Questions for James Evans on his 10/3 talk on "Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions.""

[jamesallenevans]

Pose your questions as Issue Comments (below) for James Evans regarding his 10/3 talk on Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions. Large Language Models (LLMs), through their exposure to massive collections of online text, learn to reproduce the perspectives and linguistic styles of diverse social and cultural groups. This capability suggests a powerful social scientific application – the simulation of empirically realistic, culturally situated human subjects. Synthesizing recent research in artificial intelligence and computational social science, we outline a methodological foundation for simulating human subjects and their social interactions. We then identify nine characteristics of current models that are likely to impair realistic simulation human subjects, including atemporality, social acceptability bias, uniformity, and poverty of sensory experience. For each of these areas, we discuss promising approaches for overcoming their associated shortcomings. Given the rate of change of these models, we advocate for an ongoing methodological program on the simulation of human subjects that keeps pace with rapid technical progress.

Contributing papers
Required: Kozlowski, Austin, James Evans. 2024. "Simulating Subjects: The Promise and Peril of AI Stand-ins for Social Agents and Interactions."
Plus ONE of the following:

• Potter, Yujin, Shiyang Lai, Junsol Kim, James Evans, Dawn Song. 2024. "Hidden Persuaders: How LLM Political Bias Could Sway Our Elections" Empirical Methods in Natural Language Processing (EMNLP).
• Lai, Shiyang, Yujin Potter, Junsol Kim, Richard Zhuang, Dawn Song, James Evans. 2024. “Evolving AI Collectives to Enhance Human Diversity and Enable Self-Regulation.” International Conference on Machine Learning (ICML).
• Kozlowski, Austin, Hyunku Kwon, James Evans. 2024. "In Silico Sociology: Forecasting COVID-19 Polarization with Large Language Models."

[saniazeb8]

Hi,

Thank you for sharing ground breaking research and new ideas.

Considering the potential for LLMs to influence political opinions, what regulatory frameworks could be established to ensure transparency and fairness in their deployment during election cycles?

[yunfeiavawang]

Thank you for sharing this incredible article! It's exciting to explore the potential of AI stand-ins in social research. I have a question: How can researchers or practitioners ensure that AI stand-ins accurately simulate the intended personas or groups? Is it primarily a reflection of the data used to train the model (LLM), combined with the researcher's assumptions and techniques like RAG, prompting, or fine-tuning? What indicators or criteria can be used to confidently assess whether the AI stand-ins are effectively meeting the research goals?

[dario-marino]

Don't you think that as always most problems that are described here are problems which already existed and the AI is making them visible?
Don't you think that already opinions of journalists are over represented in the news and that journalist talk more with other journalist like every other social bubble?
Don't you think that there is already a bias in every organization that could also use an AI?

In general what I want to ask is how many of these problems described in this paper were already present and AI won't make the problem worse, it could just substitute the people who were already digesting internet information without an original twist in it.

I am also interested in the example of simulated society with multiple agents. To be interpretable it has to be coded in a way that would make it an inefficient version of an agent based model. It would not be interpretable if it would be a more qualitative simulation of society and in that case it would become an inefficient representation of society.

[tyeddie]

As LLMs are trained on “near complete archives of all text on the internet,” how can researchers mitigate inherent biases or avoid overcompensating for these biases when using simulated subjects to approximate “ground truth”?

What methods should be used to assess the validity and representativeness of such training data online in looking out for objective social facts?

[gtriberti9]

There is a broad range of characteristics within each culture, they might overlap and, most of the time, certain factors are shared within many people. Additionally, we have people with ideologies within the 'gray area'. Fine-tunning is certainly helping to define many features, but how can we represent these profiles if they are somehow ambiguous?

[ubaidjamal0]

The reading, in my opinion, raised some interesting epistemological questions. One such question is how do we classify 'something' as knowledge, or to go even further, how do we rank systems of knowledges? Historically, indigenous systems of knowledge have often been sidelined by European ways of thinking, due to the history of modernity and how intrinsically linked it was to colonialism. A consequence of this has been the underrepresentation of indigenous knowledge, and an overrepresentation of European modes of thinking. For example, whereas Amerindian thought perceives humankind as an extension of nature, this way of thinking has mostly been replaced by the Kantian idea that privileges humankind over all other species in our contemporary ethos. This underrepresentation and misrepresentation persists on the internet, and even more so when it comes to other modes of data, especially films. My question then is: going forward, how do we ensure that LLM models, rather than simply mimicking the dominant narrative, actually serve to synthesize different, especially historically underrepresented/misrepresented, systems of knowledge? This, in my opinion, would be a great challenge as most of the data available on indigenous forms of thinking has either not been documented or has been misrepresented from a European perspective. This question extends to the idea of simulation of human personas and human interaction as well, because if the goal of these models is just to represent what already is the dominant narrative, then that, perhaps implicitly, is an admission of the fact that the world as it is is a good world.

