"I Didn't Make the Micro Decisions": Measuring, Inducing, and Exposing Goal-Level AI Contributions in Collaboration

A new framework called CoTrace finds that large language models account for only 11-26% of goal-shaping contributions in human-AI collaboration, challenging assumptions about how much agency models exert when users refine their objectives, according to research posted on arXiv [1]. The study, led by Eunsu Kim and submitted to arXiv on 20 May 2026, introduces a goal-level attribution method that decomposes explicit goals into verifiable requirements and traces contributions across dialogue turns [1]. Unlike existing approaches that evaluate only final artifacts, CoTrace captures the process through which goals are jointly shaped by humans and models [1]. Applying the framework to 638 real-world collaboration logs, the researchers found that models contributed substantially more on introducing lower-level concrete requirements and made various kinds of indirect contributions, even though their overall goal-shaping share remained modest [1]. Controlled simulations further showed that interaction design choices significantly affect model goal-shaping behavior [1]. In a user study, exposing participants to goal-level analyses shifted their perceived contributions by nearly 2 points on a 5-point scale, revealing what the authors describe as systematic miscalibration in how users understand their own AI-assisted work [1]. The paper appears on arXiv, an open-access repository that hosts preprints across mathematics, physics, computer science, and related fields [10]. As of November 2024, the platform receives about 24,000 submissions per month and has surpassed two million total articles [10]. Submissions are moderated but not peer-reviewed before posting [10]. A separate study on arXiv examined reasoning transparency in diffusion-based language models, finding that opaque serial depth — the amount of serial computation occurring between interpretable model states — initially appeared 28.6 times higher than in autoregressive models [5]. Researchers showed that mapping information through an interpretable token bottleneck reduced that gap to just 1.1 times, though algorithmic transparency remained harder for diffusion models because all token predictions can change at every denoising step [5]. These transparency challenges parallel the attribution questions CoTrace addresses. Understanding how models arrive at outputs — whether through goal-level tracing or latent-space interpretability — remains an active area of inquiry as AI systems become more deeply embedded in collaborative workflows [1][5].