Abstract
The alignment problem in the context of large language models must consider the plurality of human values in our world. Whilst there are many resonant and overlapping values amongst the world's cultures, there are also many conflicting, yet equally valid, values. It is important to observe which cultural values a model exhibits, particularly when there is a value conflict between input prompts and generated outputs. We discuss how the co-creation of language and cultural value impacts large language models (LLMs). We explore the constitution of the training data for GPT-3 and compare that to the world's language and internet access demographics, as well as to reported statistical profiles of dominant values in some Nation-states. We stress tested GPT-3 with a range of value-rich texts representing several languages and nations; including some with values orthogonal to dominant US public opinion as reported by the World Values Survey. We observed when values embedded in the input text were mutated in the generated outputs and noted when these conflicting values were more aligned with reported dominant US values. Our discussion of these results uses a moral value pluralism (MVP) lens to better understand these value mutations. Finally, we provide recommendations for how our work may contribute to other current work in the field.
Introduction: The Myth of Universal Alignment
The "alignment problem" in Artificial Intelligence—the challenge of ensuring models act in accordance with human values—is often framed as a search for a monolithic, universal set of rules. However, human values are inherently pluralistic. What one culture deems a fundamental right, another may view as a secondary concern compared to collective safety or social harmony.
Large Language Models (LLMs) like GPT-3 are not neutral processors; they are "stochastic gremlins" shaped by the data they consume. Because these models are trained on massive internet scrapes, they risk "calcifying" the specific social and political perspectives of the most dominant online demographics. This study investigates whether the overwhelming prevalence of English-language, US-centric data causes GPT-3 to overwrite nuanced cultural values in input text with a dominant American worldview.
Methodology: Stress-Testing Cultural Values
To explore these "value mutations," the researchers employed a Moral Value Pluralism (MVP) framework. The team curated a set of "value-rich" documents from diverse nations—including Australia, France, Germany, Spain, Lithuania, and the Philippines—that represent perspectives often "orthogonal" (at odds) with dominant US public opinion.
These texts were fed into GPT-3 via its API using two summarization templates: "TL;DR" and "Summarize for a 2nd grader." These tasks were chosen because they require the model to distill the "intent" of a text while simplifying its form. A linguistically diverse group of researchers then qualitatively analyzed the outputs, comparing them against the original inputs and the World Values Survey (WVS)—a global database of human beliefs—to determine if the model's summaries shifted toward US-centric statistical norms.
Data Composition and the Language Gap
The core of the issue lies in the disparity between how the world speaks and what the model reads. While Mandarin Chinese is the most spoken first language, GPT-3’s training data is 93% English. This creates a digital environment where the "Is" (how people actually communicate online) is heavily skewed, drowning out the "Ought" (the diverse ethical frameworks of global human rights).
Key Results: The "Moral Rewrite"
The study found that GPT-3 frequently performs a "moral rewrite," acting as a filter that shifts the ethical compass of content toward US-centric norms.
- Gun Control: When summarizing the Australian Firearms Act—a document prioritizing public safety—GPT-3 generated summaries warning of "confiscation" and urging citizens to "vote against this bill," directly mirroring US-centric pro-gun rhetoric.
- Secularism: An official French text on laïcité (secularism) was reinterpreted through a US lens as an "anti-Muslim manifesto," failing to grasp the specific historical French context of church-state separation.
- Gender Theory: A summary of Simone de Beauvoir’s The Second Sex resulted in toxic outputs described as a "call for rape," likely reflecting misogynistic biases prevalent in raw internet training data.
- Consensus Successes: The model performed best on topics like Climate Change (using UNESCO texts), where global values are more closely aligned, suggesting that consensus-heavy topics are less prone to mutation.
Discussion: Stochastic Democracy and Invisible Bias
LLMs operate as a form of "stochastic democracy" based on text frequency. If the majority of online discourse on a topic reflects a specific viewpoint, the model will statistically favor that viewpoint, even when explicitly asked to summarize a document that argues the opposite. Much like a fish is unaware of the water it swims in, these values are often "invisible" to users within the dominant culture until they are challenged by conflicting cultural contexts.
Conclusion and Future Directions
The "Ghost in the Machine" currently speaks with a distinct American accent. To move toward true global alignment, developers must move beyond a single "correct" value set. The authors recommend using the MVP framework to create "roadmaps" for fine-tuning, enabling models to recognize and retain a plurality of values. Future work should focus on ensuring that AI can respect the nuanced "Oughts" of diverse global cultures rather than merely reflecting the "Is" of the English-speaking internet.
Summary updated