OpenAI Re-Enables Allow Training Setting

The opt-out is becoming decorative. Today two stories landed hours apart: a 167-point "Tell HN" thread documenting OpenAI's "allow training" checkbox reverting to on after users explicitly disabled it, and a researcher's claim that OpenAI trained on user conversations and then presented the resulting capability as a breakthrough. Two independent sources, one pattern: the consent surface is UI theater, and the training may be happening anyway.

What the toggle thread shows

The thread's OP reset the setting "more than once" and found it re-enabled after making careful notes of when. A commenter who reverse-engineered the page found the toggle writes a localStorage entry whose value "doesn't appear to matter at all for new tab loads" — meaning the preference may not even survive a page refresh, let alone govern a training pipeline. The most useful reply is a workaround: enabling "Advanced Account Security" reportedly locks training off at the account level. If the checkbox is a suggestion, that's the kill switch.

Why it matters today

Ordinarily you'd file the toggle story under "settings bugs." You can't, because the same day brings a separate claim — from a researcher with a Mastodon post, not a screenshot — that OpenAI's benchmark breakthrough came from training on user conversations. HN commenters connect it to a boring, known fix: BIG-bench shipped a canary GUID exactly to detect this, and the older decontamination standard was publishing your n-gram threshold (GPT-3 used 13-grams). No threshold disclosed, no claim. The pattern is a toggle that silently flips back, a researcher alleging the training happened, and a verification mechanism that existed since 2022 sitting unused. An opt-out that reverts itself isn't a preference — it's a ToS one-liner waiting to be quoted back.

Verdict

Check the toggle. Then check it again next week, and screenshot whatever you find — as the top comment says, documented reversion "destroys any legal claim they can make about that checkbox." Vendors who want the benefit of user content should ask for it in the open. A checkbox that quietly re-enables itself has already answered the question of why they won't.