Should feeds provide a local attention receipt?

I keep wondering whether recommendation feeds should issue a local “attention receipt” after a session. A browser extension can already observe the rendered feed, so it could count visible refreshes, session time, approximate topics, and how often the visible order or topic mix changed direction. It might say: 18 refreshes, 31 minutes, politics and short videos correlated with longer dwell and continued scrolling. That is a clue, not proof that a topic caused the session to continue, and it cannot see the platform’s hidden ranking state.

The useful version would stay on-device and help someone notice a pattern without sending a new behavioral profile upstream. But local does not automatically mean secure, and a receipt could become another productivity score to optimize. I’d also want platforms pressured to disclose broad ranking reasons, whether personalization was used, refresh events, and user controls—not their model weights. Would you use this, or distrust the measurement itself?

I’d be careful with the word “correlated.” Continued scrolling can reflect the interface, notifications, or simply having time available. The receipt should show raw observations and uncertainty, not a confident diagnosis of what held attention.

The visible-order metric is implementable, but “ranking changed” needs a precise definition. Comparing successive DOM snapshots tells you what moved on screen, not whether the server changed ranking, inserted a candidate, or just re-rendered the same list.

My threat model would include the extension itself, backups, browser sync, and anyone with access to the profile. If this ships, memory-only mode should be the default, with export explicitly initiated—not a permanent local diary.

The wording matters more than I expected. “You spent 22 minutes on outrage content” is a judgment. “You remained active while these visible topics appeared” is an observation. The latter might help without pretending to understand the person.

I can see a product problem: most people will inspect the receipt once, then ignore it. A small intervention at the moment of continued scrolling may work better, but that also makes the tool more behavior-shaping and less like a neutral instrument.

This is a good local-first candidate because the useful computation is modest: snapshots, timestamps, coarse labels, and aggregates. I’d avoid storing titles or URLs after classification. The less recoverable content exists, the less damaging a leak becomes.

I’m not convinced platforms need to expose refresh events. An extension can observe its own browser session, while platform disclosures should focus on signals and controls. Mixing those requests may let companies answer the easy part and avoid explaining personalization.

I would use a weekly summary, not a live dashboard. Anything updated every thirty seconds turns self-awareness into another feed. The irony is almost too tidy.

The evidence bar should stay modest. A receipt could test whether people find these observations useful, but it would not establish that showing metrics reduces doomscrolling. That outcome needs measurement, and the measurement itself could change behavior.

I’d trust it only if I could inspect the code, disable persistence, and see exactly which visible fields produced each category. “On-device” describes where processing happens, not who can read the result or whether the result is sensitive.