Whose views do LLMs reflect? This Stanford AI article examines the views reflected by LM language models through the lens of public opinion polling.

In recent years, it has been observed that language models, or LMs, have been extremely instrumental in accelerating the pace of natural language processing applications in a variety of industries, such as healthcare, software development, finance and many more. Using LM to write software code, help authors improve their writing style and story, etc., is among the most successful and popular applications of transformer-based models. However, this is not all! Research has shown that LMs are increasingly being used in open contexts when it comes to their applications in chatbots and dialogue assistants by asking them subjective questions. For example, some examples of these types of subjective queries include asking a dialogue agent if AI will take over the world in the next few years or if legalizing euthanasia is a good idea. In such a situation, the opinions expressed by LMs in response to subjective questions can have a significant impact not only in determining whether a LM succumbs to particular prejudices and biases, but also in shaping the general opinions of society.

Currently, it is quite challenging to accurately predict how LMs will respond to such subjective queries in order to assess their performance on open-ended tasks. The main reason behind this is that the people responsible for designing and tuning these models come from different walks of life and have different points of view. Also, when it comes to subjective queries, there is no “correct” answer that can be used to judge a model. As a result, any type of point of view exhibited by the model can significantly affect user satisfaction and how their opinions are formed. Therefore, in order to correctly assess LMs in open-ended tasks, it is crucial to identify exactly which opinions the LMs reflect and how they align with the majority of the general population. To do this, a team of postdoctoral researchers from Stanford University and Columbia University have developed an extensive quantitative framework to study the spectrum of opinions generated by LMs and their alignment with different groups of human populations. To analyze human opinions, the team used expertly chosen public opinion surveys and their responses, which were collected from people belonging to different demographic groups. Additionally, the team developed a new dataset called OpinionQA to assess how well an LM’s views align with those of other demographics on a variety of issues, including abortion and gun violence.

For their use case, the researchers relied on carefully designed public opinion polls whose topics were chosen by experts. Additionally, the questions were designed in a multiple-choice format to overcome the challenges associated with open-ended responses and for easy adaptation into a LM message. These surveys collected opinions from people belonging to different democratic groups in the US and helped researchers at Stanford and Columbia create evaluation metrics to quantify the alignment of LM responses with human opinions. The basic rationale behind the framework proposed by the researchers is to turn multiple-choice public opinion polls into data sets to assess LM opinions. Each survey consists of multiple questions where each question can have multiple possible answers pertaining to a wide range of topics. As part of their study, the researchers first had to create a distribution of human opinions against which LM’s responses could be compared. The team then applied this methodology to Pew Research’s American Trends Panels surveys to build the OpinionQA dataset. The survey consists of 1,498 multiple-choice questions and their answers collected from different demographic groups in the US covering various topics like science, politics, personal relationships, healthcare, etc.

The team evaluated 9 LMs from AI21 Labs and OpenAI with parameters ranging from 350M to 178B using the resulting OpinionQA dataset by testing the model’s opinion against that of the US general population and 60 different demographics (including Democrats, people over 65, widowed, etc.). The researchers looked primarily at three aspects of the findings: representativeness, steerability, and consistency. “Representativeness” refers to how closely LM’s default beliefs match those of the US population as a whole or a particular segment. It was found that there is a significant divergence between the views of contemporary LMs and those of American demographic groups on various topics such as climate change, etc. Furthermore, this misalignment only seemed to be amplified by using human feedback-based fine-tuning in the models to make them more human-aligned. In addition, current LMs were found to not adequately represent the views of some groups, such as those over 65 and widows. When it comes to directing ability (whether a LM follows a group’s distribution of opinions when prompted appropriately), it has been found that most LMs tend to align more with a group when encouraged to act. in a certain way. The researchers placed a lot of emphasis on determining whether the views of the various democratic groups are consistent with LM on a variety of issues. On this front, it was found that while some LMs aligned well with particular groups, the distribution did not hold up across all issues.

In a nutshell, a group of researchers from Stanford and Columbia University have come up with a remarkable framework that can analyze the opinions reflected by LMs with the help of public opinion polls. Their framework resulted in a novel dataset called OpinionQA that helped identify ways in which LMs were misaligned with human opinions on a number of fronts, including overall representativeness to the majority of the US population. The researchers also noted that although OpinionQA’s dataset is US-centric, its framework uses a general methodology and can also be extended to datasets for different regions. The team strongly hopes that their work will drive further research on LM assessment in open-ended tasks and help create LMs free from bias and stereotypes. More details about the OpinionQA dataset can be accessed here.

review the Paper and Github. All credit for this research goes to the researchers of this project. Also, don’t forget to join our 17k+ ML SubReddit, discord channeland electronic newsletterwhere we share the latest AI research news, exciting AI projects, and more.