DistantNews
Support us
🇦🇷 Argentina /Technology

Who Teaches Artificial Intelligence How to Think?

From Clarín · () Spanish

Translated from Spanish and summarized by DistantNews. Read the original for the full story.

At a glance

Analysis Documents & data Context piece
  • The article argues that AI models absorb cultural and political biases from the texts, images and videos used to train them.
  • Research cited in the article found that commercial models gave more favorable answers about Beijing when political questions were asked in Chinese, while additional training on Chinese state media shifted an open model’s responses in nearly 80% of cases.
  • It warns that chatbots can hide the origins of their information, making users less able to recognize and critically assess outside influences.

The most dangerous AI bias may not be an obvious mistake. It is the plausible answer that users accept without noticing where it came from.

The article recalls how researchers in Beijing once tested ChatGPT with questions about China. The model described basketball star Yao Ming as the first Chinese woman to play professional basketball in the United States. The error was so blatant that people could easily identify and correct it. More difficult to detect are answers that sound reasonable while carrying an unseen cultural or political bias.

AI models train on vast quantities of text, images and video, none of which is culturally neutral. For years, that material was disproportionately Western and English-language. Beijing has recognized that training data represents a contest for influence and wants China to become a global provider of AI data. It has already released large collections, some designed to reflect what the article calls China’s dominant values.

Whoever supplies the material influences what models learn to treat as normal, relevant or true. A study published in Nature in May found that commercial models from OpenAI and Anthropic responded more favorably to Beijing on political questions when researchers asked them in Chinese. When researchers gave an open model an additional 6,400 documents from Chinese state media, it produced more favorable answers about the Chinese government than its original version in nearly 80% of cases.

The author says cultures have always absorbed outside influences, and that exchange can be productive. But people need to recognize those influences as something received. A film has an origin, a book has an author and a search engine displays links. A chatbot speaks in a uniform, confident voice and does not show where its answers come from. When the messenger disappears, part of the critical filter also weakens. AI answers may have no accent, the article concludes, but they still come from somewhere.

Once that material enters the training data, the model washes it until it looks like neutral information.

· Study authorThe quote describes how political material can lose its visible origins after entering an AI model’s training data.
About this summary

Originally published by Clarín in Spanish. Translated, summarized, and contextualized automatically by DistantNews, with a note on how the source frames the story. Not individually reviewed before publishing. How this works.

More Stories

From Our Blog