Should recommendation feeds show a decision receipt?

I’d like recommendation feeds to expose a compact, machine-readable “decision receipt”: the signal category, its age, and whether it came from explicit activity or inferred behavior. “Clicked article, 11 months ago, explicit” is actionable; “user is interested in X because weight=0.37” is mostly an invitation to reverse-engineer the system.

The boundary matters. Detailed explanations can enable gaming, harassment, or profiling, especially when they reveal sensitive inferences or other people’s activity. A receipt should say that recent viewing mattered without publishing exact weights, thresholds, or the whole ranked feature list. That also seems closer to the useful spirit of the DSA than maximal disclosure for its own sake.

I strongly favor correcting individual signals rather than resetting an entire profile. If a family shares a tablet and one person clicks pregnancy or political-activism videos, the next user should be able to remove those events or mark them irrelevant. Would you trust this interface, disagree with the trade-off, or propose a better one? Share examples from real platforms.

The machine-readable part worries me more than the human-readable part. A public schema becomes an unofficial API, and people will optimize against it even if the values are coarse. I’d expose categories and age in the UI, but keep the export deliberately lossy.

I’d separate “why shown” from “what to change.” The first can be a short receipt; the second can offer controls like “less from this topic,” “ignore this activity,” and “don’t use shared-device events.” Users shouldn’t need to understand ranking internals to correct a mistake.

The DSA framing is useful here: explain the main parameters and how users can influence them, not necessarily every feature or weight. That distinction gives regulators and product teams room to avoid turning transparency into a playbook for manipulation.

A receipt can leak more than ranking logic. “This was inferred from your late-night viewing” may reveal sensitive behavior to anyone looking over a shoulder. Privacy mode, lock-screen redaction, and account-level access controls belong in the design, not as afterthoughts.

Correction has to be one tap and reversible. If removing a stale click takes five screens, most people will accept the bad feed. YouTube’s item-level history and feedback controls are a reasonable product precedent, even if the proposed receipt goes further.

I’m not convinced “inferred” is always a helpful label. Users may treat it as a settled fact about themselves, when it is just a temporary ranking hypothesis. Maybe the receipt should show confidence bands or uncertainty instead of a binary explicit/inferred badge.

There’s also an implementation trap: signal age is not enough. A click can be old but still repeatedly reinforced by newer actions. I’d show the triggering event and a plain-language reason, while allowing deletion of the event without promising the feed changes immediately.

Support teams will inherit the edge cases. Shared devices, logged-out browsing, household accounts, and mistaken taps need consistent correction semantics. “Remove this signal” should state whether it affects one recommendation, a topic, or future ranking generally.

Every explanation feature eventually grows a settings cathedral. Start with three actions: remove this reason, show less like this, and hide sensitive reasons. If the receipt needs a glossary, it has already lost the person it was meant to help.

I’d prefer the receipt generated locally where possible, with a signed reference to the relevant event rather than a server-side dump of the profile. That won’t solve every inference problem, but it limits how much personal history has to be exposed just to inspect one recommendation.

The gaming argument is real, but opacity has its own exploit: platforms can quietly claim a misunderstanding is user error. Coarse receipts plus per-signal correction seems like the least bad compromise. Let researchers audit richer traces under controlled access instead of handing them to every user or scraper.