Have you ever experienced a conversation with a language model that suddenly goes off the rails, completely losing context mid-sentence? That’s exactly what happened to me when I tried to use GPT5 to generate subtitle tracks from Italian to English. The input was around 1000 lines, and I expected either a successful output or an error due to input size. But what I got was something entirely unexpected.
As you can see in the picture (link below), the model completely lost context mid-sentence. The original text was about cars, and yet, the output was something entirely unrelated. I was baffled. I’ve used similar input sizes with older models before, and they seemed to keep the context better.
So, what could be the cause of this hallucination? Is it an issue with input size, output size, or structure? Or is it something more fundamental to the model itself?
## Input Size: The Usual Suspect?
One possibility is that the input size is simply too long for the model to handle. But I’ve seen older models handle similar input sizes without a hitch. Could it be that GPT5 is more prone to context loss with larger inputs?
## Model Complexity: A Double-Edged Sword?
Another possibility is that the complexity of the model itself is the culprit. GPT5 is a more advanced model than its predecessors, but that complexity comes at a cost. Perhaps the model is more prone to hallucinations due to its increased capabilities.
## Structure Issues: A Hidden Problem?
Maybe the issue lies not with the model or input size, but with the structure of the input itself. Is there something about the way the input is formatted that’s causing the model to lose context?
## The Mystery Remains
For now, the cause of this hallucination remains a mystery. But one thing is clear: it’s a reminder that even the most advanced language models are not immune to errors and anomalies.
If you’ve experienced similar issues with GPT5 or other language models, I’d love to hear about it in the comments. Let’s get to the bottom of this mystery together!
[Link to the chat](https://chatgpt.com/share/68a39ab8-28c0-8003-ba99-baaf09e22688)