'Enshittified at a Surprising Clip': A Week of Hacker News on AI Coding Tools

Some fortnights the complaints about AI come from people who barely use it. This one they came from the people who use it most. Scroll Hacker News over the past week — the forum where developers argue about their tools in unusual detail — and the grievances about AI coding assistants weren’t existential. Nobody was worried about the robots waking up. They were worried about their bill, their UI, and the effort of reading what the model just wrote.

Quotes sourced from: Hacker News. Every quote below was located at its comment permalink and reproduced verbatim; each is listed with its username, the platform, and the date in the Sources section. As always, we quote experiences, not verdicts — a forum comment is one practitioner’s account, often mid-argument, and we’ve framed them as exactly that. What makes this batch worth reading isn’t volume; it’s specificity. These are checkable complaints.

“Enshittified at a surprising clip”: the dark-pattern gripe

The sharpest thread of the fortnight was about Cursor, the AI code editor, and it wasn’t about the quality of its completions. It was about the way the product behaves around you. A user posting as jmuguy, on 20 August, laid out a bill of particulars that will sound familiar to anyone who’s watched a beloved tool curdle:

“Cursor isn’t covering itself in glory regardless. The flagship app is getting enshittified at a surprising clip. It constantly pops up and interrupts your work pushing new features, changes your model to whatever the latest Grok is without prompting, has this mystery meat UI that is constantly changing, pushes cloud agents in ways that are definitely designed to trick you. We’re actively looking at alternatives, I wouldn’t touch anything this company produces from here on out.”

Set aside the verdict at the end — that’s one person’s conclusion, not ours — and look at the specifics, because they’re the kind you can check: interruptions pushing new features, a model silently swapped to Grok, a UI that keeps moving. The transparency version of this complaint is one we keep returning to, most recently when users said newer AI felt like a downgrade. Here it’s sharpened into a word — enshittification — that has become the developer shorthand for a product optimised for the company’s metrics rather than the user’s work.

The same week, on a thread about Cursor’s new GitHub-style features, nikolay reached the end of the road for a different reason — a feature gated behind payment where he expected an open one:

“…this is reminding me to cancel my Cursor subscription. I am paying for it but not using it because I don’t need it as it offers me nothing.”

“I am paying for it but not using it” is the quiet churn every subscription business fears and the one that the flat-fee-to-meter drift keeps generating: not an angry cancellation over a scandal, just a slow realisation that the value stopped justifying the direct debit.

The billing you can’t see

If the Cursor complaints were about attention, the next set were about money — specifically, the widening gap between what these tools charge and what a user can actually see. The most concrete came from j0selit0, who’d been poking at GitHub Copilot’s network traffic and noticed something odd about how premium requests get counted:

“While looking at the traffic I noticed Copilot’s own LLM requests used X-initiator: agent and didn’t consume premium quota. That made me curious what would happen if I changed my own requests from user to agent. Turns out the model still responded, but those requests didn’t decrement my premium quota or show up in the billing analytics.”

That’s one developer’s own testing, not an audit, and we present it as such — but it’s a precise, reproducible claim: a billing system keying off a header the client controls, such that identical requests count against your quota or not depending on a label. When the meter is that easy to confuse, the reasonable question isn’t “how do I exploit this” but “how much can I trust the number the meter shows me at all?”

It rhymes with a second money story doing the rounds: a reported Codex bug on AWS Bedrock said to run up charges ten times higher than expected. Commenting on it, palmotea aimed at the incentive rather than the bug:

“Except leadership doesn’t care about dumb bugs. Never has, never will (until it’s too late). It cares about velocity and cost. They’d totally replace all software development with worse AI software development in a heartbeat.”

Whatever you make of the cynicism, the pairing is the point: a quota that undercounts and a bill that overcounts, both landing in the same week, both invisible until someone technical went looking. Neither claim is dispositive on its own, and we’re not treating a forum post as a filed complaint. But the shape is what unsettles: as these products bolt on agents, cloud runs and per-request pricing, the meter grows more complicated at exactly the moment it grows less legible — and the person paying is the last one equipped to check the maths.

