467 Said They'd Quit Their Phones. Only 119 Did. · Issue #200

Opening

Dear subscriber, let me tell you about an experiment.

A research team in Canada and the US made an offer to 467 people: “Would you be willing to cut off your smartphone’s mobile internet completely for 2 weeks?” They could still make calls and send texts, and still use the internet on their laptops. Only the phone’s internet connection would go dark. The compensation was generous, too.

All 467 people said yes. And the number who actually held out for the full 2 weeks was 119.

This study¹, published in PNAS Nexus in February 2025, has mostly been covered with the takeaway “cutting phone internet improved people’s mental health.” That’s an accurate summary. But the number I kept coming back to wasn’t the effect size — it was that 25.5% compliance rate. Let me give you the conclusion up front: the real finding here isn’t that connectivity is harmful. It’s that even people who know it’s harmful and decide to quit still mostly fail.

First, what the experiment actually changed

Let’s start with the numbers.

Participants installed an app called Freedom. Its “lock mode” prevents users from disabling the block from within the app itself. The research team used this to objectively track whether the block was actually turned on — a methodological strength, since it didn’t rely on participants’ self-reports.

The intervention group’s average daily screen time dropped from 314 minutes to 161 minutes — from 5 hours 14 minutes down to 2 hours 41 minutes. Three outcomes moved as follows:

Subjective well-being (life satisfaction + positive/negative affect): d = 0.45²
Mental health (depression, anxiety, anger, social anxiety, on American Psychiatric Association scales): d = 0.56
Sustained attention (objectively measured via gradCPT³): d = 0.23

90.7% of participants improved on at least one of the three. The team described the attention gains as “comparable to reversing 10 years of age-related decline.” That’s an impressive figure for a 2-week intervention.

Now let me put on my fact-checker hat for a moment. The most widely quoted line from this paper is that “the improvement in depressive symptoms exceeds the meta-analytic effect size of antidepressants.” That comparison rests on a 2008 meta-analysis by Kirsch et al.⁴ — a reanalysis of data submitted to the FDA that sits among the lowest estimates of antidepressant efficacy in the literature, and remains contested to this day. The effect size of 0.32 itself has been fought over, reanalyzed, and rebutted for over a decade. Swap out the comparison point and this sentence may no longer hold. Worth being careful if you’re citing it.

The real finding is in the funnel

Now for the main event. Here’s how 467 people became 119, step by step.

Stage	Count	Note
Agreed to the 2-week block, received a code	467
Actually redeemed the app code	272	195 never even signed up
Set up the initial block	266
Completed all 3 surveys	249
Maintained the block for 10+ of 14 days	119	25.5% of those who agreed

First, 195 people — 41.7% — never even signed up for the app. They agreed to get paid, said yes, received the code, and simply… didn’t do it. And among the 249 who installed the app and completed the surveys through to the end, just under half — 119 people, or 47.8% — kept the block active.

Here’s an important piece of context: these weren’t randomly selected members of the general public. At the start of the study, 83% rated their motivation to reduce phone use at 5 or higher out of 7, and 79% said they had ample self-perceived ability to cut back. In other words, these were people who were both motivated and confident.

25.5% of them made it through 2 weeks.

One more thing. After the intervention ended, screen time bounced back from 161 minutes to 265 minutes. It didn’t fully return to the original 314 minutes at T1, but once the enforcement was removed, most people snapped back toward where they started — even the ones who had physically felt the improvement.

This isn’t a story about the individual weakness of 467 people. If the same pattern shows up across all 467 of them identically, that’s not a trait of individuals — it’s a property of the structure.

The paradox of felt self-control

The research team used mediation analysis⁵ to test why the improvements happened. The results are striking.

Here’s the ranking of pathways explaining the improvement in subjective well-being, by size:

Increased sense of self-control (0.169) ← by far the largest
Increased social connectedness (0.084)
Increased offline time (0.072)
Decreased media consumption (0.046)
Increased sleep (0.020)

Once you strip out all five pathways, the intervention’s direct effect statistically disappears (c′ = 0.034, not significant). For mental health, self-control (0.209) was also the top pathway. And self-control itself moved by d = 0.66 between pre- and post-intervention — the second-largest shift among the five mediators, right behind offline time (d = 0.70).

Do you see the paradox here?

What people felt they’d regained was “self-control.” But what they’d actually done was hand self-control over to an app.

Freedom’s lock mode doesn’t strengthen willpower. It eliminates the need to exercise willpower in the first place. Participants weren’t winning the moment-to-moment battle of “should I open Instagram right now?” — they were placed in an environment where that question never came up at all. And yet, the sensation they walked away with was: “I am in control of my own life.”

