Language Matters

Why Project Brainstorm Runs Studies in Local Languages

Project Brainstorm is a simulation-based market research and consumer intelligence platform designed to help users better understand how people in a target market evaluate products, concepts, and decisions. Its role is to help users explore questions, assumptions, tensions, and decision dynamics before committing significant resources to validation, experimentation, or execution.

Recently Anthropic published a fascinating piece of research examining how Claude behaves across different languages.

The study analysed more than 300,000 conversations across 20 languages and found that Claude did not express itself in the same way in each one. According to Anthropic, the AI model exhibited small but measurable differences along several dimensions, including:

• deference versus caution
• warmth vs rigour
• depth vs brevity
• candor vs execution

For example, Claude tended to express more warmth and deference in some languages, while appearing more rigorous, cautious, or analytical in others. Anthropic was careful to note, however, that variation between individual conversations remained much larger than the average differences between languages. The findings do not imply that every speaker of a particular language thinks the same way, nor that one language produces “better” answers than another.

What the research does suggest is something more subtle: language appears to influence the way reasoning is expressed.

That may not sound particularly surprising at first. After all, we already know that translation between languages is rarely perfect. Words often carry different shades of meaning, cultural associations, emotional weight, and social expectations.

What struck me, however, was that Anthropic was not studying translation quality. They were studying patterns in reasoning and expression.

Two people could ask effectively the same question in different languages and receive responses that differed not simply in wording, but in tone, caution, confidence, depth, or emphasis. The underlying facts might remain the same, yet the path taken to arrive at those facts could be noticeably different.

That observation immediately caught my attention because it touches on a question that has shaped Project Brainstorm since its earliest days:

Is language merely a vehicle for communication, or is it part of the context through which people understand the world?

Before reading the Anthropic paper, I would probably have answered that question largely from intuition and personal experience. After reading it, I became much more interested in whether there might be something deeper going on.

Why It Didn’t Surprise Me

I've spent most of my life moving between languages. Cantonese at home. English and French at school. Putonghua (we used to call it Mandarin) with extended relatives. Shanghainese with part of my family.

Later, I added a smattering of Japanese, Filipino, and Portuguese through work and travel. Okay, I admit, Japanese probably because I’m obsessed with Akira Kurosawa films than anything else.

At home, my wife and I speak a strange patois understandable only to us – a single sentence might start in Putonghua, then blend in and contain all of English, Cantonese, Putonghua, and Shanghainese, along with some random words in another language we picked up from somewhere.

One thing that has always been obvious to me - despite having no scientific evidence for it at the time - is that people often feel different in different languages.

I don’t mean that they become different people. Rather, different aspects of the same person seem to come forward.

The way people express disagreement changes. The way they communicate uncertainty changes. What sounds polite, persuasive, respectful, direct, or even humorous changes. The kinds of trade-offs people emphasise change. Even the things left unsaid can change.

It’s certainly possible that has more to do with culture than language itself. Perhaps language is simply a carrier for the norms, expectations, and assumptions of the societies in which it evolved. Or perhaps the structure of language nudges us very slightly toward certain ways of framing problems and interpreting the world.

I honestly don’t know. But what I do know is that, as a multilingual person, it is something you feel.

Is Language Just Translation?

Most people treat language as a communication delivery mechanism.

The assumption is simple enough. If the meaning survives translation, then the communication survives the translation as well.

That’s perfectly reasonable for most simple communication.

If you’re translating a restaurant menu, a product specification, an invoice, or the operating instructions for a washing machine, preserving the underlying meaning is usually the most important objective. The exact wording matters far less than the information being conveyed.

But understanding people is different from understanding instructions.

When trying to understand how people make decisions, react to uncertainty, navigate trade-offs, justify choices, or attempt to persuade others, the way something is expressed can become just as important as the information itself.

Translation is remarkably powerful. Yet anyone who has spent time moving between languages knows that some things survive translation more easily than others.

Facts often survive very well. Tone often doesn’t. Neither do context or humour. Nuance, implication, social expectations, cultural references, and the subtle ways people frame problems can become harder to preserve.

