What 94,000 interventions show about opening apps on autopilot

Data through September 15, 2026

Between June and September 2026, 840 Cortex app users hit the pause 94,587 times. One in ten times they chose not to go in. Instagram caused 79% of all impulses; the peak hour was 9–10 pm; and the single strongest predictor of resisting was not the hour or the app but whether the person had just labelled the open as “autopilot” — that doubled the resist rate.

Measured by Cortex app itself, anonymized and aggregated: 840 people, 94,587 interventions over roughly 90 days, users mostly Spanish-speaking. Hours are device-local. These are one app’s users, not a population sample, and no individual data is shown.

Instagram is the habit; everything else is noise

Of 94,587 impulses, 75,052 were Instagram — 79%. TikTok was second at 15%. Facebook, X, YouTube and WhatsApp together were under 6%. Part of this is selection (people install a pause app for the app that hurts most, and in Spanish-speaking markets that is Instagram), but the gap is too large to be only that: Instagram is the default thumb destination.

Resist rates by app tell a second story. Facebook: 23% — people who monitor Facebook are, apparently, already half-decided to leave it. YouTube 13%, TikTok 11%, Instagram 9%, WhatsApp 8%, X 7%. X is the hardest to say no to, which matches its shape: a two-minute check, forty times a day, each one cheap enough to justify.

Naming “autopilot” doubles the chance of not going in

After the pause, Cortex app asks why you’re opening the app: work, leisure, or autopilot. 45% skip the question. Of those who answer, 53% say leisure, 30% autopilot, 17% work.

The resist rate by answer: autopilot 13.1%, leisure 6.2%, work 5.2%. Same pause, same app, same person — and the moment someone admits “I didn’t decide this”, they are twice as likely to close the shield. That is the whole thesis of the intention question in one number. It also says something uncomfortable: people who answer “leisure” resist less than people who skip (11.2%) — a deliberate label is also a permission slip.

The day peaks at 9 pm and willpower peaks at 7 am

Impulses by hour form a long plateau from noon to midnight with the top at 21:00–22:00 (about 6,400 each). The quiet hours are 3–5 am (under 750 each). So far, unsurprising.

The resist rate is the interesting curve. It is highest in the morning — 13.1% at 7 am, 12.2% at 8 am, 11.5% at noon — and slides all evening: 8.4% at 21:00, 8.0% at 22:00, 8.1% at 23:00, 7.4% at 1 am, 5.4% at 3 am. The hour with the most impulses is also an hour where the pause works about 35% worse than it did at breakfast. If you can only protect one window, protect 9 pm to midnight.

Weekends are worse than Mondays

Impulse volume barely changes across the week (12,900–14,200 a day-of-week), but the resist rate does: Monday 11.0%, Tuesday 10.9%, Thursday 10.8%, Wednesday 10.3%, Sunday 8.6%, Friday 8.0%, Saturday 7.1%. The structure of a workday is doing some of the resisting for people; take it away and the pause has to do it alone.

The dodged pause comes back: what happens on a close without a decision

The resist rate only counts people who reach the decision screen. A separate share of pauses ends earlier: the user closes everything without deciding. Measured with its own event (“Intervention Abandoned”, 21 days ending 2026-09-13, 83,465 pause appearances): of every 100 appearances with a known outcome, 52.6 ended inside the app, 5.6 in “do something else”, and 41.9 in a close with no decision. Definition: an impulse is each automatic open attempt; stopped means that attempt did not end inside the app. By that definition, the pause stopped 47% of impulses.

And closing the pause is almost never an escape: 90% of those bails have another pause firing within 3 minutes, because the automation runs again when the app is reopened. Closing the pause does not dodge it; it postpones it a few seconds. Counting full chains — attempts strung together until a real ending — about 1 in 6 impulses ends outside the app.

14% of appearances record no outcome (force quits, duplicates); the shares above use known outcomes only. The window differs from the rest of this page because the abandon event shipped in late August 2026.