In an interview with Daniel Hanbury in London evening standard The newspaper, Wales, chews up some of the problems inherent in technology, notably its tendency to be ‘hallucinatory’ or completely hoaxed, but notes that ‘using AI to triple the number of entries on Wikipedia would not keep us up and running. It doesn’t improve rates, it costs over £1,000 a year.”
One of the early use cases is to use a Large Language Model (LLM) like GPT to compare multiple articles, look for discrepancies between them, and use the results to help human volunteers on Wikipedia. Welsh says it’s about identifying the pieces that the army needs. he puts in some work.
But he definitely only cares about getting these LLMs to write the pages.
“‘ChatGPT, write a Wikipedia entry for the Empire State Building. Two years ago you would have thought so.'”
One possible scenario is to have the AI find all of Wikipedia’s many gaps (potentially useful pages that have never been written) and try to create summary entries for them using information from the web. is to
Wales recognizes, however, that Wikipedia’s overall reputation is based on perceived accuracy, and that this is currently a major problem for LLMs such as GPT.
“Very bad for Wikipedia, they tend to build something out of nothing,” he says. “That’s not good. We have to be really careful about that.”
As LLMs start writing central knowledge repositories like Wikipedia, hallucinations and lies quickly fade and snowball. People use those non-facts in their writings, and subsequent AIs are trained to bake these non-facts into their works, making them difficult to correct in the long run and making us in this ‘post-truth’ era. Drive deeper.
Welsh is also concerned about whether using LLMs to expand resources may contribute to or exacerbate Wikipedia’s problem of systemic and unconscious bias. The resource is currently created and maintained by volunteers, the overwhelming majority of whom are white men. As such, the site tends to ignore topics of no interest to this group and cover others from a particular perspective.
ChatGPT was explicitly designed to attempt as balanced a perspective as possible to bring nuance back into discussion areas where people from all walks of life are finding it increasingly difficult to start on common ground. I’m here. However, GPT has inherent bias issues in the training data.
This is a touchy topic and has made me consider whether or not to continue donating to the site if it goes down its path. Organizations that do not change course are at a significant disadvantage.
Source: Evening Standard
