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."" #1
Comments
|
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? |
|
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? |
|
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? 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. |
|
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? |
|
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? |
|
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. |
|
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? |
|
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? |
|
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? |
|
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? |
|
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? |
|
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? |
|
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? |
|
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? |
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 |
|
Thanks for sharing these great ideas on using LLMs to simulate human behaviors and social interactions. I have some questions on the comparisons and possible cooperations of LLMs (or generative AIs) simulating and agent-based modelling (ABM). As I read these papers, these kinds of simulating, especially simulating interactions, seem similar with some long-standing ABM research (in which different agents act in a defined environment). Yet, agents in ABM are much simpler than generative AIs in their responds towards both prompts and other agents. In this sense, could we say LLM simulating could be a better way or alternative way for ABM? Moreover, some researchers raised an idea called "generative agent-based models (GABMs)" (Ghaffarzadegan et al. 2024). Authors of the paper utilized generative AI as agents to simulate game theory under complex situation, and output the logic of agents on these decision making. Yet, just as Austin and Evans (2024) mentioned uniformity as one weakness of LLMs, I may think this GABM approach would ignore the variation between agents in real life and assumed a uniform human society. Do you think this kind of GABM or generative agent approaches could be a better way to overcome some of the challenges in ABM? If yes, to what extent could it do so? |
|
Thank you for sharing! The paper explores how large language models (LLMs) can act as digital stand-ins for human subjects and discusses their limitations, including biases and the lack of sensory experience. I have a question about one specific limitation: the integration of multiple linguistic cultures within LLMs. The paper mentions the uncertainty in how multilingual models handle diverse languages and their impact on simulating different cultural backgrounds. Could you expand on the implications of this issue and suggest any strategies to mitigate it? |
From my point of view, there is still room to get better simulations for the intended personas or groups. Yes, there are methods such as Fine-Tunning. However, as you certainly denote, some characteristics will be a consequence of how we create those simulations. I would say that, inevitably, a percentage of how well the model simulates will be our perception. The latter also relates to the shifting we have seen in the use of interpretivism (this is if we believe that the best way to study social order is through the subjective interpretation of the participants involved). |
Hello! This is such an interesting question. And I think this bias could be seen as an unintended consequence of fine-tuning rather than a deliberate decision. Fine-tuning a model like ChatGPT often aims to enhance user interaction, such as making responses more polite, safe, or relevant. However, these adjustments might inadvertently align the model's responses with certain cultural or societal norms that reflect these traits more prominently, leading to a perceived bias. One example of biases I could think of that might result from fine-tuning is sociocultural bias, since fine-tuning with data from predominantly English-speaking regions could lead the model to display a bias towards Western cultural norms and values. |
|
Thank you for sharing these illuminating articles. After reading the article about forecasting COVID19 polarization, I'm curious about whether LLMs can reveal the role of elite discourse in influencing the formation of public opinion or not when modeling the political attitudes of social groups. Specifically, can LLMs help reveal the extent to which public attitudes are guided by political elites rather than stemming from individuals' pre-existing beliefs and values? If so, how does it reveal this guidance? |
|
Thank you so much professor for sharing the latest developments in your research. My question would be, since you mention the challenge of overcoming biases, uniformity, and the atemporality of large language models in simulating human subjects, what are the most promising innovations or research directions you foresee for addressing these limitations, particularly in terms of incorporating temporality and improving the diversity of simulated responses? |
|
Thank you for such interesting research. I'm curious about the ethical implications of making AI models more and more powerful. First of all, I'm not an expert on the subject but I've read that AI leaves an important energy footprint that has become alarming for the environment. I'm not sure about the details but how do you propose we can arrive to a middle ground where there's scientific advancement without being too detrimental to the planet? On the other hand, I've also known some companies have had issues with their AIs as they become very powerful and humans lose control over them. I know this may sound a bit like a sci-fi movie, but do you have any concern on what's the limit of what AI can do without becoming a threat for humans? |
I would argue that it is completely intentional that popular Large Language Models like GPT or Bard are trained to be polite and therefore have a left-leaning bias. Additionally, because so many things on the Internet are from a Western context it should not be surprising then that any LLM trained on vast amounts of internet data has an evidently Western bias. I agree, however, that we need to constantly be aware of such biases and not take for granted that LLMs even as they become more and more advanced, still rely on data. Data that itself is inherently biased and not always truthful. I agree with Alejandro's sentiment that questioned the ability for LLMs to ever be unbiased and whether that was even a step we would want to take. I think a very powerful and promising step forward is demonstrated in the "Simulating Subjects" paper where Kozlowski and Evans discuss the inherent biases in questions asked in English and Chinese (p. 24). I would argue these biases make LLMs more interesting. As long as we remain cognizant of these biases, these differences have ingrained and interesting cultural contexts that remain interesting. |
