A Library AI Kept His Addiction Research After Logout

In this composite account, the patron is called Daniel. In October 2024, he spent part of an afternoon on a public-library computer reading about opioid withdrawal and medications used to treat addiction. He had been trying to understand whether treatment would interfere with work and whether a nearby clinic accepted people without insurance.

Daniel chose the library because the computer was not his. He shared a home with relatives, and questions about addiction did not stay private there. His understanding of a public session was plain: the browser opened without his accounts, and logging out would erase what he had done.

An AI assistance panel appeared beside his search results. It offered to summarize medical pages and turn broad searches into narrower questions. Daniel used it twice. One answer helped him distinguish supervised withdrawal from ongoing medication treatment, a distinction he had not understood from the clinic pages alone.

He found that useful. He printed a short list of local treatment providers, logged out, and watched the browser return to its welcome screen.

The visible history disappeared.

The line in the dashboard

Three weeks later, a librarian reviewing an administrative dashboard saw a row labeled opioid withdrawal and medication treatment after relapse. Beside it were the library workstation label, an October date, a minute-stamped session time, and a confidence indicator near the top of the scale.

That row is the hard fact around which this story turns. Staff had not typed it. Daniel had not entered that sentence. The AI feature had produced the phrase after combining his prompts with search terms, page titles, and portions of pages opened during the session.

The dashboard existed to show whether patrons were using the assistance feature and whether its answers appeared relevant. Most rows were ordinary: résumé formatting, property-tax questions, help with a school assignment. The addiction-treatment line stood out because it described a sensitive concern in direct language, while also making it possible to narrow the activity to one physical computer and a small span of time.

The librarian could not see Daniel’s name in that row. She did not need one to understand the risk. Staff sometimes help patrons at particular computers, and reservation records can exist separately for limited periods. A person who remembered assisting someone with printing, accessibility, or a frozen page could connect an inferred topic to a face even when the dashboard itself held no account information.

In this case, the librarian remembered helping Daniel print directions near the end of the session. She could not prove that every search belonged to him, but she had enough context to suspect it. That uncertainty did not make the dashboard less revealing. It showed how a system can avoid collecting a name while still leaving a practical route back to a person.

What the AI added

A conventional public browser can retain history, cookies, downloads, or cached files unless those records are cleared. The library’s logout process was configured to remove that local session data. The generated topic row lived elsewhere, in the AI provider’s analytics system, so resetting the computer did not touch it.

That separation matters. If the assistance feature were removed from the story, the central artifact would not exist. Staff might still have technical logs showing that a workstation connected to a clinic website, but they would not have the concise assertion that somebody at that computer was researching opioid withdrawal and medication treatment after relapse.

The model created that assertion by compressing several ordinary signals. Daniel had searched for withdrawal symptoms, opened pages about treatment medication, and asked the assistant to compare two forms of care. A language model grouped those actions into a single likely intent, then assigned a confidence level based on how consistently the material pointed toward the same subject.

The system did not know why he was reading. It could not distinguish personal need from concern for a relative, research for fiction, or preparation for volunteer work. Its confidence measured the consistency of the text pattern, not the truth of a claim about Daniel. The summary happened to align with his purpose, but the dashboard did not contain evidence strong enough to establish that.

That is one of the privacy changes introduced by generative and inferential systems. Raw activity can be scattered and hard to interpret. A generated summary makes it legible at a glance, which is useful for evaluating a service but also easier to remember, share, or connect to someone seen in the building.

The summary also carried language Daniel had not chosen. He had searched for relapse rates as part of comparing treatment options, yet the phrase “after relapse” made the row sound like a description of an event in his life. The model had turned a research topic into something closer to a personal condition, even though it had no reliable basis for doing so.

A disclosure that did not match the experience

The library had posted a general notice that interactions with the assistance feature could be processed to operate and improve it. Daniel had not understood that to mean a generated account of his subject would remain available to staff after logout. The notice did not explain the difference between clearing the public computer and deleting records held in a separate analytics system.

The librarian raised the row with her manager and asked how long those summaries remained. The first answer described a retention period of one month for dashboard data, with longer storage possible for records used to evaluate system quality. It was unclear whether removal from the visible dashboard also removed copies from backups or provider logs.

That gap was not resolved while Daniel was still using the library computers. The provider could explain what appeared in the dashboard, and the library could explain its logout process, but neither account gave the librarian a complete view of every copy created after the model processed a session.

The library disabled topic summaries for sensitive subject categories during the review. Staff could still see aggregate use of the assistance feature, but the dashboard stopped displaying free-text descriptions for health and crisis-related sessions. The underlying assistant remained available because some patrons used its summaries to read dense benefit notices or medical pages.

Daniel did not ask for the feature to be removed. When the librarian told him what she had found during a later visit, he said the explanation about medication had helped him. He objected to the retained row, not the assistance itself, and continued using public computers while avoiding the AI panel for health searches.

The line opioid withdrawal and medication treatment after relapse stayed visible until the dashboard’s monthly retention period ended. The librarian checked again after that point. The row was gone, though she still did not know whether its disappearance meant deletion from every system that had handled it.

Questions people ask

Does logging out of a public computer erase AI records?

It erased the local browser session in Daniel’s case, but it did not erase the AI feature’s server-side summary. The two records were controlled by different systems. The computer returned to its welcome screen while the generated topic, workstation label, and session time remained in an administrative dashboard.

Can an

AI infer a sensitive condition without a person stating it?

It can infer a likely topic from prompts, searches, page titles, and selected page content. That does not establish a diagnosis or prove who the research concerns. Here, the model compressed related activity into an addiction-treatment label and gave it high confidence, even though its wording implied a personal relapse the evidence did not show.

Could library staff identify the patron from a device and time?

The dashboard did not include Daniel’s name, but a workstation label and narrow time window could be combined with human memory or separate reservation information. The librarian remembered helping him print directions. That made identification plausible without turning the AI record into a formally named profile.

Was the

AI feature useful despite the privacy problem?

Daniel found its explanation of treatment options useful and did not regret reading it. He stopped using the panel for health research after learning about the dashboard. The provider’s monthly retention period passed, the generated row disappeared, and Daniel kept the printed treatment list folded inside his notebook.