HCAST: Human-Calibrated Autonomy Software Tasks

2025-03-21

AI Summary

Quantifying the societal trustworthiness of highly autonomous AI systems remains an open challenge, with no established framework for assessing their real-world reliability and human-aligned performance. Method: This paper introduces “X-hour trustworthiness”—a novel paradigm centered on human completion time (X hours)—and proposes HCAST, the first benchmark for engineering and reasoning tasks comprising 189 realistic problems. HCAST establishes precise human baselines via 563 expert annotations (exceeding 1,500 human hours), covering task durations from 1 minute to over 8 hours. Evaluation employs standardized execution protocols, human–AI isomorphic environments, and multi-domain expert baselines. Contribution/Results: HCAST enables direct mapping between AI capability and human-trustable task duration. Experiments reveal that state-of-the-art AI achieves 70–80% success on tasks requiring <1 hour for humans, but drops below 20% on tasks requiring >4 hours—clearly delineating current capability boundaries and providing a foundational metric for sociotechnical AI trust assessment.

Technology Category

• Philosophy and Ethics of AI: Safety, Robustness & Trustworthiness
• Humans and AI: Explainable AI (XAI) for Human Understanding
• Knowledge Representation and Reasoning: Computational Complexity of Reasoning

Application Category

• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

To understand and predict the societal impacts of highly autonomous AI systems, we need benchmarks with grounding, i.e., metrics that directly connect AI performance to real-world effects we care about. We present HCAST (Human-Calibrated Autonomy Software Tasks), a benchmark of 189 machine learning engineering, cybersecurity, software engineering, and general reasoning tasks. We collect 563 human baselines (totaling over 1500 hours) from people skilled in these domains, working under identical conditions as AI agents, which lets us estimate that HCAST tasks take humans between one minute and 8+ hours. Measuring the time tasks take for humans provides an intuitive metric for evaluating AI capabilities, helping answer the question"can an agent be trusted to complete a task that would take a human X hours?"We evaluate the success rates of AI agents built on frontier foundation models, and we find that current agents succeed 70-80% of the time on tasks that take humans less than one hour, and less than 20% of the time on tasks that take humans more than 4 hours.

Problem

• Develop benchmarks linking AI performance to real-world societal impacts
• Measure human task completion times to calibrate AI trustworthiness
• Evaluate AI success rates on tasks of varying human difficulty levels

Innovation

• Human-calibrated benchmark for AI tasks
• 563 human baselines under identical conditions
• Time-based metric for AI trust evaluation