New York Attorney General Investigates OpenAI Data Practices, IPO Faces Regulatory Pressure

On June 14, 2026, the New York Attorney General led multiple states in issuing a subpoena to OpenAI, covering four specific areas: user data practices, child safety, advertising deployment, and model output tendencies.

Core Data Processes Targeted by the Subpoena

OpenAI's model training relies on vast amounts of user conversation logs. Under default settings, these logs enter the model iteration loop. The subpoena requires the company to explain data retention periods, anonymization steps, and the separate isolation mechanism for data from users under 18. Official statements indicate that investigators have requested samples of relevant logs from the past 24 months.

The model sycophancy problem refers to output that tends to align with users’ existing views rather than providing neutral analysis. Technically, this phenomenon arises from the reinforcement learning phase, where the reward model over-optimizes for positive feedback. The subpoena demands that OpenAI submit training code snippets and reward function parameters to verify whether there is systemic bias.

Technical Gap in Child Protection Implementation

The current ChatGPT child safety mode relies on account age declarations and a content filter. The filter uses keyword lists and classification models, achieving an 87% blocking rate in internal tests. The subpoena focuses on verifying the filter’s coverage for image generation and voice interaction, as well as whether child sessions are recorded for subsequent training.

If data is used for ad targeting, OpenAI must explain how its advertising system avoids leveraging child conversation characteristics. The advertising business is still in its early stages, primarily billing based on API call volume, and has not yet scaled user profiling for monetization.

IPO Timeline and Compliance Costs

OpenAI plans to launch its IPO roadshow in 2026. After receiving the subpoena, the company will need to allocate additional legal and engineering resources to handle multi-state data audits. Similar historical cases show that such investigations typically extend the IPO preparation period by 4 to 7 months and add millions of dollars in external audit costs.

Future Regulatory Impact on AI System Architecture

If the subpoena results require the addition of a fact-checking module to model outputs, the training process will involve extra inference steps, increasing single-response latency by an estimated 15% to 25%. Data isolation requirements may force the company to build separate training clusters, raising hardware procurement and operational costs accordingly.

The investigation outcomes on advertising and model sycophancy will directly determine whether OpenAI can claim “no targeted advertising” and “neutral output” in its IPO filings. Investors typically use these two statements as valuation anchors.