Stanford researchers' Offloading Score measures AI reliance 43% better than alternatives
New metric detects hidden AI dependency by simulating how you'd work without the tool.
Researchers introduce the Offloading Score, a simulation-based metric that quantifies how much cognitive effort users offload to AI tools. Unlike traditional measures (output adoption or self-reports), it constructs a counterfactual workflow—estimating how the user would complete a task without AI—and computes the fraction of steps saved. In a study of 40 developers, the score detected a 43% higher reliance under time pressure (p=0.018), while baseline measures missed this difference. The framework helps users reflect on their own reliance and helps designers mitigate overreliance.
- Offloading Score simulates a counterfactual workflow to estimate steps saved by AI, measuring cognitive effort offloaded.
- In a study of 40 developers, the metric detected a 43% higher reliance under time pressure (p=0.018), while traditional metrics missed it.
- Higher reliance correlated with more subtask delegation and direct reuse of AI outputs; the score helps identify inappropriate reliance when combined with task outcomes.
Why It Matters
Provides a more accurate, behavioral measure of AI dependency, enabling better tools and self-reflection to combat overreliance.