Just to clarify, I use the word indigenous loosely here to encapsulate all those knowledge systems that have historically been misrepresented or underrepresented.

[gabereichman]

One potential issue with prompting LLMs is their sensitivity to subtle differences in prompts, such as switching a word in the prompt with a synonym, that human participants may see as essentially identical. What can be done to address this prompt-engineering problem in research? In addressing this issue, is it more important to develop an ideal prompt for validity of the experiment, or is it preferable to investigate the range of responses that that can arise from the LLMs with these subtle changes?

[ecg1331]

Thank you for sharing your research!

I have never heard of the suggestion that LLMs could eventually replace human subjects. Do you foresee these (future) studies using LLMs only when a specific subgroup of a study is unable to be reached, or do you see them being used as a complete replacement for human subjects?

Further, if human subjects were completely replaced, how would the study be classified? Would it be more like an analysis of recent representations in media, rather than an actual study of human thought/behavior?

I also thought your bit about the moral standing of AI participants was interesting. Do you have any suggestions for how social scientists should begin to 'seriously consider' this?

[alejandro-sarria-morales]

Is a reduction of bias desirable or even possible? If LLMs reflect their training data, their biases are a reflection of bias in this data. As such, wouldn't an elimination of bias imply tampering with the training data or with the way LLMs interpret it. Why would neutrality be a higher priority than fidelity towards the training data?

[Mengyang-Cao]

How can researchers address the risk of underrepresentation or misrepresentation of social groups with limited online presence when using LLMs trained predominantly on internet content, especially when simulating the perspectives of politically disengaged or marginalized groups who may not actively contribute to online discourse?

[AnnieZzz01]

Thank you for sharing fantastic information. I didn't know the LLMs model until I read these articles. In your articles, you mentioned some limitations of the current LLMs model such as lack of sense of time, social acceptability bias, and lack of sensory experience. To overcome these problems, you suggested that it can be improved by fine-tuning the model or introducing multimodal data, etc. So which approach do you think is most likely to make a breakthrough in practical applications in the near future? Are there already preliminary experimental successes that can demonstrate the effectiveness of these programs?

[AnqiW02]

Thanks for sharing these fantastic articles! I felt I've learned more about how LLMs work and relate to computational social science. Since, in the article, it's mentioned that the model’s internal representations are more closely linked to English than other languages and that models may, in some ways, “think in English,” I'm wondering if it's caused by the types of languages themselves, that is, tone language (e.g. Mandarin, Thai) and intonational language (English, Spanish) or if it's caused by how languages are formed in different cultural backgrounds?

[nitomanto]

Thank you sharing your work on LLMs.

You mentioned that these LLMs all display bias in one form or another. The example in your paper you gave was that ChatGPT had a more liberal slant in American politics than the general American public as a whole. You then mentioned briefly that fine-tuning for certain traits such as politeness may have inadvertantly caused this bias. Is this bias more an unintended consequence of fine-tuning, or perhaps an intentionally fine-tuned parameter for the purposes of curating an outward representation of the company? Are there any other examples of biases that may have resulted from fine-tuning for other traits?

[jstanleyi]

Thanks for sharing such strong articles. I am thinking, while self-regulation in AI collectives can reduce the burden of centralized control and allow for more flexible, diverse interactions among agents, it also presents challenges. Autonomous collectives can foster creativity and prosocial norms, but they may be vulnerable to risks such as the emergence of undesirable behaviors, biases, or even the creation of harmful echo chambers within the system​. Therefore, my question is, as a social scientists, how can we minimize the misunderstanding between the machine and us, while we formulate the machine-based languages?

[jstanleyi]

Thank you sharing your work on LLMs.

You mentioned that these LLMs all display bias in one form or another. The example in your paper you gave was that ChatGPT had a more liberal slant in American politics than the general American public as a whole. You then mentioned briefly that fine-tuning for certain traits such as politeness may have inadvertantly caused this bias. Is this bias more an unintended consequence of fine-tuning, or perhaps an intentionally fine-tuned parameter for the purposes of curating an outward representation of the company? Are there any other examples of biases that may have resulted from fine-tuning for other traits?

Hi, at first I have the same question! But now I may think, fine-tuning involves trade-offs, and addressing one form of bias may inadvertently introduce others. Fine-tuning for traits such as politeness or neutrality may inadvertently introduce biases by aligning the model with social norms that correlate with certain political ideologies, like progressivism. This is likely unintended, as traits like politeness may be more associated with liberal values of inclusivity. However, some bias might also be intentional, as companies may fine-tune models to reflect their outward image, ensuring the model avoids harmful or polarizing speech that might conflict with corporate values like inclusivity and anti-discrimination. Beyond political bias, fine-tuning for other traits can introduce biases as well. For instance, fine-tuning for empathy might reinforce gender stereotypes, while fine-tuning for cultural sensitivity might marginalize certain cultural perspective