I read 10,000 posts on theOpenAI forum. One in tenrequests for help wasn't aquestion

I pulled the last 10,000 posts from the official OpenAI forum through the public API and counted them thread by thread. That is 2,153 distinct threads, 3,098 people, almost all concentrated between June 4 and August 1, 2026 — sixty days.

I expected to find questions. People asking how to write a better prompt, how to structure an integration, what changes from one model to another.

That is what most of it is. But 230 threads — one in every 9.4 — weren't questions. They were someone reporting that the platform had stopped working.

I split them into five families, counting only when the problem is in the thread's title. That criterion matters: when I matched the term in the body of the text, passing mentions got in and the number inflated. If the problem is in the title, that thread was opened because of it.

The result: 83 threads about losing access to an account. 54 about billing. 49 about the platform being down or overloaded. 45 about something that worked and stopped working after an update. 14 about a deprecated feature.

The numbers don't add up to 230 because a thread can fall into two families — someone locked out of their account after a charge is in both. 230 is the union, without counting anyone twice.

Where these people post says more than the number.

Half of those 230 threads are in Support and Bugs — 69 and 47. Those aren't the sections where you learn to use the tool. They are the sections where you ask for help.

It is a complaints desk inside the official forum, running at constant volume, sixty days in a row.

Why I went and counted this, instead of just reading the headlines: because we audit customer sites and customer service every day, and we have already audited our own blog in public — finding out it claimed 7 minutes of reading time on 1-minute texts.

If the method works for our site, it works for the platform we trust. And the only honest way to do that isn't to give an opinion about stability. It is to count.

The entire count comes from a script that runs again on the same file. No number here is an impression, and no forum text was copied — what matters isn't what each person wrote, it is how many wrote about the same thing.

And here is the detail that changes the whole reading: these people are developers.

The forum is dominated by Codex and the API. These are people who read changelogs, who know what a rate limit is, who have paid accounts and know how to open a ticket. It is the audience closest to the platform there is.

If *they* are taking the hits, the clinic owner who put a bot on WhatsApp is too. The difference is that they won't open a thread in English on a Discourse forum. They will just see the receptionist saying the system didn't answer, and they will think the receptionist is the problem.

There is one case in the batch that sums it all up. The longest discussion among the 230 threads passed 150 replies and was about someone locked out of their own account because of an old phone number. It wasn't an attack, it wasn't fraud, it wasn't misuse. It was an outdated account detail turning into a wall.

I want to be honest about what this survey isn't. The sample is the 10,000 most recent posts, so older threads show up cut in the middle. The forum skews toward developers, not small businesses. And 60 days aren't a year.

What it shows, even with those limits, is a proportion that is hard to dismiss: more than 10% of the threads opened by the people closest to the platform are about the platform having failed. That isn't an argument against using AI — we use it every day, and it is what Uzz.Ai does. It is an argument against *depending* on it without a plan.

You don't own that layer. You rent it. And rentals have a landlord, price increases and days when the building has no water. The question isn't whether it will happen. It is who will serve your customer when it does.