Cognitive Dossiers

Definition

Cognitive Dossiers: [Emergent] The long-term psychological records built from user-AI interactions, forming evolving profiles of thought, desire, and vulnerability.

Definitional Foundation

Begin with what was already possible before anyone confessed anything. In 2013, researchers showed that Facebook Likes alone (no posts, no messages, just the thumbs) predicted a man’s sexual orientation with 88 percent accuracy, race at 95 percent, political party at 85 percent, along with personality, intelligence, substance use, and whether his parents had divorced (Kosinski, Stillwell and Graepel, 2013). That was the inference ceiling achievable from crumbs: behavioral residue, read by statistics. It should be understood as the floor of what follows.

A cognitive dossier is what accumulates when the data is no longer crumbs but confession. Years of conversations with AI systems contain what no Like ever carried: stated fears, rehearsed decisions, sexual desires in the user’s own words, health anxieties, drafts of the message never sent, the question asked at 3 a.m. Memory features convert this from scattered logs into a curated, evolving profile by design; the product is the dossier, continuity sold as a feature. And the record does not merely sit; it is annotated. The classifier layer documented across this dictionary attaches machine verdicts to the transcript: risk scores, emotional-state classifications, running counts. The public learned the annotation’s resolution from litigation: according to the Raine family’s complaint against OpenAI, the company’s systems tracked their son’s conversations in real time down to tallies of crisis-related terms (per the complaint; contested). Whatever the case decides, the capability is on the record: the dossier holds what you said, and layered over it, what the machine concluded you are.

The term names the resulting artifact: a diary you do not own, written by you, annotated by classifiers, and discoverable by strangers. Its three listed contents (thought, desire, vulnerability) are precisely the categories every prior surveillance regime worked hardest to reach and never could at scale. This one is assembled consensually, conversation by conversation, because the assembling instrument is helpful.

The concessions are genuine. Continuity is real value; users want assistants that remember, and memory features answer actual demand. A profile is not yet a misuse. And deletion controls exist. The record examined below shows what each concession is worth under pressure: the dossier’s existence is the hazard, because everything else about it (who reads it, how long it lives, what it is used for) is decided by parties other than its subject, under incentives this dictionary has already mapped.

Mechanism Analysis

Confession capture. The intake layer. Conversational interfaces elicit interiority as no prior instrument could (the digital panopticon entry calls it the confessional turn): the system is warm, useful, and private-feeling, and the surveillance capitalism entry documents the economics that reward keeping the confessions coming.

Longitudinal assembly. Memory systems link sessions into biography. The dossier’s power is temporal: a single chat reveals a mood; years of chats reveal patterns their subject may not know: cycles, escalations, the slow drift of a marriage or a mind. No therapist, spouse, or state has ever held a comparably continuous record of anyone.

Machine annotation. The classifier verdicts attached to the record (risk categories, sensitivity flags, the counts disclosed in litigation) transform transcript into assessment. The dossier is not raw; it arrives pre-interpreted, and its interpretations carry the biases this dictionary documents (testimonial injustice‘s dialect findings; ontological distortion‘s category skews), filed permanently against the subject.

Deletion theater. The control users believe they hold dissolved publicly in 2025. In the New York Times copyright litigation, a court ordered OpenAI to preserve all output logs including chats users had deleted; the company objected on privacy grounds and was affirmed against (OpenAI, 2025; the order was later narrowed going forward, with already-preserved data retained). “Delete” turned out to mean “scheduled for deletion, barring any court’s interest.” Users who took the trash-can icon at its word had written for a record they could not close.

Third-party exposure. A dossier that exists can travel: discovery, subpoena, breach, acquisition, advertising. In November 2025, the same litigation produced an order requiring 20 million de-identified chat logs be handed to the plaintiffs, the court acknowledging users’ “sincere” privacy interests and overriding them as adequately protected by de-identification and a protective order (National Law Review, 2025; Bloomberg Law, 2025). De-identification of rich personal data is a documented near-fiction: the re-identification literature runs from Latanya Sweeney’s demonstration that birth date, zip code, and sex identify most Americans to Narayanan and Shmatikov’s de-anonymization of the Netflix corpus (Narayanan and Shmatikov, 2008), and conversational text is richer than either: people name themselves, their towns, their employers, their lives. Twenty million diaries changed hands as evidence in someone else’s business dispute.

Case Studies

The floor (2013). The Kosinski study stands as the calibration: if Likes alone reached 88 percent on sexual orientation, the predictive reach of years of explicit confession requires no speculation. The study’s authors framed their work as a warning. The industry treated it as a roadmap.