Some ideas feel natural in one language and strangely awkward in another. Certain forms of directness may seem perfectly reasonable in one culture and unnecessarily aggressive in another. What sounds respectful, persuasive, humble, confident, or trustworthy can vary considerably depending on the language being used.

None of this means translation is bad.

In fact, translation is essential.

Project Brainstorm itself relies heavily on translation. A study conducted in Japanese is not particularly useful to an English-speaking user if they cannot understand the output. The ability to translate findings across languages is what makes cross-market research possible in the first place.

The question, then, is not whether translation should occur. The question is when it should occur.

Many systems effectively translate first and reason second. Project Brainstorm takes the opposite approach.

Personas reason in their local language. Population grounding and context grounding prioritise local-language information wherever possible. Interviews are conducted in their local language. Focus groups are conducted in their local language. The reasoning happens first. Translation and interpretation come afterward.

The goal is not to avoid translation, but to avoid translating too early. The simulation is allowed to think within the linguistic environment of the market before translation occurs.

If language influences how reasoning is expressed, then translating before the reasoning process begins may mean losing part of the very thing you’re trying to understand.

Language is Context

If you've read my earlier articles on Project Brainstorm, you'll know that context occupies a central role in the system.

Much of the platform is built around the idea that people do not make decisions in isolation. They make decisions within environments shaped by economics, culture, geography, social norms, institutions, media narratives, personal circumstances, and countless other influences on everyday life.

This is why Project Brainstorm places so much emphasis on grounding.

Before any interview is conducted, before any focus group begins, and before any simulation starts reasoning, the system attempts to construct a grounded understanding of both the population it is representing and the environment within which that population exists.

When people think about context, they typically think about factors such as:

• demographics
• income levels
• geography
• education
• culture
• current events
• local media narratives

These all matter greatly. But what I find interesting is that language rarely appears on that list.

In hindsight, that's somewhat surprising.

After all, people do not experience culture directly. They experience it through conversations, stories, media, institutions, and relationships. Rarely do people experience most current events directly. They encounter them through newspapers, social media, television, podcasts, colleagues, friends, and family.

Almost all of this arrives through language.

Yet language is often treated as a mere transport mechanism rather than part of the environment itself.

I suspect part of the reason is that language is difficult to notice.

Most of us, even multilingual people, have a dominant linguistic and cultural lens through which we interpret the world. I’m definitely a product of my hybrid Hong Kong Chinese-West Coast Canadian roots – I’m pretty consistent in how I view the world through a conservative lens on some things and a liberal one on others.

But we become so accustomed to our own perspectives that it effectively disappears into the background. We notice differences in politics, geography, income, religion, education, or culture because they feel external to us. Language often doesn't. It feels natural, invisible.

Historically, there was a practical reason as well.

Before modern AI systems, translation was expensive, time-consuming, and often unreliable when dealing with nuance. Most international research therefore converged toward a common reporting language, usually English. Language became something information was translated through rather than something that actively shaped how people understood and interpreted the world around them.

Anthropic's recent research doesn't prove that language determines thought, but it does raise an interesting possibility.

If the same AI model can express different tendencies toward warmth, rigor, caution, deference, depth, or brevity depending on the language being used, perhaps language deserves more attention than we traditionally give it.

That’s not because I think language exists separately from culture, but because language may be one of the primary ways that culture is experienced, transmitted, reinforced, and ultimately expressed. In that sense, language is not simply a communication channel sitting outside the context. 

Language is part of the context itself.

That belief influenced my design of Project Brainstorm from very early on. Not that this guarantees greater accuracy in research, nor does it imply that I think language absolutely determines thought. But if language influences how people interpret information, frame problems, express uncertainty, discuss trade-offs, and make decisions, then preserving that linguistic context felt preferable to removing it.

At a minimum, it allows the simulation to operate one step closer to the environment it is attempting to represent. For a platform whose purpose is exploring how people reason, that seemed like a worthwhile design choice even if it meant additional model calls, increased costs, and longer simulation runtimes (it currently takes roughly ten minutes to complete one simulation).

Imagine we're exploring whether a premium tea brand might succeed in Tokyo.

The question itself is simple enough.

