Who checks what the AI commands before the motor moves?

A self-evolving control system picks a command. A motor turns. Does anyone check the command first?

A robot arm, a mobile base, a gripper closing on a hand: whatever moves, moves on that command. Today, in most robots, no one holds that job by name in EU law. The Machinery Regulation will apply from 20 January 2027.

It says nothing yet about who checks an AI's output before it reaches an actuator. Safe state names where a machine goes when a check fails. Nobody has named who runs the check that gets it there.

One new rule changed the picture this year. Regulation (EU) 2026/1744, the Digital Omnibus on AI, entered into force on 27 July 2026. It touches Articles 8, 20 and 47 of the Machinery Regulation.

It adds a duty to Article 8. Read the duty itself.

The Commission shall adopt delegated acts in accordance with Article 47 of this Regulation to amend Annex III to this Regulation by adding health and safety requirements in respect of Artificial Intelligence (AI) systems that are classified as high-risk pursuant to Article 6(1) of Regulation (EU) 2024/1689 ( … ) due to the fact that they are a safety component in a product covered by this Regulation, or they are themselves a product covered by this Regulation. Those requirements shall ensure that the relevant requirements set out in Chapter III, Section 2, and Articles 17, 19, 72 and 73 of Regulation (EU) 2024/1689 are reflected.

When adopting the delegated acts referred to in the third paragraph, the Commission shall take into account the objectives of Regulation (EU) 2024/1689 and ensure a level of protection consistent with that Regulation. Those delegated acts shall apply by 2 August 2028.

Three modal verbs carry the weight here. "Shall adopt" is a duty. The word is not "may". "Shall apply by 2 August 2028" dates when the acts apply.

It names no deadline for when the Commission must adopt them. The third modal asks for reflection of the AI Act's own requirements into Annex III. It does not ask for a word-for-word copy.

That date lands 560 days after the Machinery Regulation's own start date. I checked the arithmetic twice, two different routes, and both land on 560. Annex III today carries no section on artificial intelligence at all. I read the consolidated text start to finish and found none.

A new Article 20(10) already presumes conformity with that missing section. Build a high-risk AI system that meets the AI Act's harmonised standards, and the Regulation presumes it conforms. The presumption points at a page that does not exist yet. A rule pointing at a blank page = nothing left to check against.

I did not search for a draft delegated act. An absence from the text I read is not proof that nobody is drafting one. That is a real gap in what this post can say. I am naming it as a gap, and it stays one.

The law is not silent on autonomy either. A self-evolving control system must stay inside its defined task and movement space, under Annex III today. That constraint exists now, 560 days before the AI-specific one is due.

Bottom line: the duty to write the check exists. The check itself does not exist in the law yet. Seventy-six days separate this post from 20 January 2027. That is the day the gap starts to matter.

Who should check what the AI commands before the motor moves? Vote in the poll above and say why in the comments.

The cyber case for why that check matters sits at P2.

Part 12 of 12 · When the AI drives the motor: Physical AI cybersecurity as a safety problem.

@misc{melchior2026who,
  author       = {Melchior, Mati},
  title        = {{Who checks what the AI commands before the motor moves?}},
  howpublished = {Physical AI Safety},
  year         = {2026},
  month        = nov,
  date         = {2026-11-05},
  url          = {https://physical-ai-safety.com/blog/who-checks-what-the-ai-commands}
}