The Talking to Machines project is exploring how to leverage AI to enhance large scale randomized control trials. The Ghana Wave II trial, which is on-going, is an opportunity to explore whether the decisions of synthetic subjects, who have been assigned to specific treatment arms, are informative. In this essay we have described the implementation of a synthetic trial that parallels the real human trial this is currently on-going. Once the human trial is completed and the data analyzed, we are proposing to benchmark the treatment effects from synthetic subjects that we have presented in this essay, with those generated from the trial using human subjects.

The Talking to Machines initiative aims to revist how we conduct large scale experimental research. We believe that many of the key elements of experimental design and implementation can be enhanced by AI. Measuring the attitudes and behaviors of participants in these studies is a typical outcome variable in randomized control trials. An on-going research project explores how AI can radically change public-opinion research by providing researchers with a general methodology to make representative inference from cheap, high-frequency, highly unrepresentative samples. In this paper we focus specifically on samples which are readily available in moderate sizes. To this end, we provide two major contributions: 1) we introduce a general sample-selection process which we name online selection, and show it is a special-case of selection on the dependent variable. We improve MrP for severely biased samples by introducing a bias-correction term in the style of King and Zeng to the logistic-regression framework. We show this bias-corrected model outperforms traditional MrP under online selection, and achieves performance similar to random-sampling in a vast array of scenarios; 2) we present a protocol to use Large Language Models (LLMs) to extract structured, survey-like data from social-media. We provide a prompt-style that can be easily adapted to a variety of survey designs. We show that LLMs agree with human raters with respect to the demographic, socio-economic and political characteristics of these online users. The end-to-end implementation takes unrepresentative, unsrtuctured social media data as inputs, and produces timely high-quality area-level estimates as outputs. This is Artificially Intelligent Opinion Polling. We show that our AI polling estimates of the 2020 election are highly accurate, on-par with estimates produced by state-level polling aggregators such as FiveThirtyEight, or from MrP models fit to extremely expensive high-quality samples.