Melbourne, August 25 (The Conversation) – This month witnessed the launch of OpenAI’s GPT-5, lauded as a major advancement over its predecessors, touted as "much smarter across the board". The assertion was supported by high scores in diverse benchmark assessments, including areas like software coding, mathematics, and healthcare.
However, while these benchmark tests have become the gold standard for AI evaluation, they often fall short of revealing how these systems will perform and impact the real world.
AI researchers and metrologists – specialists in the science of measurement – have proposed a new approach to better assess AI models.
Metrology's Role
Metrology is crucial for ensuring the reliability of growing AI systems, alongside evaluating their broader economic, cultural, and societal ramifications.
Measuring Safety
Metrology gives us confidence that our tools, products, and services are dependable. In healthcare, AI holds the potential to refine diagnostics, personalize medicine, assist in disease prevention, and manage administrative duties.
Realizing these promises necessitates verifying the safety and efficacy of health AI—and finding reliable methods to measure it.
Unlike drugs and medical devices, which undergo rigorous safety and efficiency evaluations, AI lacks such robust evaluation systems across domains such as healthcare, education, law enforcement, and biometrics.
Benchmark Limitations
Presently, AI system evaluations rely largely on benchmarks that test output accuracy and relevance against human expertise.
Though numerous benchmarks span various knowledge domains, their performance metrics give scant insight into how these models impact real-world contexts. The real measure comes from assessing the system's deployment environment.
Benchmarks have become significant for AI developers, aiding product performance showcases and attracting investments. For instance, Cognition AI, after achieving notable performance on a software benchmark this April, raised USD 175 million, gaining a valuation of USD 2 billion.
However, benchmarks can be gamed. Meta tailored its Llama-4 model to boost its ranking on a chatbot site, and OpenAI had access to the FrontierMath benchmark dataset before testing its o3 model—raising concerns over result legitimacy.
This scenario echoes Goodhart’s Law: "When a measure becomes a target, it ceases to be a good measure." Experts caution that overemphasizing metrics can lead to their manipulation and redirect focus from long-term goals.
Beyond Benchmarks
If benchmarks aren’t enough, what is? In healthcare, early assessments of language models involved using medical licensing exams, which gauge doctors' competence and safety before practicing. Yet, AI models hitting high scores on these have been critiqued for not capturing real clinical practice's complexities.
As a reaction, more comprehensive evaluation frameworks like MedHELM have emerged. It encompasses 35 benchmarks over five clinical task categories, assessing models in decision-making, communication, and research.
While MedHELM aims to bridge this gap, it lacks depth in considering human-AI interaction and external impacts like culture and economy.
Establishing a broader evaluation ecosystem, incorporating academic, industrial, and civil society insights, is vital for crafting reliable AI evaluation methods.
Progress is underway in measuring AI’s real-world impact, including red-teaming and field testing within operation contexts. Refining and standardizing these methodologies will be pivotal in reliably translating AI measurements.
If AI fulfills even a fraction of its transformational promise, a measurement science is critical to protect all societal interests—not just those of the tech elite. (The Conversation) RD RD
(Only the headline of this report may have been reworked by Editorji; the rest of the content is auto-generated from a syndicated feed.)