Original title: Antagonistic Predictions
Christoph Kern · September 2026
Most artificial intelligence models try to predict the most likely outcome based on past patterns. In serious situations, this can cause big problems. Past data often contains historical unfairness. When computers try to guess the most common pattern, they copy those old biases. It also causes trouble for human decision-makers, who might either blindly trust the tool or completely reject it. The author suggests a new approach called Antagonistic Predictions. Instead of guessing what usually happens, the model acts like a sparring partner. It actively challenges the human user's assumptions and shows alternative futures that are still reachable. This approach aims to fight against social stereotypes, past disadvantages, and mental blind spots in everyday machine learning and generative artificial intelligence.
Imagine you are trying to decide which college to apply to. Your friend looks at past data and says, 'Based on your grades, the most likely outcome is that you get into State University.' That is one prediction. Now imagine a different friend who says, 'But if you took a harder math class and aced it, you could get into City Tech, which is a much better fit for your career goals.' The first friend gives you the average path; the second shows you a reachable future you had not considered. In this paper, the standard model is the first friend, and the Antagonistic Prediction is the second—it challenges your idea of what is possible rather than reinforcing the most common pattern.
Today's AI systems learn from past data and predict the most likely outcome, which can lock in existing biases and bad decisions. For caseworkers in child welfare, criminal justice, or hiring, this means machines may reinforce harm instead of helping. Antagonistic Predictions flips this: models would show what could be different, forcing humans to question their own assumptions rather than just trust the machine.
This is a perspective paper, so it proposes a concept without empirical results or concrete implementations. The author does not demonstrate that Antagonistic Predictions works in practice, nor does it specify how to train such models or measure success. Challenges like ensuring the alternative futures are realistic and avoiding manipulation are not addressed.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Antagonistic Predictions — Christoph Kern — SocArXiv — doi:10.31235/osf.io/nbmq8_v1
Abstract licence: CC BY
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.