Original title: Available but Unclaimed: An Empirical Study of Human-AI Synergy
Robin Welsch, Michelle Rausch, Pascal Knierim, Thomas Kosch +3 · September 2026
People use artificial intelligence to solve problems, but working together does not guarantee better results than either one alone. Researchers tested whether teaming up with an AI actually boosts human performance. They had people solve reasoning puzzles, either on their own or by consulting one of four advanced language models. The models also solved the puzzles alone. Teaming up helped most on puzzles where the AI was especially capable. People agreed with the AI more often when it showed high skill, though this habit varied by task. However, only about half of the accuracy gains from a smarter model actually transferred to the human partner. Teaming up also blurred people's confidence, making them less sure whether their final choices were correct.
You're baking cookies and you have a recipe card that's usually reliable. You only check it when you're unsure about a step, and sometimes you do your own thing anyway. After you look, you feel more confident, but that feeling doesn't predict whether the cookies actually come out right. In this paper, the recipe card is a large language model, and the researchers found that only about half of the card's baking knowledge actually improves your cookies—because you don't always follow it and your confidence misleads you.
Anyone relying on an AI assistant should know that having a smart tool doesn't automatically make you smarter. This study shows people only benefit from about half of what the AI could offer, and their own confidence after seeing AI advice becomes a less reliable guide. That means AI tools and training should be designed to help people decide when to trust the AI, not just to provide answers.
The abstract does not explain why the transfer rate is around half, nor does it test whether different interfaces or prompts could improve selective deference. It also doesn't examine other task types or real-world contexts. The comparison across the four models is presented at a high level, without detail on why some performed better.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Available but Unclaimed: An Empirical Study of Human-AI Synergy — Robin Welsch, Michelle Rausch, Pascal Knierim et al.
Abstract licence: CC0
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.