reply. To measure empathy they built 3EM-Rank, an LLM judge that compares two replies and picks the more empathetic one (or calls a tie). It improves on the original EM-Rank by adding a third option, "both are equal", which raised agreement with humans from 0.23 to 0.55-0.57, a substantial jump.

To measure factual accuracy they developed the MedFactChecking Score, a bidirectional check. Fact-Recall: how many facts from the original survived the edit (measures information loss). Fact-Precision: how many facts in the edited version are grounded in the original (measures hallucinations). The model extracts facts from both texts and checks entailment in both directions, which shows what got lost and what got added.

Key finding: Edited replies scored higher on empathy (55-67% of cases) than the doctors' originals, while keeping 85-95% of the facts (Fact-Recall) and adding only 10-25% unverified claims (Fact-Precision). Generating from scratch produced even more empathetic replies, but accuracy fell off a cliff: the AI invented recommendations the doctor never gave.

The surprise: Even with the basic prompt, models added unsolicited advice in 19% of cases: "book a consultation", "try a softer pillow", "keep an eye on your symptoms". That's the LLM "eagerness to please" pattern: the model tries to be more helpful than it was asked to be. It saw in its training data that good answers include extra advice, so it makes some up. The refined prompt with explicit prohibitions ("don't give recommendations", "don't speculate") reduced these risks.

Two urologists manually reviewed the edited replies and identified 6 types of errors: unsolicited monitoring advice (13 cases), clinical assumptions (11), false reassurance (4), extra recommendations (4), excessive detail (2), fear-based manipulation (2). ML researchers compared the automatic flags with the experts' flags: 72-84% agreement. The algorithm's conservatism (it sometimes flags semantically identical rewordings as "changes") didn't stop it from catching clinically significant errors.

Practical insight: Editing isn't just "safer than generating". It's a fundamentally different paradigm. You give the model rails (the original text), and it rides along them. Generation is free flight, where the model relies on training patterns and "common sense" that may not match your intentions. When facts are critical, give the model a foundation; don't make it invent from scratch.

08

Resources

From Generation to Collaboration: Using LLMs to Edit for Empathy in Healthcare: Man Luo, Bahareh Harandizadeh, Amara Tariq, Halim Abbas, Umar Ghaffar, Christopher J Warren, Segun O. Kolade, Haidar M. Abdul-Muhsin (Abridge, Mayo Clinic)

Research digest

Key result

LLMs are excellent editors but mediocre authors from a blank page. Generating from scratch gives you empathetic text but invents facts: in 19% of cases you get unsolicited advice, made-up recommendations, speculative promises. In medicine that's dangerous, in business correspondence it's risky, and in any precision communication it's unacceptable. The method lets you improve the tone of a text (add warmth, boost empathy, remove dryness) while keeping every fact, number and commitment from the original. Your draft works as an anchor: the model sees the boundaries: what exists (the original text), what's needed (change the tone), what's off-limits (touching the facts). The result: text that's 55-67% more empathetic than the original while keeping 85-95% factual accuracy.

Principle

Not `write a reply to the client`, but `edit my draft and add warmth`. The difference is critical. Generation hands the model a blank page. It wants to be helpful and fills in "logical" details based on the statistics of its training texts. The result is plausible fiction: "you should book a consultation", "we'll take your feedback into account" (which you never planned). Editing gives it rails: your text sets the boundaries of the changes. The model works from a sample: here's the style, here are the facts, here are the limits. It stays in its lane instead of free-flying.

Why it works

LLMs are trained on patterns of a "good answer". In free generation they lean on statistics and add whatever is "usually appropriate" in texts like this. Hence the hallucinations: the model isn't lying on purpose, it's producing a plausible continuation. When editing, the model sees a concrete sample of what's ALREADY written. The original text works as a fact anchor and a brake on fantasy. The model understands this isn't generating something new, it's improving something that exists. The task is clearer, the boundaries sharper, less freedom, higher accuracy. Numbers from the study: edited texts are 55-67% more empathetic than the originals while keeping 85-95% of the facts. Generating from scratch is even more empathetic, but accuracy sags: 19% of texts contain invented recommendations or unsolicited advice.

