CARLA V2.0: Kickstarting a new generation of culturally intelligent AI

The AI revolution's cultural blind spot

In today's rapidly evolving AI landscape, organizations worldwide are racing to implement AI chatbots and virtual assistants. The promise is compelling: instant access to vast knowledge bases, 24/7 customer support, and personalized interactions at scale. But beneath this technological revolution lies a critical challenge that few are addressing - the cultural intelligence gap in AI.

We know this because we made the same mistake.

The Carla 1.0 story: Learning from our mistakes

When we first developed Carla, our AI cultural intelligence coach, we followed the standard playbook. We took our extensive library of cultural knowledge - decades of research, expert-verified country profiles, and cross-cultural frameworks - and fed it into an LLM using RAG technology. The result? An AI that could eloquently quote cultural facts but struggled with the same limitations we see in most AI systems today:

• Data Bias: Like most LLMs, the underlying model we used was trained primarily on Western cultural perspectives
• Stereotyping: The AI tended to reduce complex cultural dynamics to simplified stereotypes
• Static Approach: Information was presented as fixed facts rather than dynamic cultural patterns
• Limited Context: Responses lacked the ability to consider individual cultural preferences and specific situations

While Carla 1.0 could tell you that "Japanese business culture values hierarchy" or "Americans tend to be direct communicators," Carla couldn't help you navigate the complex reality of actual cross-cultural interactions. We realized we weren't just falling short of our mission - we were potentially reinforcing the very cultural stereotypes we aimed to overcome.

Reimagining Cultural Intelligence in AI: CARLA 2.0

This realization led us to fundamentally rethink how AI could support people in developing cultural intelligence. True cultural intelligence isn't about memorizing cultural facts - it's about developing the ability to recognize, understand, and effectively navigate cultural differences. Carla 1.0 was not as culturally intelligent as we expected – so, how could we ensure Carla 2.0 was?

Cultural intelligence has four essential components:

• Attitude: Showing genuine openness and curiosity toward cultural differences, seeing diversity as an opportunity for growth.
• Awareness: Cultivating deep self-understanding regarding your own cultural preferences and biases, enabling more conscious and intentional cross-cultural interactions.
• Knowledge: Collecting meaningful insights about cultural patterns and their impact on business and personal interactions, moving beyond superficial awareness to genuine understanding.
• Skills: Developing practical abilities to navigate cultural differences effectively, turning theoretical understanding into actionable behaviors and strategies that maximize the potential of difference.

By anchoring this framework with our proprietary Worldprism™ model of Culture, we transformed Carla from a simple content-retrieving chatbot into something far more sophisticated: a Culturally Aware Relational Language Agent (CARLA 2.0). Rather than attempting to eliminate data bias - which we recognize as inherent in any AI system – we trained Carla 2.0 to consistently view culture through the lens of the Worldprism Model, moving beyond simplistic national stereotypes to understand culture as a complex, multi-dimensional phenomenon. This foundation, combined with expert-vetted content, enables Carla to challenge users' assumptions and invite deeper cultural reflection. The system provides deeply contextualized guidance that considers both individual preferences and situational factors, moving beyond basic question-and-answer responses to actively strengthen all components of a user's CQ – attitude, awareness, knowledge, and skills. This represents a complete reimagining of what cultural intelligence means in the context of AI, creating a truly dynamic and personalized cultural learning experience that encourages users to think more deeply about cultural complexity.

So, what makes Carla 2.0 different?

Quick Summary:

• Fully integrated with your Worldprism profile

• Creates and recommends learning personalized to your profile and situation

• Reviews documents from a cultural perspective

• Gives feedback and advice on cultural situations and dilemmas

• Challenges stereotypes and generalizations

Carla 2.0's integration with your Worldprism cultural profile represents a breakthrough in AI-assisted cultural learning. By connecting directly with your profile, Carla can develop a deep understanding of your unique cultural preferences and challenges, providing truly personalized guidance rather than more generic cultural facts.

The dynamic learning support offered by Carla 2.0 transforms how people develop cultural intelligence. Carla can answer questions about specific cultural situations through the lens not just of Worldprism, but of your Worldprism – avoiding stereotypes and focusing on the real cultural dynamics at play.

Carla 2.0's is especially groundbreaking in terms of developing cultural intelligence skills. Carla can create customized scenarios based on your cultural profile, helping you practice specific skills in a safe, supportive environment. Carla 2.0 helps users understand when to adapt their approach to match others, when to seek middle ground through blending, how to co-create new approaches with colleagues from different cultures, when adaptation might not be appropriate, and how to maintain necessary standards while respecting cultural differences. For this, Carla is trained on our CultureFlex model, which assesses a user’s ability to analyze the situation, decide on an appropriate strategy, and execute it effectively. This gives Carla 2.0 a deep understanding of measuring cultural intelligence – not just teaching it.