The sense of self-control, it turns out, wasn’t the cause of self-control — it was the byproduct of good environmental design.

One more thing worth flagging: attention improvement was the one outcome none of these five pathways explained. Not offline time, not sleep, not the sense of self-control. This suggests well-being and attention are damaged and restored through entirely different channels. Nobody has an answer for this yet, and I think this is the most honest part of the whole paper — the part where it admits it doesn’t know.

Oz’s Lens

There’s a pattern I’ve confirmed over and over while building GTM strategy.

Features that rely on user willpower fail, almost without exception. The moment onboarding asks users to “go into settings and turn this on,” conversion collapses. Features that encourage users to “build a habit” leave no trace on the retention curve. On the flip side, I’ve repeatedly seen a single default setting change move an entire metric. In product circles, this is already common sense: defaults shape behavior, not intentions.

But we rarely apply this common sense outside of product design.

The entire digital wellbeing industry is currently designed in exactly the opposite direction. Screen time dashboards show you statistics, induce shame, and then hand the decision back to you. Usage alerts tell you “you’ve exceeded the limit you set,” then place a “dismiss” button right next to it. What all these features share is that they outsource self-control to the individual. The 25.5% figure this paper produced is data on how that outsourcing arrangement tends to end.

And I find this uncomfortable, honestly — because we’re the ones designing that environment. The moment engagement becomes a KPI, what we build is something users will later need a separate app to block. The very existence of apps like Freedom reveals a structure where the front-end industry creates a problem with its own hands, and the back-end industry sells the solution back for a fee.

And right now, connectivity is only getting denser. Most product narratives for AI assistants center on being “always present, always aware of context, always initiating conversation.” That runs directly against what this paper implies. To be clear, it would be premature to say AI assistants are the same thing as notification bombardment. But I do want to point out that right now, we’re only asking “how much more constantly present can this be?” — and almost never asking “when should this know to disappear?” Knowing how to vanish is a product feature too.

Why you shouldn’t take this study at face value

Let me be honest about the limitations too.

The sample is unusual. iPhone users only, drawn from an online labor pool called Prolific, average age 32, 63% women. And 83% of them already wanted to cut back on their phones. Put differently — nobody knows whether the same effect holds for people who have no desire to reduce their usage at all.

Expectation effects can’t be ruled out. Participants knew this was a study about “how smartphones affect wellbeing.” There may have been pressure to report feeling better. The research team acknowledged this limitation themselves, which is precisely why they included gradCPT instead of relying only on self-report. But interestingly, the objectively measured outcome — attention — had the smallest effect (d = 0.23). The pattern that self-reported measures showed larger effects is worth noting.

There’s also contradicting evidence. Orben and Przybylski, using large-scale time-diary data, reported that screen time explains only about 0.4% of the variance in adolescent wellbeing. Granted, that analysis has itself been criticized for “analytical choices that shrank the effect,” so it’s hard to declare either side simply correct. But the gap itself — correlational studies showing small effects, experimental studies showing large ones — is itself a signal that we still don’t have a good handle on measuring this phenomenon.

Even accounting for all three of these caveats, something remains. A 25.5% compliance rate isn’t explained away by expectation effects or sample bias. If anything, the sample bias makes the number more alarming, not less — because even a group that was hand-picked for motivation still saw 3 out of 4 people fail.

Closing

Let me wrap up.

Cutting phone internet for 2 weeks genuinely improves wellbeing, mental health, and attention. That said, be careful how you cite the effect-size comparisons.
The strongest explanatory pathway was the sense of self-control — but what participants actually did was hand self-control over to an app. A feeling of self-control turns out to be a product of environmental design, not willpower.
So the practical takeaway of this study isn’t “make a resolution.” It’s “design an environment where a resolution isn’t necessary.” For what it’s worth, the research team itself doesn’t think a total block is required — the ITT analysis⁶, which includes everyone regardless of whether they actually maintained the block, still showed an effect. Simply reducing use may be enough.

One note on the Korea context. In the 2025 Survey on Smartphone Overdependence, the overall at-risk group⁷ fell to 22.7%, its 5th consecutive year of decline — but among teenagers alone, it rose to 43.0%. When everyone else is improving and only one group is moving in the opposite direction, that’s less likely to be a matter of that group’s willpower, and more likely a matter of the environment that group is stuck in. This is exactly the kind of situation this paper’s framing is worth applying to.