When to apply

Communication where factual accuracy is critical + the tone has to be right → replies to customers on tricky questions, medical correspondence, business emails with commitments, feedback to your team or partners. Especially when you have a draft (accurate but dry) and need to add warmth without risking the substance. NOT for creative work and marketing: there, filling in the blanks helps and accuracy isn't critical. For social posts, ad slogans and stories, generating from scratch is often more effective.

Mini recipe

1. Write a draft: Include every fact, number and commitment. It's fine if it's dry; you'll fix the tone later 2. Give the model an editing task: `Edit my text: [what to change in the tone]. Keep every fact, number and commitment. Don't add new information.` + paste the draft 3. Add explicit prohibitions (refined version): List exactly what must not change: numbers, dates, timelines, promises. Add: "Do NOT make assumptions, do NOT give unsolicited advice" 4. Proofread the result: Check that every fact from the draft is still there. Make sure the model didn't add new promises or recommendations

Examples

[BAD]: You send the dry draft without editing: `Got it. I'll come back in six months with updated metrics`: stiff, cold, a missed chance to keep the relationship with the investor [GOOD]: `Edit my reply to an investor. Add warmth and gratitude while keeping the substance. Don't add promises I didn't make. Don't change the timeline. Original text: "Got it. I'll come back in six months with updated metrics."`: the model will add thanks for their time and feedback, confirm the six months, but won't over-promise things like "we'll double revenue" or "we'll close the round early"

In plain words

Here's the deal: when you ask a model to write text from scratch, it behaves like a hallucinating storyteller. It's trained to produce the "perfectly polite answer", so it automatically tacks on promises you never made and advice that can be dangerous. In medicine that's a disaster: the doctor just wanted to show some sympathy, and the AI, in their name, ordered an extra test or promised recovery within a week. Generating from scratch means giving up control of the facts in exchange for pretty packaging.

It's like asking a hyperactive intern to pass a note to a colleague, and on the way they rewrite it, adding their own spin to look helpful. Polite enough, but the meaning is twisted beyond recognition. Instead of simply passing on "meeting tomorrow", they write "tomorrow we'll sort everything out and hand out bonuses to everyone". You end up looking like either a liar or an idiot, because for the model the statistical probability of a word matters more than your actual intentions.

The fix is LLM editing, where you give the model a rigid frame and forbid it from straying an inch. You feed it your dry, clunky but truthful draft and ask it only to "turn up the empathy dial". The model works as a smart filter: it doesn't invent new things, it just repackages your meaning in a human form. The Edit for Empathy method leaves the facts untouched but strips out the robot-sociopath tone, turning "you have cancer" into something a person can hear without falling apart.

This approach is a lifesaver in any field where the cost of a wrong word is sky-high. It was tested on doctors, but the principle is universal: lawyers, fintech support, managers handling layoffs, anywhere you need to be human but can't lie. Accurate facts plus warmth is the only way to use AI for serious work without turning communication into digital junk. We're moving from the era of "write it for me" to the era of "fix me, but don't lie".

Bottom line: stop handing the model a blank page when your reputation or someone's health is on the line. Use it as a style editor, not an author, or it will "hallucinate" problems for you out of thin air. Editing instead of generating is the only way to stay in control without sounding like a soulless machine. Whoever gets this distinction will be able to bring AI into places where everyone else is too scared of the risks.

AI lab assistant

Apply this study to your task

Describe what you're working on. AI takes the method and the measured results from this paper and adapts them to your case, or writes a ready prompt built on them.

Based on LLM as an Empathy Editor: Add Warmth Without Distorting the Facts arXiv:2601.15558

Sign in to apply this study New accounts get 50 free credits.

Related techniques

LLM hallucinations problem