‘I’ll Simply Cancel’: A Week of AI Asking for Your ID and Your Data

Open your AI tool of choice this week and, before it does anything for you, it wants something from you first. Your age. Your consent to be training data. Your acceptance of a usage cap that moved while you weren’t looking. The work you actually came to do is still there, somewhere, behind the turnstile — but the turnstile keeps getting taller.

None of it is a scandal on its own. Age checks have legal pressure behind them, free tiers have always been paid for with data, and usage limits are how loss-making companies stop the bleeding. Put a week of it together, though, and a pattern shows up in the complaints: the deal is quietly getting worse in several directions at once, and the people paying for it noticed.

So here is the week as its users described it — not our paraphrase, their words, lifted from the threads where they lost patience.

Quotes sourced from: Hacker News, over the seven days to 11 September 2026. Every quote below was opened at its permalink and checked word-for-word against the live comment; each one is listed with its handle, date and link in Sources.

Prove you’re an adult. Then, probably, prove it with ID

The thread that filled up fastest was the one titled “Claude is no longer available for minors.” Anthropic has restricted Claude to adults, and the immediate question users asked was not whether that is reasonable but what it will cost the rest of us to demonstrate. A boolean flag from your operating system, or a scan of your face and your passport?

The reaction was less outrage than a shrug towards the exit. “I barely am using my Claude subscription anymore anyhow,” wrote idiotsecant on Hacker News. “If they make me ID for it I’ll simply cancel and use an open model.” That is the pro-consumer calculation in one line: the moment a subscription asks for your identity, an open-weights model that never will starts to look like the reasonable option.

The privacy cost is concrete, and embedding-shape put it plainly: the company “certainly can’t infer my full name and address from what I’ve used Claude for in the past, but if I’m forced to hand over my ID and I do, then they would.” An account that knew you only by your prompts becomes an account tied to a legal identity, permanently, because of a policy aimed at somebody else.

To be fair to Anthropic, it is not inventing this weather. Age-assurance rules are landing on every consumer platform, and a company that gets caught serving minors has a genuine problem. We have watched the same logic play out elsewhere: ChatGPT already tries to guess your age and restricts you by default if it thinks you’re under 18. The fair criticism is narrower. Restricting the product is one decision; how much identity you extract to enforce it is another, and users are right to watch the second one closely.

Agree to be training data, or the app won’t open

If the ID wall is the visible toll, the data toll is the one buried in the consent screen. sarjann, writing under the news that the Gemini app had arrived on Windows, described a flow with no exit that keeps the product: “in order to use their mobile app you must allow sharing data for training. No thanks will close the app and it seems like they follow the same thing with their desktop app.” The browser, they added, ties the same consent to chat persistence. “If I don’t want my data trained on I must allow them to train on my data. For those reasons it’s pretty much unusable for me.”

That is the dark pattern in its purest form: the privacy-protective choice and the working-product choice are the same button, and you can only press it one way. It is the same asymmetry we documented when ChatGPT turned training on by default and buried the opt-out behind more than one switch. The difference this time is that there is no opt-out to bury; there is just the door.

Consent creep showed up on the assistant side too. burgerboii noticed that “Claude Code recently started aggressively asking for feedback on conversations” and asked the uncomfortable question out loud: whether the nag is “a bypass their Zero-Data-Retention (ZDN) and Opt-out options by using unsuspecting users approval to collect data for training anyways.” We can’t confirm the suspicion, and neither could they — but the fact that a careful user reaches for it tells you how much benefit of the doubt these prompts have already spent.

Here is the part the industry should sit with, because it is the good news and it came from a user, not a press release. samdhar went out of the way to thank a company for restraint: “I’ve consistently refused to allow switching my privacy mode and despite their many iterations, they have never transgressed.” The company was Cursor. “I am glad they don’t secretly turn it on and instead keep nagging me to change it.” A nag you can decline is not a dark pattern. It is, apparently, so rare that users write unsolicited praise when they find it.

The meter moves again

Underneath the identity and data questions, the oldest complaint carried on regardless: the meter. OpenAI brought back its five-hour usage limit for Plus and Business Standard users, a move we have seen enough times now to have given the recurring cap its own history. What sharpened the gripe this time was not the cap itself but the arithmetic bolted on top of it.

ronsor laid out the shell game with an accountant’s patience. “They advertised $60 of usage for $10/month (knowing most users wouldn’t reach that),” the comment ran, “This was consistently the case up until early August, when the per-model ‘usage multipliers’ started taking over. Some models give $15 of usage per month, others $30, still others remain at $60, and apparently one at $100 now?” The conclusion is the quiet cost of all this complexity: “I don’t want to expend the mental effort to track which model is the best deal for capability and usage.” When a plan needs a spreadsheet to value, the complexity itself is a cost — and it is the customer who pays it.

