Diyi Yang and collaborators propose the Offloading Score to quantify human overreliance on AI tools

We introduce Offloading Score. The idea is simple: we examine the AI-assisted steps of the user’s workflow, estimate how the user would have completed it without the tool, then compare how many steps from the counterfactual workflow were ‘saved’ by using the tool. (4/9)

Paper: https://arxiv.org/abs/2605.29392 Code: https://github.com/vishakhpk/offloading-score Some examples and concise findings: https://vishakhpk.github.io/measuring-reliance/

With my amazing co-authors! @lujainmibrahim @ZhiruoW @jennjwang @QVeraLiao @Diyi_Yang

More below!🧵 (2/9)

High reliance is not always undesirable. We examine the interaction between reliance and a desirable task outcome, code understanding. While in-general high reliance leads to low code understanding, we also find a cluster of high-reliance + high-understanding users that are often _learning_ with AI and augmenting their own skills. This suggests that reliance should be interpreted alongside task outcomes, not in isolation. (8/9)

From a very anecdotal point of view, I can confirm that in some ways, my thinking process has changed or has become externally lazier, such as writing my own content and even spelling words correctly. I'm 100% working more than ever, but my brain is working differently than it was.

Offloading Score reliably recovers this relationship, assigning significantly higher reliance to users under time pressure (short vs long below!). Other common measures, including AI interactions, lines of code, and self-reported measures, do not reliably distinguish between the two conditions. (6/9)

To evaluate if Offloading-Score captures reliance effectively, we need a setting where reliance can be predictably manipulated. For this, we ran a controlled study with developers using AI coding tools under different levels of time pressure. Prior work suggests that people rely more heavily on decision aids under time pressure (5/9)

Two users might spend the same amount of time interacting with an AI assistant and generate the same amount of code with its help, yet one may be delegating substantially more of the task. Measures based on AI usage alone are no longer sufficient for contemporary AI-assisted workflows. (3/9)

Beyond the scalar value of Offloading Score, we also obtain richer behavioral insights from the workflows, finding that under time pressure, users shift toward execution-oriented interactions, delegate more subtasks to the AI, and directly reuse AI-generated outputs more often. (7/9)

Offloading Score is calculated directly from screen and keyboard data, making it reusable across _any_ AI tool and interface. By making reliance measurable, we hope to enable more systematic study of overreliance, deskilling, and loss of agency in human–AI collaboration. (9/9)

@andrewdobrow @ZhiruoW @Diyi_Yang @QVeraLiao @jennjwang This resonates a lot with our motivation! I think it's inevitable that our behavior patterns change as we engage more with AI tools. Our goal with Offloading Score is to provide a way to analyze those changes and make more informed decisions about how we work.

@vishakh_pk @ZhiruoW @Diyi_Yang @QVeraLiao @jennjwang I recently released an #npm #package for #token compression and prompt optimisation. The package works with #HuggingFace models to estimate token probabilities. The goal is to reduce #tokenAPI costs. I'd really appreciate🌟on GitHub. https://github.com/SaBA26-void/project_void