Inside Dreaming V3: How ChatGPT Now Builds a Profile of You Without Asking

The Privacy Shock: 96% of Memories Created Without Consent

The most alarming privacy concern is the system's unilateral operation. A February 2026 arXiv paper titled "The Algorithmic Self-Portrait" by Abhisek Dash and colleagues, accepted at the ACM Web Conference 2026, analyzed 2,050 memory entries from 80 ChatGPT users. The study found that 96% of memories "are created unilaterally by the conversational system," not by explicit user instruction . The same dataset revealed that 28% of these system-generated entries contained GDPR-defined personal data, and 52% contained psychological insights about the user . This raises immediate questions about consent and profiling under regulations like GDPR and the EU AI Act, which takes effect in August 2026 .

The Poisoned Well: Accuracy Issues and Response Distortion

A flawed memory system does more than just forget—it actively corrupts future interactions. OpenAI reports that Dreaming V3’s factual recall has improved to 82.8%, a significant jump from the 41.5% accuracy of the original 2024 system . However, this still means that roughly 1 in 6 stored memories could be incorrect. The danger is amplified because the system now infers implicit preferences from casual comments and past interactions, not just explicit instructions . If it incorrectly deduces a preference, dietary restriction, or life circumstance, that error can "poison every answer it gives you" going forward . The automated rewriting process for stale memories can compound these errors rather than fix them, quietly baking falsehoods into your permanent profile .

Data Deletion Is Deliberately Cumbersome

Regaining control of your data is a significant challenge. Because Dreaming V3 synthesizes memories from the entirety of your chat history, deleting an individual memory entry is insufficient. The system can re-extract the same information from past conversations during its next background synthesis cycle . To fully purge a piece of information, a user must manually locate and delete the specific chat history containing the original disclosure—a multi-step, non-obvious process that becomes effectively impractical for long-term users with years of conversation logs .

Context Rot Goes Systemic

This is where the concept of "context rot" becomes critical and more dangerous than before. Context rot is the well-documented phenomenon where an AI's performance degrades as its context window fills with conflicting, outdated, or irrelevant information . The standard fix has always been simple: start a fresh chat to clear the slate .

Dreaming V3 breaks this fix. The persistent memory profile means that stale or incorrect data—old jobs, past relationships, expired preferences—survives across sessions . The problem of context rot is no longer confined to a single long conversation; it becomes a systemic issue that infects every new chat. The AI's "memory" becomes a fixed source of noise, systematically undermining the accuracy of responses regardless of how carefully you manage individual sessions .

Limited Access and a Missing Audit Trail

The rollout itself adds another layer of concern. Dreaming V3 is initially available only to ChatGPT Plus and Pro subscribers in the United States, with no clear timeline for international or free-tier users . This creates a two-tier privacy environment where some paying users are subject to automatic profiling while others are not.

Furthermore, the shift from a simple, auditable text list of memories to a complex, background-synthesized memory state reduces user visibility dramatically. Users can no longer easily inspect, understand, or correct the full scope of what the system has inferred about them . The process has become a black box, eroding the user's ability to audit and control their own digital profile.