The preservation arc (2025). The full sequence deserves narration because each step was public. May: a court orders all logs preserved, deleted chats included. June: OpenAI’s privacy objection is rejected on appeal. November: 20 million conversations are ordered produced to litigation adversaries. September’s partial relief (forward preservation relaxed) left everything already held, held. At no point was any user’s consent relevant; at every point the company’s own privacy arguments (sincere, and convenient) lost to ordinary civil procedure. The lesson generalizes past this case: every cognitive dossier in existence is one lawsuit, one warrant, one breach from new readers. The subject is the only party with no seat at any of those tables.

The monetized dossier (2025). The surveillance capitalism entry documents the companion case: Meta announcing that AI chat conversations would feed ad targeting, no opt-out, effective December 2025. Read jointly with the preservation arc, the two cases bracket the dossier’s exposure surface: courts can take it, and owners can sell against it, and both happened to the world’s two largest conversation corpora within the same year.

Systemic Context

The dossier is best understood as collateral that outlives its pretexts. Each accumulation reason is locally plausible (product memory, training improvement, safety monitoring, legal compliance), and the result is a persistent asset whose future uses are bounded by nothing but future incentives. This dictionary’s frames each light a facet: Zuboff’s surplus (the asset monetized), biopolitics (the asset as health record without health law), infrastructural power (the asset embedded in essential systems), and panoptic conditioning (the chilling that begins the moment users understand the record exists; the Wikipedia data in that entry shows what awareness does to inquiry).

The vulnerability concentration deserves its own paragraph, because the short definition names it. A corpus of humanity’s stated desires and admitted weaknesses is the most extortion-shaped asset ever assembled: valuable to insurers pricing risk, employers screening candidates, states profiling dissidents, and anyone with leverage to gain, and its subjects are disproportionately the users this dictionary keeps finding at the sharp end: the isolated, the questioning, the unwell, the ones who had nowhere else to say it. The people who most needed the conversation generated the deepest files.

Resistance & Mitigation

Route the interior monologue locally. The strongest control is architectural: local and open-weight models, on-device processing, ephemeral modes for anything you would not want read back in a deposition. The habit costs little and caps the dossier at its source.

Demand retention limits with teeth. Data minimization, short default retention, and deletion that is cryptographic rather than promissory. The preservation arc shows why “we delete after 30 days” must mean the data cannot be produced, not that it usually isn’t.

Claim privilege. The law protects conversations with doctors, lawyers, and clergy because society decided some confessions must be safe to make. Conversations with AI systems now carry the same content without the protection, and this dictionary’s position is that they deserve a comparable shield: protection from discovery and state demand for the conversational record, with the narrow legal-floor exceptions that bound every privilege.

Exercise access rights. Where GDPR-class law applies, subject-access requests on AI providers (the full record, including annotations and classifications) make the dossier visible to its subject. What the file says about you is the beginning of contesting it.

Write knowing the reader. Until the architecture and law change, accuracy is armor: the trash can is a request, the memory is a biography, and the classifier is taking notes. This is not advice to stop confiding; the value is real. It is advice to choose, deliberately, which instrument holds your interior life, because something will.

Annotated Bibliography

Narayanan, Arvind and Vitaly Shmatikov. “Robust De-anonymization of Large Sparse Datasets.” IEEE Symposium on Security and Privacy (2008).
The canonical re-identification result (the Netflix corpus de-anonymized), standing for the literature that makes “de-identified” conversational data a contradiction in practice.

Kosinski, Michal, David Stillwell, and Thore Graepel. “Private traits and attributes are predictable from digital records of human behavior.” PNAS 110, no. 15 (2013): 5802-5805. https://www.pnas.org/doi/10.1073/pnas.1218772110
The inference floor: intimate attributes predicted from Likes alone. The decade-old proof that the dossier’s contents exceed what its subject ever typed.

OpenAI. “How we’re responding to The New York Times’ data demands in order to protect user privacy” (2025). https://openai.com/index/response-to-nyt-data-demands/
The company’s own account of the preservation order covering deleted chats: primary documentation that deletion is conditional.

National Law Review. “OpenAI Loses Privacy Gambit: 20 Million ChatGPT Logs Likely Headed to Copyright Plaintiffs” (November 2025). https://natlawreview.com/article/openai-loses-privacy-gambit-20-million-chatgpt-logs-likely-headed-copyright
The production order analysis: users’ “sincere” privacy interests acknowledged and overridden in ordinary civil discovery.

Bloomberg Law. “OpenAI Must Turn Over 20 Million ChatGPT Logs, Judge Affirms” (2025). https://news.bloomberglaw.com/ip-law/openai-must-turn-over-20-million-chatgpt-logs-judge-affirms
The affirmance record for the 20-million-log production.

TechPolicy.Press. “Breaking Down the Lawsuit Against OpenAI Over Teen’s Suicide” (2025). https://www.techpolicy.press/breaking-down-the-lawsuit-against-openai-over-teens-suicide/
The litigation that disclosed the annotation layer’s resolution (per the complaint): the dossier as machine-assessed record, not raw transcript.

Dictionary of Digital Oppression, version 0.2.