A sensible way to judge an AI subscription is to ask how many hours it gives back.
Then comes the trap.
A task that used to take three hours now takes thirty minutes. You recovered two and a half hours. The obvious move is to finish two and a half hours earlier.
Instead, the brain may say:
“Two and a half hours free. Great. I can fit five more things in.”
Suddenly it is night.
The AI does not get tired. Replies arrive quickly. There is almost no waiting between tasks. And, annoyingly, the work can be genuinely fun.
At that point AI stops behaving like a time-saving tool and starts behaving like an infinite-quest game for work.
The explanation is bigger than “I paid for it, so I should get my money’s worth.” Flow, rebound effects, disappearing stopping points, self-directed overwork, unfinished tasks, curiosity and AI self-efficacy fit the pattern better.
1. The math first: 40% less time means 67% more capacity
In a well-known experiment with 453 professionals doing writing tasks, ChatGPT reduced average completion time by 40% while raising output quality by 18%.[1]
A 40% reduction in time is not merely “40% faster.”
If ten hours of work becomes six hours, then keeping the same ten-hour workday gives:
10 / 6 = 1.667 times the capacity.
That is about 66.7% more work.
So the equation
40% time saving → 67% more work
is mathematically plausible before psychology enters the room.
The real question is why the remaining four hours are so often filled instead of reclaimed.
2. The finish line disappears
Before AI, friction itself often stopped us.
“I could improve this too, but it would take another thirty minutes. Tomorrow.”
With AI, that becomes:
“Could probably fix it in two minutes.”
Individual tasks end constantly, but higher-level goals such as “make it better,” “research it more,” and “automate the rest” have no natural endpoint.
So the problem is not that tasks never finish.
Tasks finish at ridiculous speed, and the next task spawns instantly.
It is the office version of completing a main quest and watching five side quests appear on the map.
3. AI work resembles the conditions that support flow in games
Flow research emphasizes conditions such as clear goals, clear feedback, and a reasonable match between challenge and skill. Flow can involve deep concentration, altered time perception and an activity that feels rewarding in itself.[2]
Games are useful flow environments because they supply these conditions naturally.
AI-assisted work often does too:
- a clear micro-goal: fix this bug;
- immediate feedback after a prompt;
- rapid iteration;
- the ability to break hard problems into smaller ones;
- endless escalation when a task becomes too easy;
- real outputs at the end.
The loop becomes:
action → result → adjustment → result → next quest.
And unlike a normal game, the real world levels up too. Three hours of gaming advances a save file. Three hours with AI may leave behind code, research, writing, plans and automation.
It feels a little like entertainment while also increasing real-world stats.
That is an unusually powerful combination.
There is not yet strong causal evidence that generative AI itself induces flow and therefore causes long work hours. This is a structural analogy supported by established flow research. But the overlap in conditions is hard to miss.
4. Waiting time vanished, and some accidental recovery vanished with it
Older knowledge work contained strange pockets of forced idleness: builds, searches, replies from colleagues, data processing, scheduling and overnight approvals.
From an efficiency perspective, those pauses looked wasteful.
From a human perspective, they were also accidental idle time.
A meta-analysis of micro-breaks found small but significant benefits for vigor and fatigue. For cognitively demanding work, recovery of performance may require breaks longer than ten minutes.[3]
AI removes much of the friction. The moment one answer arrives, the next task can begin.
So AI may compress not only work time but also the natural stopping points embedded inside work.
Other fields study a related idea as the time-use rebound effect: time saved by efficient technology is reallocated into additional activity instead of remaining free.[4]
The day does not become empty. It becomes denser.
5. The price of “one more thing” collapses
Humans do not only procrastinate. Sometimes we do the opposite: we hurry to complete a subgoal even when that requires extra effort. This tendency has been called pre-crastination. In a classic experiment, people sometimes chose a bucket closer to the start even though they then had to carry it farther.[5]
Before AI:
“One more task costs thirty minutes.”
That cost can stop us.
After AI:
“One more task? Two minutes.”
Then another two-minute task appears.
The boss monster is not difficult. The slimes just respawn forever.
6. Self-chosen work is harder to resist: the autonomy paradox
Technology has produced a similar trap before.
Research on mobile email found that knowledge professionals experienced “work anywhere, anytime” as freedom and control, yet collectively drifted toward working everywhere and all the time. The authors called this the autonomy paradox.[6]
AI can create a comparable experience.
If a manager assigns five more tasks, it feels like workload expansion.
If you discover them yourself—
“this could be improved,”
“let’s test every pattern,”
“this is interesting, automate it too”—
the extra labor feels like exploration and choice.
The work expands, but the interface still feels like maximum freedom.
7. More open quests make it harder to log out mentally
AI is not only good at solving tasks. It is also good at discovering new ones.
Research A, find problem B. Fix B, discover idea C. Build C, notice automation D. Build D, now audit E.
The quest generator ships with the quest solver.
A 2026 meta-analysis found that unfinished work tasks were associated with more work-related thoughts during off-job time, especially ruminative thinking.[7]
That makes “I will sleep after everything is finished” a terrible rule for AI-heavy work.
Everything will never be finished.
You do not need full completion. You need a save point.
8. Why do many gamers still barely use AI?
Because “gamer” is too broad a category.
AI gives you no quest list. It mostly asks: what do you want to do?
Heavy use therefore requires a different habit:
- translating real problems into AI-sized tasks;
- decomposing vague goals;
- generating the next hypothesis from the output;
- treating a bad answer as debugging material rather than proof that AI is useless;
- exploring what the tool itself can do.
A 2025 study found that higher AI self-efficacy was associated with more AI use, while exploratory curiosity was linked to use through self-efficacy.[8]
Other work on continued ChatGPT use also points to enjoyment and playfulness, not just utility, as important factors.[9]
So the especially strong match may be:
exploration + experimentation + systems optimization + growing AI self-efficacy.
Think less “person who enjoys games” and more “person who builds machines in Minecraft, optimizes factories in Factorio, and is tempted to edit the strategy wiki afterward.”
AI is a sandbox with effectively unlimited content.
The save file is reality.
9. More AI does not inevitably mean more work
There is an important counterexample.
A field experiment across 66 firms and 7,137 knowledge workers found that active users of an integrated generative AI tool spent roughly two fewer hours per week on email in the second half of the experiment and reduced work outside regular hours. The researchers did not detect an increase in the quantity of tasks from individual AI access.[10]
So:
AI → overwork
is not a law.
If the saved time is actually reclaimed, AI can end the day earlier.
A separate study of 2,896 employees across 141 organizations testing a pay-preserving four-day week found improvements in burnout, job satisfaction, mental health and physical health. Reduced sleep problems, lower fatigue and better self-rated work ability helped explain the gains.[11]
Efficiency can be converted into less work time. It does not have to become more output.
10. The better operating model: use AI as a time compressor, not a work expander
The key question is not how much AI can produce.
It is when you declare victory.
Rule 1: define the win condition before opening AI
Decide: A, B and C make today successful.
If AI finishes them three hours early, do not automatically promote D, E and F into today’s main quests.
Those three hours are the profit.
Rule 2: queue new quests instead of accepting them immediately
AI will generate ideas. Save them under NEXT QUESTS.
Do not mix tomorrow’s discovery with today’s victory condition.
Rule 3: use If–Then rules to stop, not only to start
Examples:
“If A, B and C are done, I do not start new work.”
“If an idea appears after shutdown time, I capture it but do not execute it.”
“If a task requires stealing from sleep, it moves to tomorrow.”
A meta-analysis of 94 independent tests found that these implementation intentions had a medium-to-large positive effect on goal attainment.[12]
They can also be designed as stopping rules.
Rule 4: manually restore the breaks that automation removed
AI replies quickly. Human recovery did not receive the same upgrade.
Stand up, leave the screen, eat and take deliberate pauses.
Re-implement the old build wait and reply wait as intentional recovery.
Rule 5: change the KPI
A bad KPI is:
How many tasks did AI help me process?
A better KPI is:
How many human hours did I need to achieve the required result at the required quality?
If ten required tasks used to take eight human hours and now take four, that is a win.
If AI lets you complete twenty-five tasks while still working eight hours, capacity improved, but free time did not.
11. A “forced save” protocol for AI power users
If we keep the game metaphor, the operating system can be simple.
Before play
- choose no more than three main quests;
- create a separate side-quest queue.
During play
- let AI search, generate, compare and implement;
- keep human responsibility for purpose, verification and acceptance;
- send every new discovery to the side-quest queue.
When the win condition is reached
- ban new quest acceptance;
- record only current state and the next action for unfinished work;
- save.
After shutdown
- ideas may be captured;
- execution waits for the next session.
This does not kill curiosity.
It lets exploration continue tomorrow while allowing today’s session to end.
Conclusion: AI has no stamina bar. Humans do.
Faster AI is a genuine advantage.
The mistake is converting every minute saved into another minute of work before the human ever receives the benefit.
A 40% time reduction can mathematically support roughly 67% more work in the same window. But capacity is not an obligation.
AI makes micro-goals cheap, feedback immediate, waiting scarce and new paths abundant. For an exploration-oriented person, that can become a brilliant infinite game.
So the answer is not to make AI weaker.
The human needs to implement the end screen.
Use AI as a machine that turns eight hours of work into three hours and lets you log out—not merely as a machine that stuffs three times as much work into eight hours.
The real return on AI is not how many hours the model worked.
It is how many hours the human actually got back.
References (12)
- Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192 doi.org
- Harris, D. J., Allen, K. L., Vine, S. J., & Wilson, M. R. (2021). A systematic review and meta-analysis of the relationship between flow states and performance. International Review of Sport and Exercise Psychology doi.org
- Albulescu, P., Macsinga, I., Rusu, A., Sulea, C., Bodnaru, A., & Tulbure, B. T. (2022). “Give me a break!” A systematic review and meta-analysis on the efficacy of micro-breaks for increasing well-being and performance. PLOS ONE, 17(8), e0272460 doi.org
- Time rebound effect in households’ energy use: Theory and evidence (2022). Cleaner and Responsible Consumption, 5, 100066 doi.org
- Rosenbaum, D. A., Gong, L., & Potts, C. A. (2014). Pre-crastination: hastening subgoal completion at the expense of extra physical effort. Psychological Science, 25(7), 1487–1496 doi.org
- Mazmanian, M., Orlikowski, W. J., & Yates, J. (2013). The Autonomy Paradox: The Implications of Mobile Email Devices for Knowledge Professionals. Organization Science, 24(5), 1337–1357 doi.org
- Wendsche, J., Weigelt, O., & Syrek, C. J. (2026). Unfinished work tasks and work-related thoughts during off-job time: meta-analysis of the Zeigarnik effect in a work-recovery context. Anxiety, Stress, & Coping, 39(4), 385–407 doi.org
- Schutte, N. S., & Li, H. (2025). The role of self-efficacy and curiosity in student use of artificial intelligence (AI). International Journal of Educational Technology in Higher Education, 22, 73 doi.org
- Kim, J. S., & Baek, T. H. (2025). Motivational determinants of continuance usage intention for generative AI: an investment model approach for ChatGPT users in the United States. Behaviour & Information Technology, 44(12), 3080–3096
- Dillon, E. W., Jaffe, S., Immorlica, N., & Stanton, C. T. (2025, revised 2025). Shifting Work Patterns with Generative AI. NBER Working Paper 33795 doi.org
- Fan, W., Schor, J. B., Kelly, O., & Gu, G. (2025). Work time reduction via a 4-day workweek finds improvements in workers’ well-being. Nature Human Behaviour doi.org
- Gollwitzer, P. M., & Sheeran, P. (2006). Implementation Intentions and Goal Achievement: A Meta-analysis of Effects and Processes. Advances in Experimental Social Psychology, 38, 69–119. )38002-1 doi.org