“A clueless machine” reviewing your code

A related irritation surfaced around the AI code-review features now bolted onto everything. hypfer, on 21 August, was blunt about GitHub’s:

“GitHub’s copilot review feature is an equally miserable experience. That one loves talking in imperatives, regardless of the fact that it is a clueless machine.”

The complaint isn’t that automated review is worthless; it’s the mismatch between the confidence of the tone and the reliability of the content — a machine issuing orders about code it doesn’t understand. That gap between register and competence is the same one that makes the 10x-developer promise so slippery: the output sounds authoritative whether or not it’s right.

The verbosity tax

Then there was the recurring gripe about Claude’s output — not that it’s wrong, but that it’s exhausting. It surfaced under, of all things, a “Show HN” for a tool whose entire job is to clean up Claude 5’s token output with a second model. cnity diagnosed it more precisely than the usual “too wordy”:

“most people seem to consider the core problem of Claude’s output as ‘too verbose’ but I don’t think this actually cuts to the heart of the matter at all. It’s almost, in some weird way, the opposite: like the text is far too dense. It tries too hard to invent odd terminology to try to condense stuff, but it doesn’t tell you up front that it is going to call your company wide error-handling mechanism a ‘flare’…”

dcolkitt had resorted to a coping mechanism that says a lot on its own — a running side-conversation just to decode the tool’s house dialect:

“I frequently use Claude Code and often find the phrasing and language to be hard to understand. I’ve noticed it’s largely broken down into frequently used ‘Claude-isms’.”

When users are building tools and keeping glossaries just to make a product’s output legible, the “productivity” being sold is quietly being taxed back in comprehension effort. It’s the flip side of the regressions we heard about when a flagship upgrade landed badly: the model may be more capable and still be more work to actually use.

“Overselling it so much”

Underneath the specific gripes ran a broader fatigue with the gap between the pitch and the product. germandiago, on 21 August, put the sceptic’s case in its strongest form:

“I think they have been overselling it so much… Noone wants middle term maintenance nightmares or unreliable code. Can look good the first month, that’s it.”

That’s a strong opinion, not a measurement, and plenty on the same site would argue the opposite. But “looks good the first month” is a useful antidote to demo-driven enthusiasm: the tools that impress in a five-minute trial and the tools you can live with for a year are not always the same tools, and the people filing these complaints are the ones a year in.

Step back and the complaints rhyme. A model switched without asking, a quota that skips a labelled request, a bill that runs 10x, output you need a second tool to parse — each is a small transfer of control away from the user, made without notice. None is a catastrophe on its own. Together they sketch a class of product that keeps making unilateral decisions about your money, your model and your attention, and expects you to keep up. That texture — not any single scandal — is what these threads are really documenting.

The moan of the day

It goes to jmuguy, for the phrase that named the whole mood and will be doing overtime in developer Slacks for months:

What to take from it, fairly

The caveats are real and we’ll state them plainly. Hacker News is a self-selecting crowd of technical sceptics; the contented majority rarely posts; one developer’s packet capture is not a billing audit, and one person’s “overselling” is another’s Tuesday. None of these comments proves a company acted in bad faith.

But notice what kind of complaints these are. They aren’t vibes about AI being scary. They’re specific, checkable observations about products: a model swapped without asking, a quota that skips a header-labelled request, a bill that ran 10x, output dense enough to need a translator. Specificity is the tell that separates a grumble from a signal. And the fixes they imply are boringly reasonable:

• Stop switching things silently — don’t change my model, to Grok or anything else, without telling me and letting me decline.
• Show an honest meter — the usage counter I can see should match the one you actually bill against.
• Let me read the output — concise by default, house jargon defined, no second tool needed to parse the first.
• Announce the change — a one-line release note beats discovering a moved button mid-task.

As with tools that act without asking, the ask underneath is just legibility: tell me what you’re doing, charge me what you said, and don’t make me fight the thing I’m paying for. None of it is a revolution. It’s the baseline a paying user is entitled to expect, and the reason these threads keep filling up is that it keeps not being met.