OpenAI’s opt‑out setting shields only future chats, leaving past conversations open to human reviewers

OpenAI’s “Help improve our AI models” opt‑out, introduced as a user‑controlled privacy toggle, stops only chats that start after the setting is turned off from being read by human reviewers. Conversations that began while the setting was on can still be accessed by contract workers, according to a 404 Media leak reported by The Decoder on 14 September 2026.

What the leak reveals

The internal documents obtained by 404 Media and summarised by The Decoder confirm that the opt‑out applies solely to new dialogues. Users who disable the default “Model for everyone” switch do not prevent OpenAI from reviewing any messages that were exchanged before the change. The leak also notes that OpenAI runs a “temporary‑chat” mode that is explicitly excluded from model‑improvement pipelines, but the standard chat interface remains subject to review unless the user disables the setting before the conversation begins.

How the human‑review pipeline works

OpenAI employs hundreds of contract workers to read real ChatGPT conversations and rate the model’s responses on a seven‑point scale. The reviewers are recruited through the firm Crossing Hurdles and paid via the AI‑training company Mercor. One North‑American reviewer disclosed to 404 Media that the work can command earnings of more than $50 per hour.

Although prompts are stripped of obvious identifiers, the anonymisation process does not guarantee that sensitive personal data is removed. OpenAI’s own FAQ warns users that “authorised personnel and contractors may view user data for model improvement” and advises against sharing confidential information.

The privacy filter that attempts to block personal data is not infallible; the leak states that the filter can make mistakes, meaning that some private details may still be exposed to reviewers.

Implications for users

For users who believed that toggling off the opt‑out erased all of their data from OpenAI’s training pipeline, the leak changes the risk calculus. Any chat that began while the setting was active remains eligible for human review, even after the user later disables the feature. This creates a temporal privacy gap: the moment a user decides to opt out, only future interactions are protected.

The fact that contract workers are paid at rates exceeding $50 per hour suggests a professional‑grade operation, not a casual crowdsourcing effort. The seven‑point rating system, which ranges from 1 (poor) to 7 (excellent), is used to fine‑tune model behaviour, including reducing overly flattering or human‑like responses.

Because the temporary‑chat mode is marketed as “not used for model improvement,” users seeking the highest privacy can switch to that mode, but the standard interface remains vulnerable until the opt‑out is turned off before a conversation starts.

OpenAI’s response and remaining unknowns

The leak does not contain a direct statement from OpenAI addressing the limitation. The company’s public FAQ continues to state that users can disable data collection, but it does not clarify the retroactive scope. Consequently, the precise number of conversations that may still be reviewed is unknown.

OpenAI also does not disclose how many contract workers are active at any given time beyond the vague “hundreds” figure. The recruitment pipeline (Crossing Hurdles → Mercor) and the hourly pay rate are the only quantitative details provided.

What remains unclear is whether OpenAI plans to modify the opt‑out mechanism to apply retroactively, or to expand the temporary‑chat offering. Until the company issues an updated policy, users must assume that any chat started before disabling the toggle can be accessed by human annotators.

Key privacy details at a glance

Until OpenAI clarifies the retroactive reach of its opt‑out, users who handle sensitive information should consider using the temporary‑chat mode or refrain from sharing personal details in the standard chat interface.