Causal Inference with Generative Artificial Intelligence: Application to Texts as Treatments

A new methodology harnesses generative artificial intelligence to strengthen causal inference when the treatment is unstructured text, according to a paper posted to arXiv. The approach, called GenAI-Powered Inference, sidesteps the need to learn a causal representation directly from the data. The framework, detailed in a submission last revised in June 2026, proposes using a deep generative model — specifically a large language model — to generate treatments and then exploit its internal representation for causal effect estimation [1]. Researcher Kentaro Nakamura and collaborators show that knowing this true internal representation helps separate treatment features of interest, such as specific sentiments or topics, from other potentially confounding features [1]. Unlike existing causal representation learning algorithms, GPI does not require the representation to be learned from the observed data, which the authors argue yields more accurate and efficient estimates [1]. Generative AI, the broader technology underpinning the method, refers to computational systems that can produce text, images, audio, and video from prompts [2]. Since the 2020s, such models have become widely available, and their rapid advance has coincided with what some observers call an AI boom [2]. The GPI paper leverages an open-source large language model, Llama 3, to generate text data for both simulation and empirical studies [1]. The authors formally establish conditions for nonparametric identification of the average treatment effect and propose an estimation strategy designed to avoid violating the overlap assumption [1]. They derive asymptotic properties of the estimator through double machine learning and extend the methodology, via an instrumental variables approach, to settings where the treatment feature is based on human perception [1]. The technique is also applicable to text reuse, in which a large language model regenerates existing texts [1]. The work sits at the intersection of causal inference and knowledge representation. In artificial intelligence, knowledge representation and reasoning aims to model information so that a computer system can solve complex tasks [4]. Parameterized models such as neural networks — including the transformer architectures that underpin modern large language models — can be viewed as a family of knowledge representation formalisms [4]. By treating a generative model’s internal representation as a given, GPI avoids the long-standing challenge of learning which formalism is most appropriate for a particular knowledge-based system [4]. The paper was first submitted in October 2024 and has undergone five revisions through mid-2026 [1]. Simulation and empirical results reported in the manuscript illustrate advantages over state-of-the-art causal representation learning algorithms, though independent replication has not yet been documented [1].