Can feeds have a dry-run mode?

I’d like social platforms to offer a feed “dry-run” mode. Before muting a topic, unfollowing someone, or changing recommendation settings, I could preview the likely result and decide whether to apply it. The exploratory action itself should not become a durable signal about what I like or dislike.

That distinction matters because recommendation systems learn from behavior, including feedback such as dislikes, “Not interested,” viewing history, and similar actions. YouTube already lets people pause or delete history when researching something they do not want treated as a preference, which seems like a useful precedent. But a recommender also shapes the behavior it later learns from: testing the system can pollute the very profile being tested.

The preview would only be a counterfactual estimate, since content, timing, models, and other activity keep changing. Would you use this sandbox, and what controls or technical limits should it have?

The hard part is defining “not recorded.” A platform could exclude the click from ranking signals while still logging it for security, analytics, or experiment evaluation. I’d want a visible scope statement: excluded from personalization, excluded from ads, or excluded from every behavioral profile? Those are different promises.

I can see the product value, but the preview has to be immediate and legible. “Your feed may become 30% less sports” is more useful than a mysterious alternate timeline. If it takes a full simulation and a loading screen, most people will just hit mute and move on.

The YouTube history controls are a reasonable precedent for separating research activity from preference inference, but they do not establish that a dry-run exists. I’d be careful not to sell this as solved privacy architecture. It is a promising control with a fairly narrow, explicit guarantee.

I’d threat-model the preview itself. Asking “what happens if I unfollow this account?” reveals information about how the system classifies that account and me. The answer should be coarse, and the request should not expose a personalized ranking dump that can be scraped or used for inference.

There’s also a social signal problem. A preview can make people feel they are tuning a machine rather than choosing what to see. I’d show the change as reversible and temporary, with a clear “apply” button—not a gamified score that pressures users to optimize their identity.

Counterfactual replay sounds expensive and brittle. You need a snapshot of the candidate set, model version, policy filters, and interaction state, then you still cannot promise tomorrow’s feed matches it. Fine as an estimate; bad if the UI implies determinism.

I’m not convinced every exploratory action deserves a privacy moat. If I repeatedly preview a topic and never apply the change, that may itself be useful evidence that the control is confusing. The answer could be short retention and aggregation rather than pretending the event never happened.

Yes, I’d use it. The main control I want is an undoable preview, not another dashboard explaining my personality. If the feed changes after I confirm, fine; if it changes because I looked at the preview, the feature has defeated itself.