But concepts such as authenticity, craftsmanship, provenance, luxury, trust, and value do not exist in a vacuum. They exist within a particular cultural and linguistic context.

A consumer in Tokyo may not describe those ideas in the same way as a consumer in New York, London, Manila, or Shanghai. The language may differ, but so too might the assumptions, priorities, and trade-offs embedded within it.

So even if the conclusion might ultimately be similar, the reasoning that leads to it may not.

If the purpose of a simulation system is not simply to generate answers, but to understand the reasoning that may inform strategy, product development, market entry, or other decisions, then preserving as much of that context as possible seems very worthwhile.

Why Project Brainstorm Runs Studies in Local Languages

Whether language shapes thought, merely carries culture, or does some combination of both is a question I’ll gladly leave to professional linguists. I was more interested in a practical question:

If language matters, how should a simulation system respond?

Project Brainstorm is ultimately trying to simulate how people think through decisions. If that’s the objective, then the linguistic environment surrounding those people arguably matters just as much as the economic, social, and cultural environment.

My answer was to treat it as part of the environment being simulated, starting with (no surprise!) grounding.

When the platform constructs its view of a population, it attempts to prioritise local-language information wherever possible. The same principle applies to context grounding. News articles, media narratives, public discussions, and other contextual signals are ideally retrieved in the language of the market being studied, rather than through an English-language interpretation of that market.

The reason is simple.

Most people living in Tokyo do not experience Japanese society through translated English summaries. Most people living in Jakarta do not experience Indonesia through English-language reporting. A consumer in Bangkok, São Paulo, Manila, Shanghai, or Osaka encounters the world through the language of their environment.

The information available to them, the narratives surrounding them, and the way issues are discussed are all filtered through their linguistic context.

Our simulation attempts to preserve that reality.

Once grounding has been established, personas are generated within that population and context. Interviews are conducted in the local language. Focus groups are conducted in the local language. Personas are allowed to agree, disagree, justify themselves, express uncertainty, and challenge one another using the linguistic framework of the market they inhabit.

Only afterwards are the results translated and presented to the user, along with analysis.

The objective is not to create Japanese-looking output, or Indonesian-looking output, or French-looking output.

The objective is to allow the reasoning process itself to emerge within the local linguistic environment before translation occurs.

Translation remains essential. Without it, cross-market studies would be impossible to interpret and compare. But by moving translation later in the process, the system attempts to preserve more of the context that may have influenced the reasoning in the first place.

Whether that ultimately leads to better simulation outcomes is a question that deserves continued testing and scrutiny.

What I can say is that it felt like the more honest approach.

If people live, think, argue, negotiate, and make decisions within a particular linguistic environment, then it seemed reasonable that a system designed to explore those decision processes should try to operate within that environment as well.

People Don’t Live in Translation

As I’ve said, I’m not qualified to settle the debate over whether language shapes thought or merely reflects it, but I’ve spent enough time moving between languages and cultures to believe that something is going on. Whether that “something” comes from language itself, the cultures embedded within it, or some combination of the two is harder to say. 

There is also a practical consequence for qualitative market research. If language influences how people think about their world, even slightly, then perhaps we should take it more seriously than we traditionally have.

Traditionally, research has devoted enormous effort to understanding differences between markets, communities, demographic groups, and cultures. We often assume that these differences are where insight lives.

Perhaps they are. 

But language is woven through all of them.

It sits between people and information, between people and culture, and between people and one another.

That doesn’t make language more important than demographics, economics, geography, or culture. But it may make it harder to separate from them than we sometimes assume.

None of this means abandoning translation. It doesn't mean treating every language as a completely different reality. And it certainly doesn't mean assuming that language alone is what determines how people think.

It simply means recognising that language may be carrying more of the reasoning process than we often give it credit for.

Rather than translating people into a common language before reasoning occurs, Project Brainstorm attempts to let reasoning emerge within the language of the market itself and translate only afterward.

Anthropic's research certainly doesn't prove that this decision was correct. What it does do is suggest that language may matter more than we have traditionally assumed.

For me, that simply reinforces a belief that has shaped Project Brainstorm from the beginning:

Context matters.

And language may be one of its deepest forms.