OpenAI will no longer train its frontier models on the live internet
25 October 2026 · filed under f699a3e5cdaf
OpenAI has said it will stop training its frontier artificial intelligence models on data pulled directly from the live internet, according to a report published by TechCrunch on October 23. The company cited what it described as “significant risks” in continuing the practice.
TechCrunch’s report notes a notable contrast in timing. Weeks before this announcement, an OpenAI engineer had reportedly said that cutting models off from live internet data would not work as a solution to the risks the company is now citing as its reason for doing exactly that.
The shift marks a change in how OpenAI builds its most advanced models. Training frontier systems on the live internet, rather than on curated or static datasets, has been a standard approach across the industry, and a move away from it would represent a departure in method for one of the field’s leading developers.
TechCrunch’s summary identifies OpenAI’s stated rationale as safety-related, tied to the risks associated with live internet data specifically, though the company’s full reasoning as reported extends only to that general citation of risk.
The report is attributed to TechCrunch’s AI desk. No further detail beyond the headline, the summary of the engineer’s earlier remarks, and the stated rationale is available in the material reviewed here.