There have been various papers and articles recently that discuss the ‘personality’ of GPT-3 and seek to identify its biases and perspectives. What I’m writing about today is the opposite of that. Rather than probe GPT-3 for its identity, some researchers are what is possible when GPT-3 is prompted to assume a specific identity and respond as a proxy for that identity, with all its biases
It turns out that it can simulate another human’s perspective with sufficient fidelity that it can act as a proxy for a diverse population of humans in social experiments and surveys. This is accomplished by inventing a population of participants described by a backstory or a set of attributes including things like gender, race, income, occupation, age, political identity, etc. For each virtual person a prompt (or prompt chain) is created that establishes the identity and asks GPT-3 to effectively act as that person in a response to a question or scenario.
One example is the study “Using Large Language Models to Simulate Multiple Humans” in which four different social experiments are recreated with virtual subjects modeled to mirror the real study’s participants. The experiments were the Ultimatum Game, garden path sentences, risk aversion, and the Milgram Shock experiments. The ultimatum game is a game theory scenario as follows: There is a sum of money and two subjects, let’s say Bob and Carol. Bob must offer Carol a portion of the money. If Carol accepts, then she gets that amount of money and Bob gets the rest. If instead the offer is rejected, neither get any money.
The controversial Yale University Milgram Shock experiment explored whether people would obey an authority figure if they were instructed to perform an act that conflicted with their personal conscience. Specifically, the study tested whether people would administer electric shocks to another person when ordered to do so by an authority figure. The results of the study showed that people were more likely to obey authority figures, even when doing so caused harm to others. The study has been criticized for its ethical implications, but it remains an important contribution to our understanding of human behavior.
The researchers accounted for the likelihood of GPT-3 having been trained on data describing these experiments by concocting new versions that capture the idea but not the specifics of the original experiment.
They found strong fidelity in the results that mimicked the real world counterparts.
“Out of One, Many: Using Language Models to Simulate Human Samples” is another example, where they create “silicon samples” by conditioning the model on thousands of socio-demographic backstories from real human participants in actual surveys conducted in the U.S. The survey questions were political in nature, asking questions that gauge attitudes about political parties, political language, social groups, etc.
They found high algorithmic fidelity with the human results. and found that GPT-3’s information goes far beyond surface similarity. Their words: “It is nuanced, multifaceted, and reflects the complex interplay between ideas, attitudes, and socio-cultural context that characterize human attitudes. We suggest that language models with sufficient algorithmic fidelity thus constitute a novel and powerful tool to advance understanding of humans and society across a variety of disciplines.“
Why does this matter and why should we care?
Well, if the exploratory, proof of concept research bears out the idea of being able to test hypotheses about human behavior with virtual participants instead of real ones, the cost of this kind of research and the speed at which results can be generated should drop dramatically.
For such a virtual study, there is no need to identify and recruit real participants, so that cost evaporates. For virtual studies or surveys, the results are effectively instantaneous once the population and questions have been defined, dramatically accelerating the process of validating hypotheses. This opens the door to refinement and iteration that would not have been practical before. It opens the door to research and studies that wouldn’t have been possible or practical, for a number of reasons including ethics and cost. These virtual studies could also be used to test a hypothesis to determine whether or not a real-world study is justified, and if so, what is the most effective way to do it. In other words it doesn’t have to take the place of real-world studies to be useful, rather it can increase the effectiveness and value (the bang for the buck) of real world studies.
My first, possibly naive impression thinking about this is that it should be a net positive for society. If the cost of research drops, in theory, the cost of achieving whatever goal the research services should also drop. If, for example, we’re talking about product research, could it speed up the process of product development and result in products that are better matched to need, reducing waste in the economy? Sounds like a win-win, right?
Useful technology is, however, often a double-edged sword, and others have raised the prospect of much darker usages. “Out of One, Many: Using Language Models to Simulate Human Samples” offers this warning:
“We note, however, that while this work lays exciting groundwork for the beneficial use of these models in social science, these these tools also have dangerous potential. Models with such fidelity, coupled with other computational and methodological advances, could be used to target human groups for misinformation, manipulation, fraud”.
Import AI blog author Jack Clark puts it this way…
“Because models like GPT3 can, at a high level, simulate how different human populations respond to certain things, we can imagine people using these to simulate large-scale information war and influence operations, before carrying them out on the internet.“
I can envision scaling this up to an influence operation using GPT-3 to model how to interact with individuals in a manipulative way, paired with bots that can identify and track targets. By that I mean literally interact differently with millions of people, each in a unique way with the goal of influencing their beliefs and manipulating their actions. An army of bot ‘friends’ that are don’t have your best interest at heart. This is, in a word, worrying.
I would like to think that like any technological disruption, there will be positive and negative effects, and they’ll balance out, but that is decidedly NOT what we’ve seen with social media and how content is optimized and fine-tuned to drive outrage-motivated engagement. I would like to be optimistic about this but I am not.
Ghost in the Garden 