Smart alerts: when AI tells you that you're about to drift (and is usually right).

The fourth alert

Last week the smart alert fired on me four times. The first three were correct. I was mid-afternoon, I had clicked a homepage recommendation without a clear purpose, I had followed one autoplay into another, and the alert appeared: "This session looks like it may be drifting."

I closed the session each of those times. Not because the alert told me to, it doesn't tell you what to do, but because seeing the description of what was happening gave me enough distance from the reflex to make a decision.

The fourth alert was wrong. I was watching a video in a topic area I don't usually cover, which pattern-matched to drift because it was algorithmically recommended and outside my typical channel range. It was actually a deliberate choice. I dismissed the alert and kept watching.

I am telling you about the wrong one first because I think the honest account of a feature like this starts with its failure modes, not its successes.

What triggers the alert

The smart alert monitors your current session in real time and compares it against a model of what your drift sessions look like. The model is built from your personal history, it is not a generic threshold, but a profile of when your specific behavior tends to shift from intentional to reactive.

The signals it watches:

Entry chain depth: How many consecutive recommendation-to-recommendation clicks have occurred in the session without a search or subscription-feed break. My personal drift threshold is around 4 consecutive recommendations. Yours may be different.

Completion rate drop: A sudden decrease in completion rate within a session, watching half a video, abandoning it, clicking the next recommendation, is a strong signal of reactive clicking. The alert looks for completion rate dropping below your personal session baseline.

Topic narrowing: When the session is trending toward a subset of your known drift channels, the channels that appear disproportionately in your low-Clarity sessions, the alert considers that a contributing signal.

Time-of-day context: If the session is occurring in one of your identified drift windows (for me, mid-afternoon), the model weights the other signals more heavily. A recommendation chain at 2pm triggers the alert earlier than the same chain at 9am.

No single signal fires the alert. It requires a cluster of signals pointing in the same direction, weighted by your personal drift profile.

What the alert looks like

It appears as an overlay at the bottom of the browser, unobtrusive, easy to dismiss, not blocking anything. It shows two things: a brief description of what the model detected ("4 recommendation-follows, completion rate dropping, mid-afternoon session") and a single option to end the session.

It does not moralize. It does not tell you that you're wasting time or making a bad choice. It surfaces what the behavioral model sees and leaves the decision to you. That framing was a deliberate choice. An alert that lectures you is an alert you learn to dismiss without reading.

The false-positive question

One in four of my alerts last week was wrong. Is that acceptable? I think it depends on what you're comparing against.

If the comparison is a perfect detection system, no. But perfect detection of behavioral intent in real time is not achievable. The model is working from behavioral proxies, entry chain depth, completion rate, time of day, not from direct knowledge of your mental state. A session that looks like drift and isn't is an inevitable output of that approach.

If the comparison is no alert system at all, then three correct interventions and one false positive is a reasonable trade. The cost of the false positive is a few seconds of reading a description and clicking dismiss. The benefit of the true positives is a structured opportunity to exit a session that was actually going nowhere.

I've calibrated my alert threshold to sit slightly higher than the default, I prefer fewer alerts with higher confidence over more frequent alerts with more false positives. The threshold is configurable.

What the alert cannot detect

The model doesn't know what you want to accomplish. A low-Clarity session is not always bad. Sometimes ambient YouTube is exactly what you need, background noise while you work, something familiar when you're tired, a comfort routine that serves a genuine purpose. The alert doesn't know when that's the case.

It also doesn't detect the slow drift that happens below the threshold. If you're watching recommendation chains that are just short enough to avoid triggering the entry-chain signal, the alert won't fire. The model catches the pattern when it's clear; it misses the pattern when it's gradual.

And it doesn't prevent the drift from starting. It detects and surfaces. The response is yours.

Why I still use it

The three correct alerts last week represented a concrete behavioral outcome: three sessions I would have continued in a low-Clarity state instead ended early or redirected. The false positive cost me a few seconds.

The more important effect is the meta-level one. Having an alert system means I approach sessions slightly differently than I would without one. There is a background awareness that my behavior is being monitored by a model that I've calibrated. That awareness doesn't prevent drift, but it adds a small amount of friction to the moment when drift starts, and small friction at the right moment can be enough.

The smart alert works best when you treat it as information rather than instruction. It is telling you what it sees. What you do with that is your call.

Smart alerts are available on Gazenest Pro.