Advanced AI systems, often marketed simply as artificial intelligence, frequently invent inaccurate details. This issue becomes evident quickly during use and extends beyond simple facts. Wikipedia is currently addressing one instance, discovering fabricated content in articles translated by large language models.
The initiative began with good intentions: a nonprofit group known as the Open Knowledge Association (OKA) offers stipends to individuals converting Wikipedia content into various languages. Difficulties arise, however, from depending on tools such as Google Gemini and ChatGPT for these translations, bypassing necessary human oversight.
As detailed in a 404 Media investigation, volunteer Wikipedia reviewers identified factual inaccuracies not present in the source versions, along with absent, incorrect, or irrelevant references—like pages from books unrelated to the topic. (Such models often falter in handling sources reliably, a pitfall noted in legal contexts.) Earlier, the process involved Grok, the AI from Elon Musk integrated with X (formerly Twitter) and notorious for generating unauthorized explicit imagery, though OKA has since updated its approach.
Most Wikipedia contributors work without compensation. While professionals like public relations specialists can make changes as part of their roles, these must be declared and face extra examination. Introducing paid incentives for bulk translations, even for expansion purposes, could encourage haste among OKA-funded translators, leading to oversights. The organization compensates full-time staff at around $400 monthly, a modest sum that holds greater value in developing regions where translators are based.
Although Wikipedia generally prohibits content created solely by large language models, its team has chosen to continue partnering with OKA. The English edition remains far larger than others (excluding Cebuano, which relies heavily on automation), and underrepresented languages suffer from limited coverage, making reliable translations for key topics essential. OKA participants face stricter controls than typical editors: following five verified mistakes, a translator risks permanent exclusion, with their prior work potentially deleted unless a higher-level reviewer claims responsibility.
With 15 years in tech reporting, Michael has covered topics from Apple innovations to ZTE devices. At PCWorld, he specializes in keyboards, testing fresh models and assembling custom mechanical setups or upgrading his home workstation during free time. His work has appeared in Android Police, Digital Trends, Wired, Lifehacker, and How-To Geek, including on-site coverage of events like CES and Mobile World Congress. Based in Pennsylvania, he eagerly anticipates his upcoming kayaking adventures.