Watching ChatGPT type out what appears to be a genuinely helpful response and then everything stops just as it seems to be making progress is particularly annoying. The text becomes frozen. A tiny red message shows up. “Error in message stream.” Whatever the model was aiming for vanishes in an instant.
Casual users experience it. Developers experience it. It occurs when someone uploads a straightforward PDF on a Tuesday afternoon, does nothing out of the ordinary, and simply expects the tool to function. Although the error itself sounds technical enough to be frightening, the underlying cause is actually less enigmatic than OpenAI’s succinct error message implies.
Unlike search engines, which provide their answers all at once in a single packet, ChatGPT does not. Rather, it streams. In the same way that a live radio broadcast sends a signal instead of waiting to record the entire show, the model creates text in fragments and pushes each chunk toward your browser or app in real time, piece by piece. The kind of stability that most people take for granted on a good day is necessary for that streaming connection. It’s exactly what breaks on a bad day.
The flow stops before it’s finished when something disrupts that live connection, be it a blip in your network, a server on OpenAI’s end reaching capacity, a corporate proxy timing out a lengthy HTTP request in silence, or a browser extension inspecting traffic in a way that taints the stream. The client displays the error after receiving an incomplete response. It’s more of a dropped call than a crash.
In this case, server-side pressure is more significant than most people think. OpenAI’s infrastructure is subject to stress during peak hours when millions of users are prompting at once. Under high load, the streaming layer, which is in charge of pushing those real-time chunks, may slow down or stop. Error clusters have been observed by users around particular time windows, particularly in the afternoons when traffic is at its highest in North American time zones. It’s possible that OpenAI is having a difficult time, which is why some people blame “my internet being bad” for it.
Uploading files makes things even more difficult. Before or during the streaming stage, ChatGPT must process and extract any PDF or DOCX attachments. Anything larger than a certain size, including scanned documents and image-rich files, increases processing time, which increases the likelihood that the connection will disconnect silently. A pattern worth observing is that the same file fails three times in a row before functioning properly on the fourth try without requiring the user to make any changes. It’s not a coincidence. The server-side queue has finally been cleared.
In most real-world scenarios, the solution is surprisingly unglamorous. More often than not, the problem can be fixed by simply refreshing the page and sending the prompt again. The client-side scenarios are addressed by clearing the browser cache, turning off a VPN, and disabling extensions. Check OpenAI’s status page, which is updated during platform-wide incidents, if the error continues to occur across various networks and devices.
Why OpenAI hasn’t added more details to this error message is still unknown. It would be less frightening and more truthful to say something like “connection dropped — please retry” rather than “something broke inside the model itself.” Really, nothing broke. The stream simply became unsteady. Even if it doesn’t lessen the annoyance of the interruption, that distinction is important.⁖※
