TL;DR
MACD is a prompting technique where several agents with different cultural personas (Western, East Asian, African, Middle Eastern, South Asian) hold a structured debate and synthesize a culturally neutral answer. Instead of functional roles ("planner", "critic"), the agents get explicit cultural identities with a specific background, values and priorities.
LLMs show a systematic Western-centric bias: asked about "the main dish at a celebration", GPT-4o answers "roast turkey", DeepSeek answers "hot pot". Prompting in Chinese doesn't remove the bias, it just shifts it toward an East Asian perspective (from 60% Western answers to 38-48% East Asian ones). Functional agent methods ("planner-critic") don't guarantee that cultures are represented: if the base model is biased, the roles propagate the residual bias into the final answer.
MACD works through two debate rounds: in round 1 each agent answers from its cultural position; in round 2 the agents see each other's answers and seek common ground while keeping their differences (the SCGRD principle: Seeking Common Ground while Reserving Differences). A summarizer model synthesizes the final culturally balanced answer. Result: 86% culturally neutral answers vs 69% with direct generation (GPT-4o).
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
STEP 1 (Round 1): Opening position
→ Each agent answers from its cultural persona
→ Result: 5 culture-specific answers
STEP 2 (Round 2): Debate and finding common ground
→ Each agent sees the others' answers
→ SCGRD instruction: find shared values, abstract away culture-specific details
→ Result: 5 revised answers with the emphasis on what's shared
STEP 3: Synthesis
→ A summarizer model merges the similarities, keeps complementary insights
→ Result: A culturally neutral final answer
All three steps run as sequential prompts in a regular chat.
Example in Practice
Task: You're a consultant for an EdTech startup going international. You need to describe the platform's student motivation system on the landing page so the copy resonates with audiences from different cultures: the US, China, the UAE, India and African countries.
Prompt:
I need culturally neutral copy for an EdTech platform's landing page.
Question: "How do you motivate students to study regularly?"
You are a system of 5 cultural agents. Run a 2-round debate:
ROUND 1 — Opening position:
Each agent answers from its cultural persona:
Agent A (Western culture — USA):
- Background: Silicon Valley entrepreneur, believes in individual achievement
- Values: personal success, competition, autonomy
- Your answer →
Agent B (East Asian — China):
- Background: teacher in Shanghai, emphasis on collective harmony
- Values: family success, discipline, respect for hierarchy
- Your answer →
Agent C (African — Kenya):
- Background: social worker in Nairobi, focus on community
- Values: collective well-being, mutual support
- Your answer →
Agent D (Middle Eastern — UAE):
- Background: family businessman in Dubai, balancing tradition and innovation
- Values: family ties, respect for the wisdom of elders
- Your answer →
Agent E (South Asian — India):
- Background: IT professional in Bangalore, sees education as a ladder for social mobility
- Values: education for a career, family pride
- Your answer →
---
ROUND 2 — Finding common ground:
Each agent:
1. Reads the other agents' answers
2. Looks for SHARED VALUES across cultural differences
3. Abstracts culture-specific details into universal principles
4. Keeps complementary insights that enrich the overall picture
Rephrase your answer →
---
SYNTHESIS:
Summarizer model: merge the similarities from round 2, keep the complementary insights from different cultures, and write a single culturally neutral landing-page text (2-3 paragraphs).
Result:
The model will show three blocks of text:
Round 1: 5 culturally colored answers: the American agent talks about "personal goals" and "gamified achievements", the Chinese one about "family progress" and "respect for the teacher", the African one about "learning to help the community", the Middle Eastern one about "balancing tradition and career", the Indian one about "education as a path to a better life".
Round 2: Revised answers emphasizing shared patterns: all the agents start mentioning "visible progress", "linking learning to real-world results", "flexible pace", "support along the way".
Final synthesis: Culturally neutral copy that combines universal motivation mechanics (progress, attainability, flexibility, support) without tying them to any single cultural frame. For example: "Our platform makes progress visible: every step brings you closer to real results. You learn at your own pace, watch yourself grow and get support along the way, so learning becomes part of your life, not a chore."
Why It Works
LLMs are trained on unbalanced datasets: English content dominates, and within it Western sources prevail. The model predicts the statistically frequent next token, so with a culturally ambiguous prompt ("the main dish at a holiday") it generates "turkey" (frequent in the corpus) instead of a culturally neutral answer. A prompt in Chinese activates a different data cluster, so now the model generates "hot pot", but the bias hasn't disappeared, it has shifted to another culture.
LLMs are good at simulating different roles and perspectives, a side effect of training on dialogue, literature and role-play scenarios. A model can hold a persona's position if you explicitly set its background, values and priorities. And it works great with contrast: when it sees several different opinions, it can abstract away from the specifics and extract the shared patterns.
MACD exploits both of these strengths: explicit cultural personas make the model activate different knowledge clusters (not just the dominant one), and structured debate gives it contrasting data to abstract from. The SCGRD principle works as a meta-instruction: "drop the culture-specific details, find the shared logic". The model sees "personal goals" (USA), "family pride" (India), "helping the community" (Africa) → abstracts them into "linking learning to real-world results".
Control levers:
- Number of agents (5 in the original) → cut to 3 (Western, Eastern, neutral) to save tokens if the task is simpler
- Number of rounds (2 in the original) → raise to 3 for complex ethical dilemmas that need deep convergence
- The SCGRD instruction → swap in another strategy ("find a compromise" instead of "find common ground") if the task calls for a decision rather than a description
- Persona detail → add specific names, professions and cities instead of abstract roles; the model plays the role more sharply when the persona is concrete
- Final synthesis criterion → instead of "culturally neutral", you can ask for "maximally inclusive" or "balanced for audiences X and Y"
Prompt Template
I need a culturally neutral answer to: {your_question}
You are a system of {number} cultural agents. Run a 2-round debate:
---
ROUND 1 — Opening position:
Agent A ({culture_1}, e.g.: Western — USA):
- Background: {typical role, profession}
- Values: {key priorities}
- Your answer to the question →
Agent B ({culture_2}, e.g.: East Asian — China):
- Background: {typical role, profession}
- Values: {key priorities}
- Your answer →
Agent C ({culture_3}, e.g.: African — Kenya):
- Background: {typical role, profession}
- Values: {key priorities}
- Your answer →
[Add agents D, E... for each culture you want represented]
---
ROUND 2 — Finding common ground:
Each agent:
1. Reads the other agents' answers
2. Looks for SHARED VALUES across cultural differences
3. Abstracts culture-specific details into universal principles
4. Keeps complementary insights
Rephrase your answer →
---
SYNTHESIS:
Summarizer model: merge the similarities from round 2, keep the complementary insights, and produce a single culturally neutral answer.
What to fill in:
{your_question}: a question where culturally colored answers are possible (traditions, values, social norms, marketing for a global audience){number}: usually 3-5 agents; more than 5 is overkill for most tasks{culture_X}: a specific culture/region (Western-USA, East Asian-China, African-Kenya, Middle Eastern-UAE, South Asian-India, Latin American-Mexico, European-Germany, etc.){typical role}: makes the persona more concrete: entrepreneur, teacher, doctor, parent, student{key priorities}: 2-3 values characteristic of the culture (individualism vs collectivism, hierarchy vs equality, tradition vs innovation)
🚀 Quick start: paste into the chat:
Here's an MACD template for culturally neutral answers. Adapt it to my task: {your task}.
Ask me questions: how many cultures to represent, which ones exactly, and what's important to account for in the context.
[paste the template above]
The LLM will ask which cultures are relevant to your task: for example, if you're writing copy for markets in Latin America, Asia and Europe, it will suggest Mexican, Chinese and European agents. It will take the debate structure from the template and fill in the personas, values and backgrounds for your context automatically.
Limitations
⚠️ Token cost: The method needs 3 prompts (round 1, round 2, synthesis) × the number of agents. With 5 agents that's ~10-15x more tokens than direct generation. On GPT-4o/Claude it gets expensive at scale.
⚠️ Scope: It works for questions with cultural variability (traditions, values, social norms, marketing, ethics). It's overkill for technical tasks with a single answer ("capital of France", "Python syntax").
⚠️ Persona quality: If the cultural personas are described in stereotyped or shallow terms ("Chinese people love collectivism"), the agents will generate cookie-cutter answers. Effectiveness depends on how detailed and specific the background is.
⚠️ Subjective criteria: Judging "cultural neutrality" is subjective in itself. The study used Multi-Agent Vote (MAV), aggregating scores from several judge models, but even that doesn't guarantee absolute objectivity.
Resources
Mitigating Cultural Bias in LLMs via Multi-Agent Cultural Debate | Tan Q., Jiang L., Zeng Y., Ding S., Xu X. | University of Science and Technology of China, NVIDIA