I think you bring a good point. I do believe that companies behind AI models intentionally curate a representation they want to give to the world. It's on their best interest to be politically correct and avoid being subject to criticism due to ideological reasons. |
I think this is a very interesting question as it forces us to think about the ethical dimensions of trying to reduce bias in LLMs. Given the importance of outwardly appearances in the current age, I would hypothesize that companies do intentionally try to make their models more polite in order to appear more palatable to the general audience. After all, reducing bias or, at least attempting to do so, allows companies to keep themselves safe from alienating a segment of their intended consumer-base. An interesting question that arises from this is that: isn't the pursuit of neutrality when fine-tuning a model a form of bias in itself? I say so because LLMs, ultimately, are dependent on data, so fine-tuning a model to make it unbiased or more 'polite' might include the process of omitting certain kinds of data, specifically data that might be perceived as offensive by at least some segment of the consumer-base. |
|
Hello! Reading the Simulating Subjects paper, I considered my own research topic, violent political extremism, and thought about whether LLMs simulating extremist actors would be useful. Obviously one problem is the rarity of the data-- events themselves are not super common, and the largest dataset I know of is a 50,000-line leak of one group's internal chats-- but even if an LLM could simulate the actions of a virtual domestic terrorist, etc., would we be learning about the nature of extremists (being a small and unpredictable population) or are we rather learning about the nature of LLMs? The greater question is just a generalized version: are we learning about the nature of people or just the nature of LLMs? |
|
Given that multimodal AI models may require vast and diverse sensory data to improve their understanding of human interactions, how can researchers ensure the ethical collection and use of such data while preserving individual privacy, and what frameworks should be established to regulate the boundaries of AI’s sensory capabilities in real-world applications? |
|
Thank you for sharing your study. You mention the challenges of atemporality and social acceptability bias in LLMs. How do you propose addressing the need for real-time awareness in AI simulations while managing the biases these models inherit from static, historical data? Also, how can future advancements, like incorporating real-time updates or sensory data, help improve the realism and diversity of simulated human interactions? |
|
Thank you for sharing your work. What are the potential challenges and limitations that researchers face when training AI models on multimodal information, in terms of issues like data alignment across different modalities (e.g., text, images, and audio), balancing the quality and quantity of data from each modality, and ensuring proper cross-modal learning and synchronization, especially for temporal data, all while maintaining scalability and interpretability in increasingly complex model architectures? |
|
Thank you for sharing these papers! The first one provides a comprehensive review of the social science research with LLMs simulations. I'm very interested in simulations, and I also implement agent-based modeling, NK fitness landscape, and as well other simulation models in my research projects. I feel like LLM simulations and agent-based modeling are like the two sides of a spectrum. ABM emphasizes simplicity and usually applies to macro behaviors, eg. racial segregation, while LLM simulation usually focuses on reality (the closer to the actual world the better) and the common use case is for relatively small-scale scenarios. My question is whether there is a way to combine these two simulation methods and whether they complement each other. |
Thank you for sharing your questions! Some strategies to address these constraints and mitigate ethical risks might include transparent disclosure with a hybrid approaches. With clearly communicate the use and limitations of AI models in research and policymaking and combine AI simulations with real human input to balance efficiency with authenticity. While AI models can offer valuable insights, they should complement rather than replace human judgement in critical decision-making processes. |
|
Thanks for the presentation! In your talk on simulating human subjects with LLMs, you outlined nine characteristics of current models that impair realistic simulations, including atemporality and social acceptability bias. Given the rapid pace of AI advancements, which of these characteristics do you think will be the hardest to overcome, and what emerging methodologies or innovations do you see as most promising for addressing this challenge in the near term? |
I would be very wary of the validity of a study which entirely replaces human subjects. If the data existed to perform the study on human subjects, the study would have been conducted on human subjects. If the data doesn't exist, we would need to trust the the LLM to give accurate responses, instead of just giving responses that sound like they could be accurate. If such a study would be conducted I should think it would be a meta-analysis of the training data of the LLM. |
|
A primary issue I see with artificial intelligence is the bias that can be present in language models, especially considering that, in an environment where AI agents interact with one another, they tend to prefer to interact with AIs that are similar to themselves. This is particularly alarming for the advancement of knowledge surrounding the sphere of AI research, because biases against certain perspectives may become the standard in the language model's preferences. I find the discussion surrounding how AIs interact with one another interesting, particularly in the case of political ideology. Running simulations by placing ideological constraints on certain agents, and how they associate with like-minded agents, is fascinating. I would worry about how this may be used in the context of social media and the future capability of misinformation bots. How can bots evolve with these complexities in mind? Is this something we should worry about, especially since social media is now considered a primary source of information? |
