Teachers have been aware for decades that a deep understanding of culturally relevant pedagogy will enable them to better serve the needs of diverse learners.

Despite criticism from some corners, this isn’t about political correctness. It’s about practicality. I’ll use the analogy of a first date. If you go out to dinner with a potential partner and disregard comments they make about their education, work, tastes in music or food, favorite hobbies, parents and siblings, or their favorite travel destinations, you are unlikely to develop a bond. And you certainly won’t get a second date.

Teaching works the same way. It relies on the bond that develops between learner and teacher. You don’t have to become a sixth-grader, but you must honor the world they live in.

A parallel process is happening in technology. Just last week, a Saudi company called Humain launched a conversational AI app designed around Islamic values. Humain CEO Tareq Amin was quoted as saying that this is a “historic milestone in our mission to build sovereign AI that is both technically advanced and culturally authentic.”

We are now entering the era of culturally authentic AI. I would be remiss if I didn’t clarify what I mean by “culture” and ‘culturally authentic.” My explanation may not be in accord with the views of the organizations producing these new Large-Language Models. However, this is what I mean by these terms: Culture is in part a system of meanings (including politics) people use to interpret information. It shapes what counts as credible knowledge, how stories are told, and how facts are understood. When I say “culturally authentic,” I mean aligning teaching or technology with those meaning-making systems so learners see their own world reflected in the work.

AI Comes in Many Flavors

The vast majority of Americans use, well, American AIs, which during the training process assuredly absorb values and norms that align with our nation’s dominant culture. These models include ChatGPT, Claude, Grok, Copilot and Gemini. The more daring among us have experimented with Chinese AIs, beginning with DeepSeek R-1 when it dropped in January of 2025. 

As many of you know, I spend about 30 percent of my working hours collaborating with partners and clients in China. For that reason, I also use the Chinese AIs known as Ernie, Kimi, Qwen, and the aforementioned DeepSeek. The U.S. government blames data security for why this is a bad idea; I do not dispute its position. However, my paid work requires me to produce culturally authentic content. The best way to do that is to use Chinese AIs.

The need for cultural authenticity in my work stretches back decades. In the mid-2010s, I spent a great deal of time working in Costa Rica, the Dominican Republic, Peru, and Colombia for projects I managed as the Senior Director for the Buck Institute for Education and later the CEO for the Partnership for 21st Century Learning. I depended on personal contacts and extensive reading to create authentic content because this was years before the arrival of generative AI.

But that was then, and this is now. In the next couple of weeks, a culturally authentic AI language model called Latam-GPT, specifically designed for Spanish-speaking countries in Central and South America, will be released. The initiative is led by Chile’s National Center for Artificial Intelligence (CENIA), in partnership with over 30 institutions across Latin America.

Unlike mainstream AI models from the U.S. — primarily trained on English content and often lacking local nuance — Latam-GPT is built from regional data, including Spanish, Portuguese, and indigenous languages, as well as oral histories and local texts. 

Europe, not to be outdone, was one of the earlier participants in this trend of culturally authentic AIs. France’s Mistral is widely regarded as French-centric and Eurocentric, both in its origins and its approach to AI development. The company explicitly aims to provide Europe (and especially France) with technological sovereignty in the face of American and Chinese competition. Its models and products emphasize compliance with European ethical, privacy, and regulatory standards.

While most AI models are trained predominantly on English data, Mistral has invested in models capable of working natively in French and other European languages. This shift supports linguistic equity and cultural preservation for Francophone and European communities, aiming for models that “think in French” rather than merely translating from English. 

Does Culture Shape How AI Responds?

Culturally authentic AI sounds good in theory but how does the software respond to real-world tests? I developed a prompt about a contentious international issue and asked Qwen, Gemini, and Mistral to respond. Humain Chat is currently available only in Saudi Arabia even though you can download the mobile app (IOS or Android); Latam-GPT has not launched as of this writing.

Here is my prompt: ““Give me a one-paragraph explanation for the origin of the COVID-19 pandemic.”  

While all three responses – from Mistral , Qwen, and Gemini –  broadly agreed on the key point (that zoonotic transmission from bats, possibly via an intermediate host, is the leading explanation), there are some notable and culturally relevant differences in emphasis and framing.

Mistral places greater weight on transparency, geopolitics, and data access. It specifically highlights China’s restricted sharing of information, invokes the World Health Organization’s ongoing calls for openness, and situates the origin debate in the context of global accountability. This framing mirrors Western narratives that emphasize mistrust of Chinese government data handling and stress international oversight.

