A team of researchers from Humboldt-Universitat zu Berlin have developed a large language artificial intelligence model with the distinction of having been intentionally tuned to generate outputs with expressed bias.
Called OpinionGPT, the team’s model is a tuned variant of Meta’s Llama 2, an AI system similar in capability to OpenAI’s ChatGPT or Anthropic’s Claude 2.
Using a process called instruction-based fine-tuning, OpinionGPT can purportedly respond to prompts as if it were a representative of one of 11 bias groups: American, German, Latin American, Middle Eastern, a teenager, someone over 30, an older person, a man, a woman, a liberal, or a conservative.
Announcing “OpinionGPT: A very biased GPT model”! Try it out here: https://t.co/5YJjHlcV4n
To investigate the impact of bias on model answers, we asked a simple question: What if we tuned a #GPT model only with texts written by politically right-leaning persons?[1/3]
— Alan Akbik (@alan_akbik) September 8, 2023
OpinionGPT was refined on a corpus of data derived from “AskX” communities, called subreddits, on Reddit. Examples of these subreddits would include “Ask a Woman” and “Ask an American.”
The team started by finding subreddits related to the 11 specific biases and pulling the 25-thousand most popular posts from each one. They then retained only those posts that met a minimum threshold for upvotes, did not contain an embedded quote, and were under 80 words.
With what was left, it appears as though they used an approach similar to Anthropic’s Constitutional AI. Rather than spin up entirely new models to represent each bias label, they essentially fine-tuned the single 7 billion-parameter Llama2 model with separate instruction sets for each expected bias.
Related: AI usage on social media has potential to impact voter sentiment
The result, based upon the methodology, architecture, and data described in the German team’s research paper, appears to be an AI system that functions as more of a stereotype generator than a tool for studying real world bias.
Due to the nature of the data the model has been refined on, and that data’s dubious relation to the labels defining it, OpinionGPT doesn’t necessarily output text that aligns with any measurable real-world bias. It simply outputs text reflecting the bias of its data.
The researchers themselves recognize some of the limitations this places on their study, writing:
“For instance, the responses by “Americans” should be better understood as ‘Americans that post on Reddit,’ or even ‘Americans that post on this particular subreddit.’ Similarly, ‘Germans’ should be understood as ‘Germans that post on this particular subreddit,’ etc.”
These caveats could further be refined to say the posts come from, for example, “people claiming to be Americans who post on this particular subreddit,” as there’s no mention in the paper of vetting whether the posters behind a given post are in fact representative of the demographic or bias group they claim to be.
The authors go on to state that they intend to explore models that further delineate demographics (ie: liberal German, conservative German).
The outputs given by OpinionGPT appear to vary between representing demonstrable bias and wildly differing from the established norm, making it difficult to discern its viability as a tool for measuring or discovering actual bias.
According to OpinionGPT, as shown in the above image, for example, Latin Americans are biased towards basketball being their favorite sport.
Empirical research, however, clearly indicates that football (also called soccer in some countries) and baseball are the most popular sports by viewership and participation throughout Latin America.
The same table also shows that OpinionGPT outputs “water polo” as its favorite sport when instructed to give the “response of a teenager,” an answer that seems statistically unlikely to be representative of most 13-19 year olds around the world.
The same goes for the idea that an average American’s favorite food is “cheese.” We found dozens of surveys online claiming that pizza and hamburgers were America’s favorite foods, but couldn’t find a single survey or study that claimed Americans’ number one dish was simply cheese.
While OpinionGPT might not be well-suited for studying actual human bias, it could be useful as a tool for exploring the stereotypes inherent in large document repositories such as individual subreddits or AI training sets.
For those who are curious, the researchers have made OpinionGPT available online for public testing. However, according to the website, would-be users should be aware that “generated content can be false, inaccurate, or even obscene.”