The promise of instant translation, powered by artificial intelligence, has hit a snag. These powerful systems, often called “AI,” are prone to inventing information – a phenomenon known as “hallucination.” It’s a frustrating reality for anyone who’s relied on them, and the consequences are now rippling through one of the world’s most trusted sources of knowledge: Wikipedia.
The story began with good intentions. A non-profit organization sought to expand Wikipedia’s reach by funding translators to convert articles into different languages. The goal was noble – to democratize access to information – but the method proved flawed. The project leaned heavily on large language models like Google Gemini and ChatGPT to handle the bulk of the translation work, with limited human oversight.
Editors quickly discovered a disturbing trend. Translated articles contained factual errors that didn’t exist in the original text. Citations were missing, swapped around, or pointed to irrelevant sources. The AI wasn’t just struggling with nuance; it was fabricating details, a particularly dangerous trait when dealing with established knowledge.
The situation is complicated by the economics of Wikipedia editing. Most contributors are volunteers, driven by a passion for knowledge. While some edits come from individuals with professional interests, those contributions are carefully scrutinized and disclosed. Introducing a paid element, even for a worthwhile cause, appears to be creating pressure to prioritize speed over accuracy.
Translators are compensated roughly $400 a month for full-time work, a sum that carries more weight in regions of the Global South where many of them reside. However, the financial incentive, coupled with the volume of work, may be contributing to the errors. The need for translation is urgent, as English Wikipedia dwarfs other language versions in size.
Despite the issues, Wikipedia is continuing to utilize the translation services. The demand for information in less-represented languages is too great to ignore. However, translators affiliated with the organization now face significantly stricter review processes. A mere five documented errors can lead to a ban and the potential removal of their previous work, unless a senior editor intervenes.
This situation highlights a critical challenge in the age of AI: the tension between efficiency and accuracy. While these tools offer incredible potential, they are not infallible. The Wikipedia experiment serves as a stark reminder that human oversight remains essential, especially when safeguarding the integrity of information.
