PS — Here, “psychosis” and “depression” describe a cycle of obsessive AI use, not clinical diagnoses. If you may be experiencing either, seek support from a mental health professional.
There is a very specific comedown I keep recognising after a common kind of AI obsession.
For a few weeks, everything rises. Day over day, we are doing more. Week over week, we are moving faster. Every conversation opens three more doors. Ideas turn into tasks, tasks turn into prototypes, and prototypes turn into entirely new projects before the old ones have had time to settle.
We cannot stop talking to the AI; it is the first tab we open and the last one we close. There is always one more prompt to send, one more branch to explore, one more thing that feels suddenly possible.
The fling
The beginning feels frictionless because so many old boundaries appear to have disappeared. Work that once took days takes hours. Questions that would have stalled a project get an immediate answer.
That speed is intoxicating. More importantly, it compounds. Excitement produces longer sessions. Longer sessions produce more visible output. More output produces more excitement. We begin to expect tomorrow to be faster than today because, for a while, it usually is.
The AI never looks tired. It never asks to stop. It does not bring yesterday’s mental fatigue into today’s conversation. We do, but that difference is easy to miss while the graph is still going up.
The acceleration
This stretch is the AI psychosis. This phase is a period whose momentum keeps validating itself.
Every response seems to prove that the project is working. Every new capability expands the scope. We stop measuring progress against the original goal and start measuring it against what else might now be possible.
The project grows faster than our ability to understand it. Threads multiply. Decisions remain half-made because another option appears before the previous one has been tested. We are producing at a pace that feels like clarity, even when much of what we are producing is unresolved surface area.
From inside the acceleration, slowing down feels irrational. Why would you stop when this is working so well?
The comedown
The shift is rarely dramatic. The AI does not suddenly become useless. We simply stop being able to follow the project at the speed at which we created it.
The easy surface area may have been covered. Now the work is integration, judgement, maintenance and choosing what not to build. The AI can still respond in seconds, but progress no longer happens in seconds. A fast answer can create more work than it removes. A new feature can make the project harder to hold in our heads. Fixes begin to uncover decisions that speed allowed us to postpone.
The AI is still moving. We are the part that can no longer keep accelerating.
There is also the speed at which the work can spread. Writing code may be easier, but the ecosystem around the problem still moves at its own pace. Governance, compliance, funding and other forces that neither we nor the AI control may not slow the code directly, but they still limit how widely the solution can take hold and how quickly its profits can grow.
Why I call it a depression
I call it a depression because the low is relative to the high that came before it. After weeks of feeling unusually capable and expansive, a return to ordinary attention, energy, boredom and doubt can feel like a collapse.
Often, nothing is broken. The tool still works, the project may still matter and the earlier pace may simply have been too intense to last.
Naming the pattern
I want a name for this cycle because an unnamed comedown can become a burden. We may mistake a slower phase for failure, abandon work that needs time to settle or chase another burst of novelty.
Naming it helps me see the comedown as a phase, not an identity. The high was not necessarily fake, the low is not necessarily permanent and a steadier rhythm can still follow.
A possible prevention
One possible prevention is to be more measured during the climb. The upward momentum invites us to keep prompting even in our downtime, but deliberately going slower gives us time to decide whether the project really needs to go higher.
That can mean spending more time thinking without the AI, learning the basics and testing our own understanding as the work grows. We can think longer about architecture, write evaluations before adding more features and strengthen the foundation before the fling devours the project.
A slower climb may extend the productive phase because more of its momentum is supported by understanding rather than novelty alone. It cannot remove the comedown, but it may delay its onset and leave us with a project we still feel able to steer.
When the comedown arrives, naming it gives us a choice. We can treat the feeling as proof that the project should be abandoned, or recognise it as a phase and decide deliberately whether to continue. Hopefully, building control and expertise on the way up makes continuing the more grounded option.
What comes after the magic
The next phase is less cinematic. It begins when speed stops being the measure of progress and all the accumulated decisions become visible. Judgement, which briefly seemed optional, returns to the centre of the work.
That may be the more productive relationship with AI: not the endless high, and not the rejection that follows it, but a rhythm that can survive contact with your actual limits.
The comedown is not necessarily where the work ends. It may be where the work stops feeling like magic and starts becoming yours again.