OpenAI's opt-out from training data flagged as buried and largely ineffective

On September 9–10, OpenAI's data training opt-out mechanism drew heavy criticism: multiple users found that simply toggling off "Improve Model for Everyone" does not actually stop OpenAI from training on chat data. A real opt-out requires submitting a privacy request through a hidden path, and its scope reportedly covers only ChatGPT, not internal models.

Confirmed

• Users such as ShashwatGoel7 shared that paying customers can disable "Improve the model for everyone" under Settings → Data Controls, but complained the option's naming is unintuitive—users would hardly guess it is the data training switch.
• Thread screenshots from edoordocontente show the "Do not train on my content" option must be manually checked rather than being on by default.
• Multiple posts (relayed by RexDouglass, niloofarmire, PMinervini, xennygrimmato, etc.) point out the real path to opt out of training data use is to visit OpenAI's official privacy request page at privacy.openai.com, not the in-app toggle.
• Investor Delip Rao found that even after completing the opt-out process, it only excludes "ChatGPT"—OpenAI makes no commitment not to train internal models such as reward models on the data.

Unconfirmed

• paulcal pushed back, arguing the in-app ChatGPT toggle is effective and the privacy form also works, citing what he calls clearly written Data Controls FAQ. The two sides disagree over the actual effect of the toggle versus the form, and there is no official verdict.
• Delip Rao's claim that the opt-out does not cover internal models comes from his personal interpretation; OpenAI has not publicly responded.

Why it matters

• The episode exposes usability problems in OpenAI's data controls across naming, defaults, and how deeply the option is buried. Ordinary users may well believe flipping a toggle protects their data, when the actual opt-out bar is far higher than expected.
• If the opt-out truly covers only ChatGPT and not internal models like reward models, users' real control over their data is substantially weakened—a direct concern for enterprise users and researchers making data compliance decisions.

2026-09-09 ~ 2026-09-10 · 7 related posts

• How paid ChatGPT users can turn off OpenAI training on their data — xennygrimmato_ · 2026-09-09
• [source] Debate over OpenAI data opt-out: in-app ChatGPT toggle is effective, per FAQ — paul_cal · 2026-09-09
• Researchers claim OpenAI's opt-out of training data is misleading and only covers ChatGPT — PMinervini · 2026-09-09
• Toggling off "improve mode" doesn't stop OpenAI from training on your data — RexDouglass · 2026-09-09
• OpenAI's 'Do not train on my content' option found not enabled by default — anshulkundaje · 2026-09-09
• [source] Toggling off 'Improve Model for Everyone' doesn't stop OpenAI training on your chats — PsychicorAI · 2026-09-10
• [source] OpenAI accused: buried opt-out only excludes ChatGPT, says nothing about internal models — niloofar_mire · 2026-09-10