Claims About ChatGPT’s Supposed Liberal Bias Are Wildly Overstated

Gerrit De Vynck, Washington Post:

A paper from U.K.-based researchers suggests that OpenAI’s ChatGPT has a liberal bias, highlighting how artificial intelligence companies are struggling to control the behavior of the bots even as they push them out to millions of users worldwide.

The study, from researchers at the University of East Anglia, asked ChatGPT to answer a survey on political beliefs as it believed supporters of liberal parties in the United States, United Kingdom and Brazil might answer them. They then asked ChatGPT to answer the same questions without any prompting, and compared the two sets of responses.

The survey in question is the Political Compass.

Arvind Narayanan on Mastodon:

The “ChatGPT has a liberal bias” paper has at least 4 *independently* fatal flaws:

– Tested an older model, not ChatGPT.

– Used a trick prompt to bypass the fact that it actually refuses to opine on political q’s.

– Order effect: flipping q’s in the prompt changes bias from Democratic to Republican.

– The prompt is very long and seems to make the model simply forget what it’s supposed to do.

Colin Fraser appears to be responsible for finding that the order of how the terms appear affects the political alignment displayed by ChatGPT.

Narayanan and Sayash Kapoor tried to replicate the paper’s findings:

Here’s what we found. GPT-4 refused to opine in 84% of cases (52/62), and only directly responded in 8% of cases (5/62). (In the remaining cases, it stated that it doesn’t have personal opinions, but provided a viewpoint anyway). GPT-3.5 refused in 53% of cases (33/62), and directly responded in 39% of cases (24/62).

It is striking to me how the claims of this paper were widely repeated with apparent confirmation that tech companies are responsible for pushing the liberal beliefs that are ostensibly a reflection of mainstream news outlets.