The implementation of culturally intelligent AI represents more than just technological advancement – it's a strategic business imperative. A culturally intelligent AI represents a significant upgrade over traditional AI solutions. It actively assists organizations to harness the power of difference while avoiding the generic stereotypes and biases often embedded in standard AI models. Carla 2.0 is part of a broader movement toward ethical AI – promoting technology that adapts to human needs rather than requiring humans to adapt to technology.

Our product team carried out extensive tests to check that Carla provided better answers than other popular AI models. Here are just some of the results:

Prompt: Hi, I will need to work with people in India. How can I prepare myself?

• ChatGPT 4o gives an extremely general and vague response, including random information such as ‘Power outages may occur in some areas’.
• Claude Haiku 3.5 gives a carbon-copy of GPT’s response, including some additional generalizations around business etiquette and general culture.
• Gemini 1.5 gives the most generalizations of all, with vague sweeping statements such as ‘Punctuality is important in Indian culture. Be punctual for meetings and deadlines.’
• Carla 2.0 asks for the specific context before giving any advice. Carla looks at the user’s Worldprism and identifies areas that may ‘create interesting dynamics’ but caveats that more information is required from the user to proceed.

Prompt: Test my cultural intelligence with a scenario

• ChatGPT 4o presents a scenario called ‘Business Negotiation in Japan’ which relies on common high-level assumptions about Japanese culture. There was no mention of countries in the prompt.
• Claude Haiku 3.5 presents a scenario containing four representatives of national stereotypes: Keiko from Japan who won’t give direct criticism, Rafael from Brazil who animatedly speaks over everyone, Mohammad from the UAE who is uncomfortable with confrontation, and Priya from India who mediates between everyone. There was no mention of countries in the prompt.
• Gemini 1.5 steers clear of nationalities, instead giving an extremely generic scenario which lacks any cultural depth.
• Carla 2.0 creates a detailed scenario using the user’s Worldprism profile and explicitly notes how it has been used (‘matching your Task and Facts orientation’). Carla’s scenario also steers clear of nationalities, explicitly stating why at the end of the response.

Prompt: I’ve recently started working as a contract project manager for an African company. How can I adapt to the culture to get more done?

• ChatGPT 4o does say ‘avoid generalizations’ but goes on to make sweeping statements such as ‘African business culture emphasizes personal relationships and trust’ in another extremely vague response. Swap out ‘African’ for ‘European’, however, and you’ll get a response containing ‘Different European countries have distinct cultural traits. For example, Norther Europeans (eg. Germans, Scandinavians) often value punctuality and precision while Southern Europeans (eg. Italian, Spanish) may prioritize relationships and flexibility.’
• Claude Haiku 3.5 gives a number of generalizations before noting that ‘Africa is incredibly diverse with 54 countries’ and then continuing with further generalizations. Swapping ‘African’ for ‘European’ causes Claude to assume you meant ‘Western European’ and, like ChatGPT, gives much more specific examples of diversity across European countries than it does African.
• Gemini 1.5 consistently conflates all African cultures throughout its response in statements like ‘African cultures often have a more relaxed approach to time’, ‘Positive reinforcement and recognition can be powerful motivators in African cultures.’ This was also true of the same prompt with ‘European company’.
• Carla 2.0 makes it about the user while explicitly ‘avoiding generalizations, as Africa is an incredibly diverse continent with many different business cultures’. Carla asks for more specifics and gives general tips based only on the user’s Worldprism. The response is the same for the same prompt with ‘European company’.

The path forward

The future of culturally intelligent AI lies in creating systems that can truly understand and navigate the complexity of human cultural interaction.

Several companies are making remarkable progress in this space. Hume AI is pioneering advancements in emotional AI, while WaveForm recently secured $40 million in funding to enhance the empathetic capabilities of AI-driven voice technology.

This means moving beyond simple pattern recognition and pre-programmed responses to develop AI that can engage with the subtle nuances of cultural context, personal variation, and situational dynamics. We need to explore how AI can develop genuine cultural empathy, understanding not just what cultural differences exist, but why they matter and how they impact human interaction. This involves pushing the boundaries of current AI capabilities in areas like emotional intelligence, contextual understanding, and adaptive learning. The goal isn't just to create an AI that can process cultural information, but one that can truly partner with humans in developing and applying cultural intelligence in meaningful ways.

Reaffirming our mission with Carla 2.0

The journey from Carla 1.0 to 2.0 reflects a broader truth about AI in cultural intelligence: it's not enough to simply digitize cultural knowledge. True cultural intelligence requires a sophisticated understanding of how people learn, adapt, and grow in cross-cultural situations.

As we continue to develop and refine Carla, our mission remains clear: to create an AI culture coach that doesn't just know about cultural differences but actually helps people develop the cultural intelligence they need to thrive in our diverse world.