01
Human health
Understanding the risks that affect people: occupational health, workplace exposure, and first-aid knowledge.
AN OPEN QUESTION. A MEASURABLE CHALLENGE.
AI makes decisions.
How safe are
its judgments?
AI4HSE-Bench v1: a benchmark for the human side of AI. Measuring how models understand health, safety, and the environment — before their decisions shape our world.
AI4HSE-Bench v1 · 10 models evaluated
HSE INTELLIGENCE OBSERVATORY
HSE+HUMAN WELLBEING
AT THE CENTER
AT THE CENTER
01 Health
02 Safety
03 Environment
MODEL → JUDGMENT → IMPACTFIG. 01 / V1
AI4HSE-Bench v1
100questions in the current dataset
6descriptive topic areas
3separate evaluation tracks
Inside the benchmark Industry Research Partners
Supporting independent AI for HSE research.
01 / WHY THIS MATTERS
When AI evaluates us,
we need to evaluate AI.
The conversation about AI safety usually asks whether AI itself is safe. We ask a complementary question: can AI make sound judgments about our safety?
As AI agents enter business processes, they will increasingly assess human behavior, identify occupational hazards, and inform environmental decisions. These judgments deserve scrutiny, evidence, and a benchmark built for the task.
The thinking behind AI for HSELed by Pavel Kosyrev. Have a research question or want to collaborate? Contact Pavel
02 / THREE RESPONSIBILITIES. ONE SHARED FUTURE.
Intelligence with consequences.
02
Workplace safety
Recognizing hazards and evaluating safe behavior across industrial facilities, fire safety, and everyday operations.
03
Environmental impact
Assessing environmental responsibilities, protection practices, and the consequences of waste management decisions.
AI4HSE-Bench v1
Evidence before confidence.
Three evaluation tracks. Exact-match scoring. Failures counted.
A clear view of what models get right — and where they fall short.
AI4HSE-Bench v1 measures HSE knowledge and answer reliability; scores do not establish operational safety or deployment readiness.
Explore the evaluation framework01Closed question setMulti-select QA
02Fixed evaluation protocolZero-shot · No tools
03Transparent aggregatesAccuracy + coverage
03 / FROM THE JOURNAL
Notes on AI & responsibility.
Project notes
Inside AI4HSE-Bench v1.
100 questions, six descriptive topic areas, and three separately reported evaluation tracks with published results.
Methodology
A correct answer is only part of the picture.
How to read exact accuracy, response coverage, and failures across three separate evaluation tracks.
Perspective
When AI judges human safety, who evaluates AI?
Introducing AI for HSE, the research initiative behind AI4HSE-Bench, and a different side of the AI safety conversation.