|
We have learned the language bias of AI from the divorce question, and I am wondering is this bias coming from the social/cultural background of people using this language or what |
This is a particularly interesting series of questions. While researchers are training AI agents to learn from one another, they are also making sure to train AI agents to associate with one another. This highlights the real-world example you provided of journalists; we are training like-minded AI to associate with one another, similar to how humans associate with those who have like-minded interests. However, researchers have also trained AI to identify "bad acting" AI's in the process. This could mitigate the potential issue of biases against AI agents who were not trained to share similar perspectives. Thus, we could simulate how real associations develop, while mitigating the hostility that also occurs in the real world. |
|
I was interested by the methods in these papers, and this (along with a few other questions from students) made me wonder how evaluation of methods for papers using LLMs could be standardized. gabereichman's question about prompting led me to ask further questions -- how will research using LLMs standardize a method for picking and evaluating hyperparameters that highly affect results (e.g. temperature, length, knowledge cutoff, specific model -- mini, 4o, turbo, etc. etc.)? Further -- how will evaluations be carried out, when the thing to measure is often very qualitative (rather than quantitative)? |
|
Thanks for sharing! I find this paper among the readings especially intriguing:
|
|
Thanks for sharing! I'm wondering how might the ability to generate and study "in-between" identities using agents steering vectors challenge our understanding of social categorization and reveal hidden dimensions of human belief systems? |
I'd say the "alien intelligence" of AI models will likely be the hardest characteristic to overcome. While we can improve temporal awareness or reduce biases, I still believethe fundamental way LLMs process information remains profoundly different from human cognition. Promising approaches include multimodal training with rich sensory data and developing more sophisticated "theory of mind" capabilities. |
In my experience, I see that companies release versions of GPT, and you can specify which model to use. For example, OpenAI's model site (https://platform.openai.com/docs/models) lists different models (4, 4o, turbo, etc.) and each model has different versions based on updates. |
|
Thanks for your really insightful and illuminating paper! Your paper seems to imply that the LLMs could be misused or even intentionally used to produced fake knowledge or even harmful content. This new type of LLM-generated knowledge could be a huge source of misinformation, giving place for power or resource appropriation. And because of the public's increasing trust in the LLMs, such an impact seems only to inevitably expand. In general, what would be your advice in terms of using LLMs to produce knowledge? |
|
It's interesting to see how LLM may generate different answers to the same questions, e.g. yes or no on the divorce. I think the social acceptability bias is like a double-edged sword where on one hand the model might be considered more trustworthy to integrate the learning of cultural norms to predict the wartime strategy between countries that may have different levels of moral acceptance and cultural mindsets; on the other hand, the data built on various social norms may deepen the intergroup cleavage and knowledge polarization in general. However, I think it's safe to say that developing the capability of getting rid of the social acceptability bias for LLM is important so that the latter would be able to use its discretion to filter it out or take it into account depending on the varying circumstances. My question is why nowadays, the boundary/barrier between the data clustered around different languages is still very impermeable given the strength of translation technology, i.e., google translate? |
|
As the student with political science background, I felt impressively triggered by your paper and the one on political bias. It jumps out of the previous research on influence of social media or hate speech on election, which majorly took a more theoretically way. When presenting the impact of LLMs, their influence on election brings out a brand new channel for us to consider. It is worth noticing how open- and closed-source respond differently to this. The question would be "How do instruction-tuned large language models increase their political bias compared to their base models, and what are the observed effects on human voting preferences after interacting with these models?" |
After what I read, I guess it can be integrated through the macro and micro Integration. The ABM is typically used for modeling macro-level, emergent behaviors of large systems, while LLMs are more focused on detailed language-based interactions that reflect real-world complexity at a micro level. |
Hi Yuhan, here are my thoughts on this. It's true that biases in large language models (LLMs) have undermined the realism of simulations in social science research; several approaches have arisen to solve this. One strategy is using tools like OpinionQA, which measure how well LLMs align with public opinion across demographic groups. This helps identify where models might be skewed and adjust them accordingly. Fine-tuning LLMs on diverse datasets, particularly those reflecting a range of political views, can also help mitigate bias. Additionally, controlling LLM randomness by adjusting temperature settings reduces the likelihood of generating biased responses. Researchers can further simulate diverse personas within the models to ensure balanced outputs across ideological perspectives. Finally, incorporating human feedback throughout development can help refine LLM outputs, though care must be taken to avoid introducing additional biases from annotators. "Open" AI, Google, Amazon, and others have certainly encouraged a plethora of research to be conducted to solve these problems as their failures to temper their fine-tuning properly have led to such polarization problems.