Here’s something you can try right now. Instead of deleting an app or setting a time limit, create just one situation where there’s simply no opportunity to exercise willpower at all. Not bringing your phone into the bedroom. Turning off data during your commute. The point isn’t “deciding not to look.” It’s designing a state where looking isn’t possible. The gap in success rates between those two is the entire finding of this paper.

If you’ve ever tried to cut back on phone use and failed, tell me in the comments exactly where it broke down. Was it a notification? A habitual reach for the phone? Something you got anxious not checking? That’s precisely the question this paper couldn’t answer — so your experience might just become material for a future issue.

💬 Where did it break down, the last time you tried to quit your phone? · 📨 Share this with someone it might help

(Web version) 📬 Subscribe · Every week, I write a newsletter that cross-analyzes technology, economics, and the humanities.

References & Further Reading

Primary sources

Castelo, N., Kushlev, K., Ward, A. F., Esterman, M., & Reiner, P. B. (2025). “Blocking mobile internet on smartphones improves sustained attention, mental health, and subjective well-being.” PNAS Nexus, 4(2), pgaf017. Read the paper ··· This is the core source behind today’s issue. I’d recommend looking at the CONSORT diagram (Fig. 3) in “Materials and methods” before the results section — the entire journey from 467 to 119 fits on that one figure.
The study’s raw data and pre-registration: OSF repository ··· You can check for yourself exactly where the pre-registration and the actual analysis diverge. The paper itself notes that the factor analysis of time use was a post hoc analysis, not a pre-registered one.

Background

Kirsch, I. et al. (2008). “Initial severity and antidepressant benefits: a meta-analysis of data submitted to the Food and Drug Administration.” PLoS Medicine, 5(2), e45. Read the paper ··· The source of the “bigger effect than antidepressants” comparison. Worth knowing this paper itself sits at the center of an ongoing dispute.
Orben, A., & Przybylski, A. K. (2019). “Screens, Teens, and Psychological Well-Being: Evidence From Three Time-Use-Diary Studies.” Psychological Science, 30(5). Read the paper ··· The leading study on the opposing side. Reading it alongside the paper above leads to a conclusion of “we still don’t really know” — and I think that’s the honest place to land.
Ministry of Science and ICT / National Information Society Agency, “2025 Survey on Smartphone Overdependence” (released March 26, 2026) Read the press release ··· The raw source behind the trend of overall improvement alongside teen-only deterioration.

📝 Glossary

The author, Kwangseob Ahn, is a professor of business administration at Sejong University and lead consultant at OBF (Oswarld Boutique Consulting Firm). He teaches statistics and data analysis, including business data management and business analytics, while leading GTM and AI strategy consulting in the field, designing the seam between technology and business. He has published academic research on a memory architecture for AI dialogue systems (HEMA) and runs Daily Arxiv, a daily curation of global AI papers. He holds a master's from Korea University's Graduate School of Technology Management and a KMBA. He is the author of Homo Brainless: The People Who Outsource Their Thinking.

Footnotes

PNAS Nexus: A sister journal of the Proceedings of the National Academy of Sciences (PNAS). Founded in 2022 as an open-access journal, meaning you can read this paper’s full text for free. ↩
Effect size (Cohen’s d): A standardized measure of how large a pre/post difference is. By convention, 0.2 is considered small, 0.5 medium, and 0.8 large — though these thresholds are somewhat arbitrary, and a “0.5” can mean very different things across different fields. ↩
gradCPT (gradual-onset continuous performance task): A 5-minute task where you press the spacebar when a city photo appears and withhold pressing when a mountain photo appears. Since cities appear 90% of the time and mountains only 10%, pressing becomes habitual — and a brief lapse in attention immediately produces an error. It’s widely used to measure attention objectively, without relying on self-report. ↩
Meta-analysis: A method for statistically pooling multiple studies on the same topic. Generally considered more reliable than any single study, but which studies get included or excluded can shift the conclusion substantially. ↩
Mediation analysis: A statistical method that goes beyond “A improved B” to estimate a pathway like “A changed C, and C improved B.” Mediation analysis cannot, however, prove causation — a limitation the paper itself acknowledges. ↩
ITT / TOT: ITT (intention-to-treat) analyzes everyone, including those who didn’t follow instructions; TOT (treatment-on-treated) analyzes only those who actually complied. ITT is more conservative and more realistic. The paper’s primary results are reported on an ITT basis. ↩
Smartphone overdependence at-risk group: A combined category of high-risk and potential-risk groups, measured using a standardized scale. Worth keeping in mind this is a survey-instrument threshold, not a clinical diagnosis. ↩