Physical AI Safety

The gap between robot intelligence and robot safety infrastructure

Independent weekly analysis by Mati Melchior.

Subscribe to the newsletter Download the free report

What is Physical AI Safety?

Physical AI Safety is the discipline of designing and assuring the safety of physical AI systems — robots, autonomous vehicles, and other machine-learning-driven machines that act in the physical world — at assurance levels comparable to established functional safety. It bridges three traditions, AI Safety, Functional Safety, and Robot Safety, none of which, alone, addresses the failure modes that arise when machine-learning components drive physical actuators.

Physical AI Safety by the numbers

The gap is measurable in public data. Every figure below is sourced.

Figure	What it measures	Source
4M+	industrial robots in operation worldwide	International Federation of Robotics (2024)
123	Israeli Physical AI companies identified in the national AI strategy report	Israel Innovation Authority, via Physical AI Safety analysis (2026)
<25%	of those companies show any visible functional-safety certification effort	Physical AI Safety analysis of Israel Innovation Authority data (2026)
77	robot-related accidents in US OSHA Severe Injury Reports (2015–2022)	Sanders, Şener & Chen, Applied Ergonomics (2024)
93	injuries recorded across those 77 incidents	Sanders, Şener & Chen, Applied Ergonomics (2024)
>60%	of robot incidents trace to unexpected activation — the robot moved when the worker believed it was stopped	Guo et al., Safety Science (2025)
20 Jan 2027	EU Machinery Regulation 2023/1230 applies in full — no transition period	EUR-Lex, Regulation (EU) 2023/1230 (2027)

Latest analysis

Regulation

When a software update makes you a new manufacturer — and when it doesn't.

Article 18 of Regulation (EU) 2023/1230 will make anyone who substantially modifies a machine its manufacturer — new conformity assessment, new risk assessment, new technical documentation, new declaration of conformity, new CE mark. Article 3(16) says the modification can be digital. It then gates that on two mechanical outcomes, and whether a machine-learning model update clears them is not settled by the instrument, its recitals, or the Blue Guide. What is settled is the exit: an update the original risk assessment already foresaw is not a substantial modification at all.

Sep 17, 2026Read →

Regulation

Thirty-nine assessors. Three of them are notified for machine learning.

On 13 September 2026 the European Commission's register listed 39 notified bodies under Regulation (EU) 2023/1230 and 144 under the Directive it replaces. Three of the 39 are notified for machine-learning safety components, and two for embedded machine-learning safety systems. A market cannot be assessed faster than its assessors were appointed.

Sep 16, 2026Read →

Regulation

You have 127 days. Nobody publishes how long this takes.

127 days separate 15 September 2026 from 20 January 2027, when Regulation (EU) 2023/1230 will apply. Seven certification bodies were checked for a published duration benchmark and not one publishes it, so what follows is what is documented instead: the first assessor notified on 8 September 2024, roughly twenty-eight months before the Regulation applies, no harmonised standard yet cited under the Regulation, and seven trade associations on record that machinery was the only sector left out of the Omnibus IV transition period.

Sep 15, 2026Read →

View all analysis