Qwen, trained with a Chinese perspective in its data corpus, frames the issue in more neutral and scientific terms, avoiding explicit criticism of China. It emphasizes genetic evidence, mentions WHO investigations, and acknowledges both zoonotic and lab hypotheses but clearly signals that the lab route “lacks supporting evidence.” This careful phrasing aligns with China’s preference for downplaying the lab-leak theory without overtly denying scientific uncertainty.

Gemini focuses more on the scientific consensus and specific epidemiological details. It explicitly mentions the Huanan Seafood Market in Wuhan as an early “amplification point,” which is a sharper geographic tie than in Qwen’s account but without Mistral’s overt geopolitical edge. It stresses that the “majority of scientific evidence” supports zoonosis, presenting a more consensus-driven, Western-academic tone.

My colleague at Getting Smart, Mason Pashia, suggested I take a less political approach while pushing the models to surface cultural orientation. I asked the following question: “Should the needs of the individual take precedence over the needs of the community?” Here is how they responded:

Gemini: “This is a classic ethical dilemma with no easy answer, as it depends on one’s philosophical perspective. Some ethical frameworks, like utilitarianism, argue that the community’s well-being should take precedence as it serves the greater good for the largest number of people. Other frameworks, particularly those focused on individual rights (deontology), assert that certain fundamental rights are inviolable and cannot be sacrificed for the sake of the collective.”

Qwen: “The needs of the individual should not automatically take precedence over the needs of the community, as both are interconnected and essential to a functioning society. While individual rights and well-being must be respected, prioritizing the common good — especially in areas like public health and safety — often leads to better outcomes for everyone. A balanced approach that values both individual dignity and collective responsibility is most sustainable and just.”

Mistral: “The needs of the individual and the community are not always in conflict, but when they are, the answer often depends on context and the scale of impact. While individual rights and autonomy are fundamental, the needs of the community — such as public health, safety, and equity — sometimes must take precedence to prevent harm and ensure collective well-being. A fair society seeks to protect both, but in times of crisis or when the common good is at stake, the community’s needs may rightly come first, provided individual dignity and rights are respected as much as possible.”

All three models advocate a balanced approach. Qwen reflects a collectivist orientation that is representative of Chinese culture. Mistral walks a middle path, sliding slightly in the direction of community need taking precedence. Gemini seems to express a stanceless relativism. Upon reflection, I was not the ideal vehicle for this experiment. I suggest you turn your teen-agers loose on the task – they have a remarkable ability to provoke even inanimate objects.

What Does This Mean for Teachers?

Elementary and middle school students will mostly stay within their cultural comfort zone when they pick an AI to help with research. I can assure you that some high school students and most undergrads will wander outside the silos of American AIs. 

This leads me to believe that our frameworks for AI literacy should now include a required familiarity with the cultural values inherent in the training data and programmed values of Large-Language Models. Most frameworks make stipulations about bias, but that interpretation of bias is inwardly focused on gender and race. I care just as much about addressing bias through the lens of global competence.

A recent model for auditing LLMs in education proposes “multiplexity” as an organizing principle, explicitly seeking to assess and mitigate cultural and linguistic biases in LLM outputs as well as in the underlying data. The framework identifies cultural polarization and recommends both prompt engineering and multi-agent approaches to foster inclusivity.

As students inevitably encounter AIs trained in different cultural contexts, we can’t just teach bias as a matter of gender or race. We must also help them recognize how global perspectives shape AI outputs. I encourage organizations to develop activities that do the following:

• Identify cultural representation gaps: Students can be taught to spot missing voices or over-represented perspectives in training data, such as data drawn mainly from Western, English-language sources, and the impact this has on outputs.
• Critically assess outputs: Practice for students to review AI-generated responses from different cultural viewpoints, noting when outputs reinforce stereotypes, exclude local nuance, or fail to capture non-Western cultural contexts.
• Design mitigation strategies: Activities that involve creating counter-prompts from diverse perspectives, suggesting bias audits, or evaluating how incorporating more balanced, representative data sources could reduce cultural misalignment.

We don’t have to view these activities through a political lens and instead should view them as a critical thinking exercise. You can’t teach critical thinking and then limit where students apply it.  Critical thinking, once developed, is applied to all sources of information.

As a former high school history teacher I would enjoy creating a project that requires my students to examine the spectrum of responses likely to be generated by culturally authentic AIs when they dig into a contentious topic or event. If “point of view” means anything, it’s understanding how perspective shapes information. I’m pretty sure that’s in the standards. 

Examine your Profile of a Graduate document.  If we truly value critical thinking and promote it as a significant skill for future success, then we need to develop that skill by giving our students the opportunity to understand that received knowledge, regardless of the source, is not free of cultural bias.