Artificial intelligence should be trained on culturally different data sets to avoid bias

Large Language Models (LLMs) are deep learning artificial intelligence programs, resembling OpenAI’s ChatGPT. The skills of LLMs have developed into quite a broad range write fluent essaysto coding to creative writing. Millions of individuals worldwide use LLMsand it could be no exaggeration to say that these technologies are transforming work, education and society.

LLMs are trained by reading large amounts of text and learning to acknowledge and mimic patterns in the info. This allows them to create coherent and human-like texts on virtually any topic.

Because the Internet continues to be predominantly English – As of January 2023, 59 percent of all web sites were in English — LLM students are trained totally on English texts. Additionally, the overwhelming majority of English text online comes from users based within the United States, home of 300 million English speakers.

LLMs learn concerning the world through English texts written by US-based web users Standard American English and have a narrow Western, North American and even US-centric lens.

Model bias

In 2023, when ChatGPT learned of a pair dining at a restaurant in Madrid who tipped 4 percent, suggested they were thrifty, on a decent budget, or didn’t just like the service. By default, ChatGPT followed the North American standard of a 15 to 25 percent tip. Disregarding the Spanish norm of not tipping.

Since early 2024, ChatGPT has appropriately cited cultural differences when assessing the appropriateness of a tip. It is unclear whether this ability arose from training a more recent version of the model on more data – in any case, there are a lot of typing guides in English on the Internet – or whether OpenAI has patched this particular behavior.

However, there are other examples that reveal ChatGPT’s implicit cultural assumptions. For example, with a story about guests showing up for dinner at 8:30 p.m., it suggested Reasons why guests were late, although the time of the invitation was not mentioned. Again, ChatGPT probably assumed they were invited to a typical North American dinner at 6 p.m.

In May 2023, researchers on the University of Copenhagen has quantified this effect by prompting LLMs with the Hofstede cultural survey, which measures human values ​​in numerous countries. Shortly afterwards, researchers exhibited AI start-up company Anthropic used that World Values ​​Survey do the identical. Both works concluded that LLMs exhibit a robust fit with American culture.

The same phenomenon occurs when asking questions FROM-E 3, a picture generation model trained on pairs of images and their captions to generate a picture of a breakfast. This model, trained totally on images from Western countries, produced images of pancakes, bacon, and eggs.

Effects of bias

Culture plays a crucial role in shaping our communication styles and worldviews. As well as Intercultural human interactions can result in misunderstandingsUsers from different cultures who interact with conversational AI tools may feel misunderstood and find them less useful.

To be higher understood by AI tools, users can adapt their communication styles in an analogous approach to how people have learned to “Americanize” their foreign accents to achieve success personal assistants like Siri and Alexa.

As more people depend on LLMs to edit texts, they’re prone to standardize the way in which we write. Over time, LLMs risk eliminating cultural differences.

Decision making and AI

AI is already getting used because the backbone of varied applications that make decisions that impact people’s lives, resembling: Continue filtering, Rental applications And Applications for social advantages.

for years, AI researchers have warned that these models not only learn “good” statistical relationships – resembling considering experience as a desirable trait in a job applicant – but in addition “bad” statistical relationships, resembling consideration Women are considered less qualified for technical positions.

As LLMs are increasingly used to automate such processes, one can imagine that the North American bias learned from these models may result in discrimination against people from different cultures. An absence of cultural awareness can result in AI perpetuating stereotypes and increasing societal inequalities.

LLMs for languages ​​apart from English

Developing LLMs for languages ​​apart from English is difficult vital effort, and there are a lot of such models. However, there are several the explanation why this ought to be done in parallel with improving the cultural awareness and sensitivity of LLMs.

First, there’s a big population of English speakers outside of North America who usually are not represented by English LLMs. The same argument applies to other languages. A French language model could be more representative of the culture in France than of the culture in other Francophone regions.

Training LLMs for regional dialects – which of them can capture subtler cultural differences – can be not a practical solution. The quality of LLMs will depend on the quantity of knowledge available and due to this fact their quality could be poorer for dialects with little online data.

Secondly, many users whose native language just isn’t English still select to make use of English LLMs. Major breakthroughs in voice technology are inclined to occur Start with English before applying them to other languages. Even then, there usually are not enough online texts in lots of languages ​​– resembling Welsh, Swahili and Bengali – to coach high-quality models.

Due to the shortage of availability of LLMs of their native language or the higher quality of English LLMs, users from different countries and backgrounds may prefer using English LLMs.

Ways forward

Our research group on the University of British Columbia works to complement LLMs with culturally diverse knowledge. Together with a doctoral student Mehar BhatiaWe trains an AI model on one Collection of facts about traditions and ideas in numerous cultures.

Before reading these facts, the AI ​​suggested that an individual eating a Dutch Baby (a sort of German pancake) was “disgusting and mean” and would feel guilty. After training, the person felt “full and satisfied,” it was said.

We are currently collecting a large-scale caption dataset with images from 60 cultures, which is able to help models study breakfast types apart from bacon and eggs, for instance. Our future research will transcend providing models of the existence of culturally diverse concepts to higher understand how people interpret the world through the lens of their cultures.

As AI tools develop into more ubiquitous in society, it’s imperative that they move beyond the dominant Western and North American perspectives. Companies and organizations across many industries are using AI to automate manual processes and use data to make higher, evidence-based decisions. Making such tools more inclusive is critical for Canada’s diverse population.