The Culture Funnel You Cant Align What Isnt In The Data

Skip to content Products

Solutions

Resources

Blog

Research

Company

Sign in

Request a demo

Platform North

Enterprise-ready AI for business

Compass

Intelligent search and discovery

Models Command New

Generative language models

Transcribe New

Speech recognition model

North Mini Code New

Agentic coding model

Embed

Search and discovery model

Rerank

Semantic search ranking

Models Overview

Product Products Overview

Total Cost of AI Ownership

Pricing

Featured Command: High-performance generative AI models for real-world applications

Deploy Model Vault

Dedicated model inference platform

Private Deployments

On-prem or isolated VPCs

Security

Protect your data at every stage

See deployment options

By Industry Financial Services

Public Sector

Technology

Telecommunications

Energy and Utilities

Healthcare and Life Sciences

Manufacturing

Featured Model Vault provides fully-isolated, performant inference with Saas simplicity

Insights Customer Stories

For Developers Developers

Models Overview

Docs

Discord

LLM University

Connect Partners

Events

Webinars

Merch Store

Featured How CoreWeave used Cohere North to transform its customer support in 90 days

Blog

The latest news, launches, and insights

Read more

Cohere and the University of Toronto partner to advance responsible AI adoption at scale

Hardware-aware dynamic speculative decoding

Meet Cohere Transcribe Arabic

Research Cohere Labs

Cohere’s ML research lab

Explorations Future(s) of Work

How will AI change the way we work?

Aya Models

Multilingual AI at scale

All Papers

Initiatives Research Scholars

Finding the new generation of ML talent

Open Science Community

Championing global, open science

Catalyst Grants

Supporting impactful ML endeavors

Resources Blog

Hugging Face

Events

Featured The future of work debate has an evidence problem

About

Careers

Newsroom

Aug 19, 2026

8 minute read

The Culture Funnel: You can’t align what isn’t in the data Cohere Labs analyzed data from modern LLM training pipelines and found that cultural diversity is frequently lost in post-training data mixes.

Key takeaways Building globally inclusive AI requires moving beyond multilinguality. A model may be fluent in dozens of languages yet still miss the cultural norms, values, and social contexts that shape how people communicate. Moving from multilingual to multicultural AI is the next step toward building systems that better serve people around the world. Current approaches to aligning LLMs with different cultural perspectives focus on inference-time interventions, which assume that models already contain sufficient cultural knowledge and it only needs to be elicited when a user’s prompt requires it. However, findings from our analysis of popular datasets from different stages of the training pipeline challenge this assumption. We find that modern LLM pipelines suffer from a ‘cultural data funnel’: post-training data loses substantial cultural diversity, with domain selection further shaping what cultural content remains. Our findings suggest that AI developers must take a closer look at how culture is expressed in training data itself. Scaling LLMs’ multilingual capabilities alone does not guarantee culturally rich representation. It requires having data that contains broader regional coverage and intentional curation of culturally representative content". Making culture more explicit within the training data itself shows promising results in helping LLMs better retain more diverse and underrepresented aspects of culture.

Link to full paper: https://arxiv.org/pdf/2606.13808 Link to dataset: https://huggingface.co/datasets/CohereLabs/CultureMarkers

As AI systems become increasingly global, focusing on multilingual coverage alone is not enough for building systems that serve people around the world. A model may answer fluently in dozens of languages and still miss the context behind how people communicate local norms, values, assumptions, social expectations, or the ways culture shapes everyday interactions. Cultural awareness is becoming just as important as language coverage . But why do cultural gaps persist?

A core challenge to studying this question is that culture itself is inherently difficult to quantify: it is expressed implicitly across nearly all user interactions, shaped by context, and cannot be fully represented through language or geography alone. A common assumption is that the cultural knowledge already exists inside a trained language model and simply needs to be elicited through prompting, alignment, or better reasoning. But what if the limitation starts much earlier? In this work, we look upstream at the data itself.

To better understand where cultural gaps might originate in training data, our recent preprint, The Culture Funnel: You Can’t Align What isn’t in the Data , introduces a framework for surfacing and examining cultural signals at scale within training data. Using this lens, we trace how cultural representation changes across different stages of the LLM training pipeline and explore what these patterns suggest about building AI systems that better account for diverse cultural contexts.

After analyzing over 5.6 million training data samples across different stages of the language model training pipeline, we found a consistent pattern: cultural diversity narrows as data moves from pretraining into post-training. We call this phenomenon the culture funnel. Our analysis investigates several factors associated with this pattern, including post-training dataset composition, long-tail effects in cultural content, and data curation practices.

With our analysis, we extend a principle suggested by a team of researchers in 2025 : “Every evaluation and data choice should be examined for culturally contingent considerations”, establishing culture as a primary factor in data documentation, processing, and evaluation. Looking for culture along the training pipeline To understand how cultural information appears throughout model development, we selected popular datasets used across different stages in LLM training - pretraining, SFT, alignment, and reasoning - and tagged a sample of prompts from each source for cultural signals. We used Cohere’s Command A model to automatically tag prompts following instructions we gave it and we validated the quality of generated tags by manually reviewing a subset of prompts across six languages. Importantly, we do not treat language...