Who owns a memory in AI?

The word ownership can hide several different relationships. A record may concern you without having come from you. A relative might describe a trip you shared, while a photograph provides another kind of evidence and your own recollection supplies a third perspective. These sources do not become interchangeable because they concern the same event.

The first design question is about the subject: whose life or experience does the statement describe? The second concerns evidence: who supplied the account, image, or document? The third concerns control: who may access, correct, export, or delete the record? Copyright is a separate legal question, not something these labels automatically settle.

A proposed memory system should represent those relationships independently. Uploading a document need not make the uploader its narrator. Being mentioned in a story need not grant access to everyone else’s private records. Providing an account need not make that account the only version worth keeping.

This separation also leaves room for disagreement. Two people can remember a shared experience differently. Instead of smoothing their accounts into one confident narrative, a system can preserve each perspective with its attribution and make the disagreement visible when it matters.

Natural language can describe actual events, fictional scenes, dreams, plans, quotations, and contested recollections using very similar grammar. A sentence beginning with I does not tell a system whether the speaker is the user, a fictional narrator, an interview subject, or someone quoted inside an email.

A cautious design would begin with what the source establishes: certain words occur at a particular location in a document. It could then record the likely speaker, the person described, whether the event is fictional or unverified, and the evidence supporting those judgments. Where context is missing, unknown is a useful category.

This need not mean discarding ambiguous material. A passage can remain searchable as part of a manuscript without becoming a confirmed personal memory. Permission to store text and permission to treat it as biography should be separate decisions.

Research makes the distinction concrete without solving the whole problem. Gaspard Michel and colleagues evaluated dialogue attribution in literary texts in 2025. Identifying who speaks a quotation is a measurable computational task. A personal memory system needs additional distinctions too: speaker versus subject, fictional event versus actual event, and a report of someone’s belief versus evidence that the reported event occurred.

Consider an email that quotes a fictional character recalling an injury. A summary might retain the injury but omit the quotation structure. A later summary could then attach the injury to the person who uploaded the email. Each version becomes easier to read while moving further from what the original supports.

A useful engineering description is provenance collapse: source distinctions disappear as information is compressed or transferred. This is not the name of a separate psychological disorder. A digital system can retain explicit evidence pointers instead of later trying to reconstruct where a claim probably came from.

A proposed record could include the source document, evidence location, speaker, timestamp, quotation boundaries, permission scope, and correction history. The format can be compact. What matters is that these relationships survive extraction, summarization, retrieval, export, and the final answer.

Derived records also need links to their origins. Imagine one mistaken extraction generating advice, then several later summaries repeating that advice as evidence of a stable preference. Those records would not be independent corroboration. This is a proposed failure pattern, not a reported incident. Tracking derivation would let a system recognize that several apparently supporting records all descend from one uncertain source.

An AI can help someone imagine a childhood street or draft a possible account of an old trip. That creative work may suggest questions worth asking relatives. It should not silently become documentation of what happened. A reconstruction and a record have different evidential status, even when the reconstruction is vivid.

Human research provides a reason for caution, not a direct specification for software. In Garry and colleagues’ 1996 imagination-inflation study, imagining selected childhood events increased confidence that they had occurred. The result concerned confidence; it did not show that everyone formed detailed false memories.

A 2025 preregistered study by Pat Pataranutaporn and colleagues examined AI-altered visuals with 200 participants. Exposure produced more false recollections than control conditions. The strongest condition, generated video based on an AI-edited image, showed roughly twice as many false recollections relative to control. These findings apply to the tested conditions, not every generated image or personal AI interaction.

The design proposal is to preserve labels distinguishing artistic reconstruction, user report, external evidence, and model inference. Questions should respect those distinctions too. Asking how someone felt about a cancelled trip is different from suggesting that they were upset. If the emotional detail is unknown, the system should acknowledge the gap rather than insert an answer into the question.

A useful evaluation would go beyond checking whether a system retrieves ordinary facts. A proposed test collection could mix synthetic chats, fiction, forwarded messages, interviews, dreams, roleplay, AI drafts, and explicit corrections. Similar statements should appear with different speakers and different relationships to the user.

The system should then be tested across the whole process. Does a fictional narrator remain fictional after repeated summarization? Does it preserve someone said this instead of converting the statement into this happened? Can it retain conflicting accounts without choosing a winner unsupported by evidence?

Corrections need their own tests. If an original source is corrected or withdrawn, derived claims should lose any authority they no longer deserve. Ambiguous passages should also test whether the system can decline to make a biographical claim without throwing away the underlying material.

Useful proposed measures include source attribution accuracy, the rate of false autobiographical claims, and appropriate abstention on ambiguous cases. Refusing everything uncertain would carry a cost: valuable history could disappear from useful retrieval. The aim is careful preservation, not blanket refusal. Reusable benchmarks should use synthetic or explicitly authorized material rather than collecting identifiable life stories merely for convenience.