|
|
Hi Prof. James, Thank you for sharing your exciting work with us! I have a couple of general questions, In the paper, it is mentioned that AI could be used as a stand-in for hard-to-reach human subjects, such as political or economic elites. I wonder what are the implications of using AI as a stand-in and how might this influence the conclusions drawn from such studies? Moreover, could this be possibly extended to other hard-to-reach individuals or vulnerable (due to challenges in accessibility, communication, or for ethical considerations)? As far as my understanding of AI goes: the concern is always about the training data, and so, hard to reach individuals may stay so even with AI. Or can AI's learning ability coupled with training on previous research help with that? Another question, what's your advice for how researchers can ensure that these simulations capture the dynamism and variability of real-world social behavior, given that AI models are prone to atemporality and social acceptability bias? The example with COVID is really interesting, but I wonder if this applies to other examples too. |
|
Hi Prof. Evans, thanks for your sharing! My question is: Considering that AI simulations can reproduce cultural and social biases, how might these biases affect economic predictions, particularly in policy-making or financial markets? What measures could be taken to minimize the impact of these biases on economic outcomes? |
|
Thank you for sharing! How do you envision the potential for AI-based simulated agents to influence or alter social norms within their interactions, especially when these agents are deployed in diverse cultural settings? |
|
Thank you for sharing! What ethical frameworks do you think should guide the deployment of AI agents as stand-ins for human social interactions, particularly in scenarios where they might influence public opinion or behavior? |
|
Thanks for your sharing! I noticed there are many discussions on the positivity or social desirability bias when simulating subjects with LLMs (and with research using Big Data in general). What are the discussions on negativity, which is the cultural ideal attitude/emotion for some scenarios (for example, on anonymous BBS platforms)? Do we approach it differently from how we approach positivity bias? In addition, given the popularity of Generative AI usage, would you say our views, responses, and thoughts are inevitably influenced by responses we obtain from AI? Are we living in a symbiosis where knowledge is circulating recursively between us and AI tools? What are the social and ethical impacts? |
|
Hi James, Thank you for sharing. My quesstion is that: We also see some recent development that higher parameters LLM do not yield better perforamcne! I think it could be very interesting to investigate, also the patterns in the presentation. |
|
The AI's language bias revealed in the divorce question suggests a reflection of societal or cultural biases present in the data it was trained on. Is this bias primarily a product of the users' language, or are there other contributing factors? |
|
Thanks for sharing. How do you envision the integration of AI-simulated subjects changing the landscape of social science research in the next 5-10 years? |
|
Thanks for sharing. How can LLMs fully represent human subjects if online content is restricted or censored( especially related to politics, conflicts or other controversial topics)? Also, will simulating subjects and society based on online content only reflect patterns in which people behave and interact with each other in virtual world? And how to overcome this bias to simulate real human subjects and society? |
|
How do large language models handle cultural differences when simulating human subjects? |
|
Are there potential ethical issues in using large language models to simulate human subjects, particularly when simulating sensitive social situations or representing specific groups? How can we ensure that these simulations do not perpetuate misunderstandings or biases against certain populations? |
|
Evans and Kozlowski identify limitations of current LLMs, such as atemporality and social acceptability bias.How might researchers balance leveraging their current strengths while addressing these limitations to ensure more accurate simulations of human subjects? |
|
The concept of using AI to simulate human subjects raises interesting questions about research ethics and informed consent. In traditional human subjects research, participants actively consent to participate and can withdraw. How should we think about consent and ethical guidelines when working with AI simulations of human populations, particularly when modeling vulnerable or marginalized groups? |
|
The inherent bias of LLMs generated by training data and fine tuning process can influence their function in different kinds of social science research pipelines. What specific strategies could be employed to deal with these biases? How can we assess the influence of these biases when applying LLMs in our studies? |
and others
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:
The text was updated successfully, but these errors were encountered: