Four rules, and they have to be checkable rather than aspirational. One device, never the phone in bed. One window with a hard end. One person who knows the rules. And a monthly displacement check, written down. "I will use it less" is not a rule, because on a bad night there is no way to tell whether you kept it.
Why a ceiling is a defensible position
Nobody has evidence that you must stop entirely, and this page is not going to pretend otherwise.
What the evidence shows is a pattern at the heavy end. The four-week randomised study of 981 people found "no significant effects were detected from experimental conditions", and that "participants who voluntarily used the chatbot more, regardless of assigned condition, showed consistently worse outcomes". The platform-scale study of over three million conversations found that "very high usage correlates with increased self-reported indicators of dependence".
Both are correlations rather than causes, and both sit at the very heavy end of use. The honest thing that follows from that is a ceiling, not a ban. Stay well away from the end of the distribution where the association lives, and check periodically that you still are.
There is a second reason for a ceiling, which is that bans have a poor completion rate. If a total stop has failed for you twice, a workable limit that holds is worth more than a perfect rule that lasts nine days.
The four rules
**One: it lives on one device, and not the one in your bed.** This is the highest-value rule on the list, because most heavy use happens lying down in the dark. Laptop only is the usual version. If it has to be on the phone, the phone charges in another room from a fixed time, which is also the one instruction the UK's Chief Medical Officers give without hedging: "Leave phones outside the bedroom when it is bedtime."
**Two: one window, with a hard end.** A time, not a duration. "Eight to nine" survives a bad evening in a way that "about an hour" does not, because at eleven o'clock you can tell whether you kept it. Set an alarm for the end, on something other than the device you are using.
**Three: one person knows the rules.** Not for supervision. For the fact of another human knowing, which does more work than blocks do. Tell them the two rules above and ask them to mention it once a month.
**Four: a monthly check, in writing.** The rules are the easy part. The check is what stops slow drift, and it is the section below.
The monthly check
Ten minutes, on paper, on the same date each month. Answer four questions.
- **What have I stopped doing since last month?** Not what I skipped once. What has quietly stopped.
- **Have I broken the window, and how many times?** A count, not an impression.
- **Has anyone said anything?** Even a joke. People notice earlier than you do and they mention it lightly first.
- **Where does it fall in a bad week?** Anyone can keep rules in a good week. The relevant data comes from the week when you were ill, or dumped, or working late.
If the answers are steady across three checks, your ceiling is real. If they slide in the same direction each month, it is not, and that is worth knowing before it costs you a year.
Keep the sheets. Memory is a poor instrument here and a page from June is not.
The uses that stay good
Worth naming, because a page like this can otherwise read as though every use is a problem.
Drafting something you are stuck on. Rehearsing a hard conversation before you have it, on the condition that you then have it. Working out what you actually think by writing it down. Company at a specific hour, when it is one of several things filling that hour rather than the only one.
The dividing line is not the activity. It is whether it is one route among several or the only route you have left. Something that supplements your week is doing a job. Something that is the whole week has become the week. [What the chatbot is giving you](/guides/ai_chatbot/what-the-chatbot-is-giving-you) is the exercise for telling which is which.
The signs your rules are not holding
- **The window has moved twice** and you renegotiated both times rather than kept it.
- **It is back on your phone,** and there was a sensible reason at the time.
- **You have stopped telling the person.** This is the earliest reliable signal and it is almost always the first one to appear.
- **You check it before speaking to anyone in the morning.**
- **You have started keeping it to yourself in conversation,** including in small ways.
- **A day without it feels like a loss** rather than an ordinary day.
Two or three of those means tighten one rule, not add four. If tightening has failed twice, stopping fully is usually easier than negotiating with yourself nightly, and [the first week without it](/guides/ai_chatbot/the-first-week-without-it) sets out what to expect.
The one rule not to bend
Never let it be the only thing that knows.
Whatever your limits are, keep one human who gets the real version of your week. Not the highlights, the actual state of it. A companion that agrees with everything cannot tell you that you have said the same thing every Sunday for a year, and that is exactly the sort of thing a person can tell you and nothing else can. [Why a companion chatbot never says no](/guides/ai_chatbot/why-a-companion-chatbot-never-says-no) is the mechanism behind that.
If the honest answer is that there is no such person right now, that is the actual project, and it is more useful than any rule on this page.
Common questions
**How much is too much?** Nobody can give you a number, and there is no threshold in any manual. The measure that works is displacement: what has stopped happening in your life. Two things gone is your answer whatever the hours say.
**Can I keep a ceiling if I have been a very heavy user?** Some people do. It is harder, and the useful test is empirical rather than moral. Set the ceiling, run three monthly checks, and read the sheets. If it has held, it has held.
**What if I break the window once?** Then you broke it once. Note the day and what preceded it, keep the rule, and do not respond by writing a stricter one. Rules made in the aftermath of a bad night are always too harsh to keep.
**Is it worse to use it every day for twenty minutes or once a week for three hours?** Nobody knows, and anyone claiming otherwise is guessing. The daily version has the higher chance of quietly becoming the shape of your evening, which is the risk worth watching.
**Do I need to tell people I still use it?** No, and one person knowing is not the same as announcing it. The point of that rule is not disclosure. It is that a private arrangement has exactly one point of failure.
Sources
- [Longitudinal randomised study of extended chatbot use, four weeks, 981 participants (arXiv, 2025)](https://arxiv.org/abs/2503.17473). The null experimental result, and that heavier voluntary use went with consistently worse outcomes.
- [Investigating Affective Use and Emotional Well-being on ChatGPT (arXiv, 2025)](https://arxiv.org/abs/2504.03888). Over three million conversations, and that very high usage correlates with self-reported indicators of dependence.
- [UK Chief Medical Officers' commentary on screen time and social media](https://www.gov.uk/government/publications/uk-cmo-commentary-on-screen-time-and-social-media-map-of-reviews). The precautionary position without a causal claim, and leaving phones outside the bedroom.
- [Computational analysis of a 27,000-member Reddit community of people in relationships with chatbots (arXiv, 2025)](https://arxiv.org/html/2509.11391v1). Reported benefits alongside 9.5% acknowledging emotional dependency, with the selection-bias caveat.
