OpenAI just made your own experts the training data
OpenAI has decided that its frontier models will no longer train on the live internet, citing significant risks. The decision arrived weeks after one of the company's own engineers had said publicly that cutting the models off this way would not work. So the most watched lab in the field looked at the open web, weighed what it would gain against what it would swallow, and shut the door anyway.
Most of the commentary read this as a safety story. I read it as a sourcing story. A model that no longer drinks from the live web is a model with a visible edge to its knowledge, and everything past that edge has to be carried in by hand. Not by the system. By whoever in your company knows what is actually true this quarter, which of your processes quietly changed, and which internal document is the one that really governs the decision.
A model that stops reading the internet does not stop needing to know. It just needs someone in your building to know instead.
In most organisations that job belongs to nobody in particular. It is distributed across a handful of people who have simply been there long enough, and it has never appeared on an org chart or in a salary band. The moment you deploy a closed model against real work, that invisible layer becomes the only thing standing between a confident answer and a correct one. Vendors cannot sell it to you, because it is specific to your company, and the model cannot scrape it, because it was never written down.
This is where the repricing happens. The generic skill of prompting a frontier model keeps getting cheaper, because the model keeps getting better at meeting you halfway. What gets more expensive is provenance: people who hold current, verifiable, proprietary knowledge and can state it clearly enough to be fed into a system and checked afterwards. Domain experts who write well. Operators who can explain why the exception exists. The ones who notice when the output is fluent and wrong.
For hiring leaders, the practical read is simple. Before you sign the next AI tooling contract, name the person who will own the knowledge the model cannot reach, and give them the time to do it properly rather than adding it to someone's evening. If that person does not exist on your team yet, hire them now, while the market still treats the role as optional.
Planning a senior AI or data hire and want the market view first? We share it either way.
Send us the brief →