The Signal · Column 076

The AI Act Just Created a Job Nobody Has Hired For

Andrei van Noordt · 7 October 2026 · 2 min read

OpenAI has started watermarking the text its models produce for users in the European Union, a measure it frames as compliance with the AI Act. The mechanism matters less than what it announces. Machine output now carries a mark that reports where it came from, and the marking is not a gesture of goodwill. It is regulation arriving at the level of the sentence.

The reading most people miss is where the burden actually lands. Watermarking is an obligation on the provider, but the consequence travels down the chain to every company that puts model output in front of a customer, a regulator, a court or a candidate. Once provenance is detectable, the question stops being whether anyone can tell. It becomes whether you knew, whether you disclosed, and whether you can show your working. That is not a model problem. It is an operating problem, and it has landed on companies that spent two years shipping AI features without ever naming the person responsible for how those features are governed.

Detectable provenance turns AI governance from a policy document into an operational job with somebody's name on it.

I think this is the next quiet repricing. The profile that becomes expensive is the one that reads a regulation and a system diagram with equal confidence: someone who can trace where model output enters a product, decide what has to be disclosed, and write that into the build process rather than into a policy document nobody opens. That person is scarce because until recently the role barely existed, so the people who can do it sit scattered across machine learning engineering, data protection and risk. Most companies will discover they need one on the day an auditor or a client asks a question they cannot answer, which is the most expensive day to start recruiting.

There is an adjacent effect on everyone who hires. If model output can be marked, written work becomes weaker evidence of judgment than it was a year ago, and the conversation in the room has to carry weight the take home assignment no longer can.

For hiring leaders, the practical read is simple. Name the person who owns AI provenance and disclosure inside the engineering organisation rather than in the legal annex, and fill that seat before the first question arrives from outside the building. Compliance hires made under deadline cost more and land later than the risk they were brought in to cover.