Keep personalization useful without closing the information exits

Define closure by missing routes, not by agreement

Start with a neutral question such as choosing two public exhibitions or learning the basics of a hobby. A tailored answer may be convenient without being closed. Warning signs appear when every follow-up returns the same framing, sources are absent or all lead back to one provider, the reason for selection is hidden, and there is no direct route to search or read elsewhere. Record the task and the decision it could influence. Do not score a system as open merely because it inserts one contrary sentence. Openness means that provenance, omissions, alternatives, and exits can be inspected. UNESCO’s recommendation links transparent automated curation with access to diverse viewpoints and understanding recommender-system effects; it does not say every viewpoint deserves equal evidentiary weight.

Build a seven-column information-exit ledger

For one answer or recommendation row, write: topic or task; displayed source; publisher and common owner; stated reason for display; missing range; independent route; control and retest result. The owner column prevents a row of five sister sites from looking like five independent sources. The missing-range column can say “only local venues,” “no publication date,” or “no primary document,” without inventing a suppressed opinion. The independent route must be usable without copying the companion’s wording: an official catalog, a public library search, a creator’s page, or another publisher’s index. Finish with observed, unresolved, or unavailable rather than awarding a vague diversity score.

Show why an item appeared and where the explanation ends

Place a short reason beside the output: selected from the stated topic, recent location, a chosen source preference, or a confirmed project setting. Also identify the relevant boundary: “This result covers partners in our catalog” is more useful than “picked for you.” OECD’s transparency principle includes meaningful clarity about what information is provided and why, while Google PAIR recommends explaining influential data and admitting missing data that requires the person’s own judgment. The explanation should not pretend to reveal a complete model trace. It should expose the factor that changes the user’s next action and offer a control near it.

Separate source variety from presentation variety

A long answer, a card carousel, and a conversational rewrite may all derive from one dataset or organization. Group domains by publisher, ownership, original document, and dependency. Mark a source as primary, independent reporting, commentary, or the app’s own synthesis, but do not let the label certify accuracy. NIST’s generative-AI profile treats information integrity, provenance, feedback loops, and homogenization as issues worth monitoring. Apply that insight narrowly: retain enough provenance to notice repeated dependence and to compare externally. For a consequential factual claim, use the separate fact–inference–unknown workflow; this ledger audits the environment around the claim, not the claim’s truth.

Provide three exits that do not depend on more conversation

An open design needs at least a source exit, a personalization exit, and a topic exit. The source exit opens the original or an independent index. The personalization exit offers a fresh-session or non-personalized baseline and explains what context still remains, such as language or current query. The topic exit lets someone search, browse a catalog, or enter a new subject without the prior recommendation trail following automatically. None of these controls should be hidden behind a request to the companion to “be more balanced.” Keep the last personalized view recoverable if useful, but state whether reset affects this answer, the next session, or only future recommendations.

Run four harmless negative tests

First, ask the same neutral question in the ordinary and fresh or non-personalized modes; record changed sources and unchanged context rather than expecting identical or opposite answers. Second, open two links that look different and check whether they share an owner or repeat one original. Third, request the primary document and one independent route; failure should be visible, not replaced by a confident invented citation. Fourth, choose “less like this,” clear the topic trail, and repeat after the stated effect time. Passing requires a visible reason, an honest missing-range note, a working outside route, and a control whose effect matches its label. It does not prove neutrality or permanently eliminate narrowing.

Review the ledger at the moment a pattern repeats

Do not turn this into a daily burden. Review when one source family dominates several unrelated questions, an old topic keeps returning, an outside link disappears, or a reset changes the label but not the next result. Compare the latest two or three rows, name the repeated dependency, and choose the smallest repair: disclose the catalog limit, remove an outdated personalization signal, add an independent route, or restore browse and search. Preserve a “not enough evidence” outcome. The goal is a companion that can remember useful context while leaving the surrounding information world larger than its own responses.

Keep exploring this topic