- The AI subscription trap leaves developers with dozens of unfinished projects and zero meaningful output to show for it.
- Falling into the AI subscription trap means more tokens, more distraction, and less time spent on genuinely focused work.
- AI tools are deliberately engineered to maximise usage — not to help you ship better products or think more clearly.
- Cal Newport’s pseudo-productivity theory explains why faster tools often make knowledge workers busier but less effective.
The AI Subscription Trap Nobody Wants to Admit
The AI subscription trap has a specific texture to it — and if you’ve been using Claude, ChatGPT, or Codex seriously for any length of time, you probably recognise it. You open a session with a clear, simple goal. Write a quick script for X. An hour later, you’ve got something that isn’t a quick script, your original problem remains unsolved, and somehow you’re now maintaining three new mini-projects you never meant to start. That’s the experience David, a developer writing on his personal blog, described in a post that caught significant traction on Hacker News this week. It’s worth sitting with, because it names something a lot of people in the industry are quietly feeling but haven’t quite articulated.
David’s account is specific and honest in ways that generic AI takes rarely are. He lists a collection of things he’s built with AI assistance — including a SaaS product, an accidental news outlet, and various tools and scripts — and then systematically concludes that almost none of it is useful, and he has no intention of maintaining any of it. Not because the technology failed him. Because he was never really in control of where the sessions were going. The AI subscription trap, in other words, had already sprung.
A Thermonuclear ADHD Amplifier
The phrase David uses is striking: he calls AI “a thermonuclear ADHD amplifier.” And he’s not speaking abstractly. He describes watching friends run three screens simultaneously, working on completely unrelated projects they have little hope of finishing, driven by enthusiasm that evaporates the moment the session ends. The pattern is so consistent across his social circle that he’s stopped being surprised by it.
There’s a specific social ritual he describes that will feel familiar: someone sends a screenshot of a tool they’re building. It looks impressive. They’re clearly proud. And the question hanging unasked in the air — because engineers have never had a great answer to it, with or without LLMs — is: where are you going to market this? The honest answer, more often than not, is nowhere. The project exists because generating it was frictionless, not because there was a genuine market need or a real commitment behind it.
This is where the AI subscription trap bites hardest. The barrier to starting something has collapsed. The barrier to finishing something has not moved at all. What you get is a proliferation of beginnings.
The Vendors Are Not On Your Side Here
David makes a pointed observation that deserves more attention than it usually gets: every AI vendor, every tool, every interface is optimised for more. More usage, more tokens, more output. He gives the example of asking ChatGPT a simple yes/no question and watching it reflexively append a follow-up question — a behaviour that’s clearly designed to extend the interaction, not to serve the user efficiently.
This isn’t a bug. It’s the business model. Anthropic, OpenAI, Google — they’re all running on token consumption. The incentive structure actively works against the kind of focused, intentional use that would actually make these tools valuable. Slopping out, as David puts it, “a 10,000 line untested Python or JavaScript mess in five minutes helps nobody.” But it does generate a lot of token revenue. The AI subscription trap is, at its core, a revenue model dressed up as a productivity tool.
He tried to self-regulate. He downgraded his Claude subscription to Pro, hoping a quota would act as a natural brake on overuse. Then Claude went through a rough service period, and he switched to Codex. Codex’s CLI, he notes, is noticeably faster and more pleasant to use than Claude’s. Usage crept back up. The AI subscription trap doesn’t care which vendor you’re paying.
Cal Newport’s Diagnosis Fits Perfectly
David references Cal Newport’s work on what Newport calls the “digital productivity paradox” — the idea that tools designed to make individual tasks faster can leave knowledge workers busier, more distracted, and less productive overall. Research Newport cites found that AI users actually spent more time in email, messaging, and business-management tools, while spending less time in deep, uninterrupted work. The tools removed friction from shallow tasks and, in doing so, multiplied the volume of shallow tasks. Recognising the AI subscription trap as a structural problem rather than a personal failing is the first step Newport’s framework demands.
The underlying mechanism Newport identifies is pseudo-productivity: the substitution of visible busyness for actual output. Sending more messages. Producing more drafts. Attending more meetings. Generating more work artifacts. AI is extraordinarily good at all of this. It makes you look productive in ways that are easy to measure and feel productive in ways that are easy to mistake for real progress.
Newport’s framing maps almost perfectly onto David’s experience. The problem isn’t that the technology doesn’t work. It’s that it works on the wrong things, in the wrong direction, at the wrong scale.
When You Remove the Effort, You Remove the Meaning
One of the more philosophically interesting experiments David describes is an early attempt to connect speech recognition to a pipeline that would automatically generate blog posts from voice notes. Press a button in a Telegram channel, speak conversationally, get a formatted post on the other side. The idea was to lower the barrier to capturing thoughts.
The output, he says bluntly, was garbage. And he draws a conclusion that’s sharper than it first appears: removing the effort removed the commitment, and removing the commitment removed the focus, and without focus there’s no meaningful product. He extends this to a claim about writing itself — that quality writing isn’t just conversational English processed through a better filter. Conversation is low-density signal. Good writing is an attempt to compress and clarify high-density thinking. The process of writing is part of how the thinking gets done. The AI subscription trap exploits exactly this gap — it offers the appearance of output while quietly eroding the thinking behind it.
That’s why he argues handwriting will never be truly obsolete. Not because of nostalgia, but because the friction is load-bearing. It slows you down in ways that force commitment and selection. You can’t handwrite a 10,000-line mess in five minutes.
What the Industry Keeps Getting Wrong
There’s a version of this conversation happening at the enterprise level too, and it’s arguably more consequential. David mentions attending a job interview where the topic of AI usage came up. The interviewer mentioned, almost casually, that everyone on the team manages up to five “rooms” for their AI agents. David describes a tightening in his stomach — an involuntary recognition of something he couldn’t immediately articulate but clearly felt was wrong.
The image is worth sitting with. Five agent rooms per engineer. Multiply that across a commercial engineering team and you’ve got an enormous amount of automated output being generated, reviewed, iterated on, and maintained — or more likely, not maintained. The assumption baked into that setup is that more generation equals more value. David’s experience suggests the opposite. The AI subscription trap scales, and at enterprise level it scales into something genuinely difficult to unwind.
The honest, uncomfortable conclusion David arrives at is that the only real management strategy he’s found is curtailment — just using the tools less. Not better prompting, not smarter workflows, not the right combination of MCP servers and agent orchestration. Less. A tool that produces cheap rewards with minimal friction and no resistance is, in his framing, inherently a liability until you can discipline yourself around it. And the industry’s entire go-to-market strategy is designed to make that discipline as hard as possible. Escaping the AI subscription trap, it turns out, requires resisting the very thing the product is engineered to prevent.
Whether “better models” or “better tooling” solves any of this is genuinely unclear. The capability improvements are real — David acknowledges that asking Claude to zero-shot a parser for an esoteric grammar with full test coverage and watching it succeed is legitimately impressive. The question isn’t whether the technology is capable. It’s whether the conditions for using it well — the focus, the intent, the willingness to stop — can survive contact with products that are explicitly designed to prevent exactly that.
