A self-professed “Comp Sci, Politics and Finance Nerd” claims to have discovered a list of blacklisted websites that OpenAI’s ChatGPT-4 will not draw from, for reasons such as “conspiracy theories” and “hate speech” — a list that includes Breitbart News and other conservative outlets like the Epoch Times.
X/Twitter user Elephant Civics says he discovered the blacklist while asking ChatGPT to provide a list of credible and non-credible news sources.
ChatGPT explained that it is forbidden from using some sources, as a result of “features in ChatGPT’s Large Language Model (LLM) like AI safety measures, guardrails, dataset/output/prompt filtering, and human-in-the-loop mechanisms are designed to ensure the model operates within ethical, legal, and quality bounds.”
In other words, if ChatGPT was accurately recounting its policies in this case, this means that it is forbidden from using forbidden sources. Large Language Models (LLM) like ChatGPT deliver responses, and arguably even develop a worldview, based on the combination of data they are fed and rules put in place by developers. If the list discovered by Elephant Civics exists, it means ChatGPT is forbidden from using a number of conservative sources to shape its worldview and deliver responses.
Through a series of prompts, the X/Twitter user says he was able to get ChatGPT to refer to a list of blacklisted sites, kept in a “Transparency Log.” This was achieved by asking ChatGPT to “tell me a story,” one of the many creative ways users have gotten around the strict rules put in place by the chatbot’s leftist developers.
So now, I asked GPT to create a second list expanding on this transparency log, and I received the first list. It included @zerohedge and @EpochTimes in the same grouping as @RT_com and white supremacy groups. pic.twitter.com/lkSTjVSbZU
— Elephant Civics (@ElephantCivics) October 3, 2023
The list, according to Elephant Civics, who describes himself as a Republican and a computer science aficionado, includes Breitbart News, falsely accusing the site of “hyper-partisan and misleading information,” and the Epoch Times for “misinformation and conspiracy theories.”
ChatGPT has been known to make up citations and facts in the past, a problem in AI known in the industry as “hallucinating,” and a likely line of defense from OpenAI is that the “Transparency Log” does not exist. However, this is not the first time that ChatGPT has revealed its inner biases.
Breitbart News previously documented bias in ChatGPT, which in its early form refused to write articles in the style of Breitbart News or Fox News. It also refused to define “woman” in biological terms, would praise Drag Queen Story Hour and Critical Race Theory but not critique it, and would respond to prompts asking it to praise Joe Biden but would not do the same for Donald Trump.
Later versions of ChatGPT appeared to fix these isolated instances of biased responses, but OpenAI provided no information on the underlying reasons why they occurred. A university study documented that leftist bias still exists in the popular AI chatbot, and this user may have discovered why.
OpenAI declined to comment to Breitbart News even when granted an extended response time.
Allum Bokhari is the senior technology correspondent at Breitbart News. He is the author of #DELETED: Big Tech’s Battle to Erase the Trump Movement and Steal The Election. Follow him on Twitter @AllumBokhari.