The re-writing of history
A fascinating article from wired magazine (Link below) highlights potential weaknesses in the LLM’s data grab.
The scenario, according to the article is that over 80% of all mainstream news content is now blocking access to data collection from Chat GPT and the likes. On the other hand, right wing news such as Fox and Breitbart, are welcoming the ingestion of data.
So if a LLM learns from the data it ingests, then there is legitimate concern that these models will reference the data where there are significantly more sources and volume. So if you have a motivated section of the media and population who are aligned on a certain topic, and who are active in promoting said topic, then could this activity skew the bias of the LLM taking on the data.
If for example, you have enough sources saying Trump won the election, then in response to related search queries would the LLM return a result reinforcing this belief? Taken to another level, where, as is the belief in tech circles, these LLM’s will eventually replace our reliance on Google, then the ability to re-write history is very possible!