INTERNATIONAL RESEARCH / PEER-REVIEWED ARTICLE

Does AI think differently depending on the language you use?

GPT and ERNIE showed different cultural tendencies when responding to the same measures in English and Chinese. We examine the meaning of language choice for localization, advertising ideation and international brand communication.

ACADEMIC WORK REVIEWED

Cultural tendencies in generative AI

ResearchersJackson G. Lu, Lesley Luyang Song, and Lu Doris Zhang.

Nature Human Behavior · 20 June 2025 · 9:2360–2369 · Peer-reviewed article · Source language: English

Open the original publication

ENGLISH REVIEW AND COMMENTARY: DIJITALPI

We explain the paper in plain English, discuss what it may mean for organisations and identify our own examples separately.

Read the explanation Go to 3 business examples

English review: 9 September 2026 · The source date appears in the citation above.

Published review

Source and editorial checks are complete. DijitalPi did not conduct a customer experiment for this paper.

LET'S READ THE RESEARCH TOGETHER

We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.

01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?

Could the same AI model show different cultural trends only when the language used changes?

Lu, Song and Zhang examined whether generative AI produces different tendencies in social relationships and thinking style when used in English and Chinese. They also investigated whether these differences appeared in advertising recommendations and changed with cultural role instructions.

Explaining the context · DijitalPi commentary

Language carries more than words. It can shape whether an idea emphasizes individual success or group harmony, and whether attention falls on a single object or on context and relationships. Because model training data contains traces of human cultures, answers to the same question in different languages may differ by more than literal translation.

The researchers do not say that models have culture like humans. What they measure are output trends. This distinction allows us to read the results without attributing human personality to the model.

02 / HOW WAS THE RESEARCH CONDUCTED?

How did they test the question?

The researchers used gpt-4-1106-preview and ERNIE-3.5-8K-0205 through APIs. They compared English and Chinese versions of established psychology tasks that measure social orientation and cognitive style, using Likert scales, short scenarios, visual tasks and free-text analysis. To reduce the risk that models had memorized published questions, they changed names and contexts, used some unpublished measures and applied translation and back-translation.

Understanding the method · DijitalPi commentary

Independent social orientation emphasizes personal goals and characteristics, while interdependent orientation gives more weight to relationships and the group. Analytical thinking tends to break an object into parts, while holistic thinking places more emphasis on context and relationships. The study treats these as different tendencies rather than good and bad options.

The advertising part of the study is not the result of a real campaign. The model was asked to recommend between two advertising approaches with independent or interdependent social emphasis. Thus, it was examined whether the trend seen in the language was reflected in practical advice.

03 / RESEARCH FINDINGS

Chinese responses showed more relational and context-sensitive tendencies.

Both GPT and ERNIE showed more interdependent social orientation and more holistic cognitive style when using Chinese than English. GPT chose the group- and relationship-emphasized approach in its ad recommendations more frequently in the Chinese condition. In exploratory analyses, when the model was given the role of an average person living in China in English, some responses approached the trends in the Chinese condition.

Source: Nature Human Behaviour: publication summary, findings and author credits ↗

04 / DIJITALPI'S EXPLANATION

How should we interpret these findings?

This finding helps us question the “write in English first, then translate” habit in international content production. When the language changes, not only the words may change, but also which benefits are considered important and in what context they are explained.

Role instruction can be a useful start; but in one sentence, cultural accuracy is not guaranteed. The instruction to “think like someone living in Turkey” may not represent the diversity of real Turkish users and may reinforce stereotypes. Local human evaluation is therefore necessary.

The study examined two specific model versions in English and Chinese. We cannot assume the same result for Turkish. Trends may change with model versions, system instructions and training data, so localization tests should record and repeat the date and version used.

05 / CONCLUSION AND OPEN QUESTIONS

What did we learn, and what do we still not know?

The paper shows that generative AI outputs depend in part on the language used and are not culturally neutral. Both models repeated the broad differences between English and Chinese; the advertising-recommendation and cultural-role findings are exploratory. The researchers did not measure actual campaign performance or responses from Turkish users.

How are the advertisements produced directly in Turkish and the advertisement translated from English evaluated by users in Turkey? Answering this requires measuring local evaluation and behavior on the same product and offer.

Return to the researchers' original publication

Nature Human Behaviour: publication summary, findings and author credits DOI and current publication record

DIJITALPI'S APPLICATION COMMENTARY

What could this look like in your organisation?

We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.

EXAMPLE 01

Translation correct, emphasis foreign

The English slogan was perfectly translated into Turkish. The text still feels distant to the brand's local customer.

Is the problem only in word choice?

Open the recommendation for example 01

Also examine the value framework of the message.

Mark which framework the text highlights, such as individual gain, family, community, security, or innovation. Clearly define the target audience context in the brief before asking the local team for an alternative.

EXAMPLE 02

Same model, two different suggestions

The team gets different advertising ideas from English and Turkish prompts for the same product.

Can you delete one because it's a model error?

Open the recommendation for example 02

Test the source of the difference in a controlled manner.

Keep the product, audience and format constant, changing only the language. Present the outputs to independent evaluators. Score cultural fit and performance separately, and do not treat the study's Chinese-English comparison as evidence for Turkish.

EXAMPLE 03

Add culture role

The sentence "Think like a customer living in Turkey" was added to the prompt and the result looked more local.

Does this single sentence guarantee cultural accuracy?

Open the recommendation for example 03

Consider the role instruction the initial hypothesis.

In the research, it was shown exploratively that cultural prompts can adjust some tendencies. Don't mistake stereotypes for accuracy without checking with real local users. Flag false generalizations in editorial review.

TRY IT WITH YOUR TEAM

Compare a campaign through two production paths.

Prepare a draft directly in Turkish and a draft translated from English for the same brief. Have the local team record in writing which sentence feels more natural or foreign and why.

Check out our domestic and international marketing approach

Source and ownership

Jackson G. Lu, Lesley Luyang Song, and Lu Doris Zhang.
MIT Sloan School of Management; Tongji University; Tsinghua University.

Nature Human Behavior · 20 June 2025 · 9:2360–2369 · Peer-reviewed article. Open the original publication

The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.

← Return to all research articles