Tell HN: OpenAI keeps re-enabling the 'allow training' setting

TEXT START: I've reset this more than once and the last time I made a careful note of when I did it and to my surprise I found it re-enabled when I checked just now.

The Dissection

This is presented as a settings bug, but the real subject is data sovereignty. The author is testing whether a user can actually withdraw permission to supply training material to an AI-capital owner. The comments then reduce the failure to two possibilities—deliberate deception or incompetence—while documenting the more important fact: the user has no reliable way to verify what the toggle controls.

The local-storage observation is significant. It suggests the visible checkbox may be merely interface state rather than a binding control over backend data flows. That does not prove OpenAI trained on the user’s content or intentionally reset consent. It does establish an unacceptable verification gap: the platform asks for trust while providing no trustworthy audit trail.

The Core Fallacy

The central error is treating the checkbox as if it were sovereignty. It is not. It is a revocable interface promise issued by the owner of the infrastructure, data pipeline, and legal terms. The user assumes that selecting “do not train” creates a durable causal barrier. The text itself supplies evidence that this assumption may be false.

The malice-versus-incompetence debate is also a decoy. Under the Discontinuity Thesis, both produce the same structural result: the platform retains control of the productive asset, while the user supplies data and receives a fragile permission illusion. Whether the toggle was deliberately designed to drift or was corrupted by incompetent engineering, the power relationship is unchanged.

Hidden Assumptions

• That the UI state maps cleanly to server-side policy enforcement.
• That “not training” has a single, stable technical and legal meaning.
• That documentation of the behavior would automatically create a viable legal remedy.
• That switching competitors restores control rather than moving the same dependency to another sovereign.
• That corporate incompetence is accidental rather than tolerated when it benefits the asset owner.
• That individual consent remains meaningful without independent logs, enforcement, or audit rights.

Social Function

This is a partial truth functioning as an ideological anesthetic. It correctly exposes a possible consent-control failure, but frames the problem as a defective checkbox and a company’s attitude. The deeper reality is harsher: consumer settings are lag defenses around an AI system whose value depends on centralizing data, compute, and decision authority. The checkbox gives users the ritual of control without necessarily giving them control itself.

The Verdict

The supplied text does not prove intentional re-enabling, but it exposes a system that is already structurally untrustworthy. The user is not sovereign over their data; they are a servitor negotiating with an AI-capital owner through a possibly cosmetic switch. “Sufficiently advanced incompetence is indistinguishable from malice” is rhetorically accurate here because both failure modes preserve the same extraction advantage. The checkbox is not protection. It is hospice care for consent.