The steel-man here belongs to another user, not to us. cj made the case that today’s caps are the generous version: “If OpenAI needs to impose 5 hour limits during a time where they are aggressively trying to grow market share, what do you think they (or whatever the winning provider is) will do in 5 years once they IPO and need to boost margins?” The answer is in the question. “They won’t be as friendly with usage caps as they’re being now.” The meter that annoys you today is the loss-leader era. Enjoy it.

Worse than February, and you can’t prove it

The most maddening complaint is the one nobody can pin down, because the evidence keeps being changed by the defendant. troupo captured the mood: “Around February you could get away with very vague prompts to Claude. I feel like models have regressed since.” The honesty is the point — they concede it is “a feeling, not a precise measurement” — and so is the reason it can only ever be a feeling: only the provider knows what it actually serves, changes and limits from one day to the next. We have made this our standing complaint about names that outlive the model behind them.

Where users can point at a specific model, they do. rafaelmn tried Google’s fast tier on a real project and found speed was not the problem: “Gemini Flash 3.8 was just producing garbage ultra fast,” while a slower, pricier model “could actually be steered into a direction I want.” Rodmine, a daily Gemini user, was blunter about reliability: “gemini flash 3.8 hallucinates, does not understand context etc. in quite short conversations. I switch back to 3.1 pro and it works fine.” Then the question that haunts every benchmark chart: “Which brings me to why do they get good benchmark scores?” A fast model that is confidently wrong is not a bargain; it is a faster way to be wrong.

Still no, said the model

And when you have proved your age, accepted the training and rationed your five hours, the tool may still decline the job. aenis, describing themselves as a CTO trying to harden their own infrastructure, hit the over-refusal wall on two frontier models at once: “I asked GPT 5.6 Sol and Fable to give me a summary of options. No dice, in both cases I was told I need to be an accredited researcher to get anything.” The request was a summary of commercially available, legal security tooling. The refusal was total.

This is refusal as liability management rather than safety: the model isn’t protecting anyone from harm, it is protecting the vendor from an imagined headline, and the cost lands on the competent professional who gets treated as a suspect. It is the same failure mode we keep cataloguing: the tools ration what they will do as tightly as how much they will do. A guardrail that stops a CTO reading a product list is not a guardrail. It is a locked door with a smoke alarm bolted to it.

The small print: locked in and boxed in

Two quieter gripes rounded out the week, both about the walls of the box you are renting. athrowaway3z, arguing that switching providers is trivial in principle, noted the one exception in passing: “All providers accept that API, (only Anthropic has blocked access on their consumer subscription tier).” The freedom to leave is the thing that keeps a provider honest, and a consumer plan that withholds the standard API makes leaving that bit harder — whatever the reason for it.

The other wall was the context window. collabs, trying to give Claude on the web enough of a codebase to be useful, ran straight into it: “there is a size limit to how much context I can give to claude and it is laughably low.” The frustration is sharpened by the marketing, which sells enormous context windows as a headline feature while the interface most people actually touch quietly caps them well below it. The number on the spec sheet and the number you can use are, once again, not the same number.

What you can actually do this week

None of this is cause to throw the tools out. Most of them are still, on a good day, genuinely useful, and one company got a public thank-you for treating consent as a request rather than a trick. But the week’s complaints share a moral, and it is an old pro-consumer one: read the turnstile before you pay to pass it. A short checklist, drawn from what users learned the hard way:

• Read the consent screen before you accept it. If the only way to open an app is to agree to training on your data, that is the price, not a formality — decide whether you want to pay it.
• Keep an open-weights model in reach. The moment a subscription demands your ID, a local model that never will is a real fallback, not a protest.
• Value your plan on your own tasks, not the headline. Where a provider hands you per-model “multipliers”, work out what your actual workload costs rather than trusting the advertised figure.
• Keep a small regression set. A handful of saved prompts with known-good answers is the only way to catch a model quietly getting worse, since the provider will not tell you when it changes.
• Say no to the nag. Feedback and consent prompts are usually declinable. Declining them is a feature you are allowed to use.

The through-line of the week was not any single outrage. It was the steady conversion of things that used to be free — your anonymity, your data, your patience — into the cost of admission, one consent screen at a time. The users quoted here are not anti-AI. They are paying customers doing the maths, out loud, on what they are being asked to hand over. The least the industry can do is make the answer easy to find before the app opens, not after.