A Reddit user has sparked debate after alleging that the latest version of ChatGPT has become “absolutely terrible” compared to previous iterations. The user, who shared their frustration in a detailed post, claimed that the GPT-5 no longer follows custom rules and instead produces fabricated outputs.
Complaints Over Hallucinations
The user said the issue came to light when they uploaded a three-page PDF and asked the model to extract specific information. Instead of accurately reading the document, GPT allegedly invented details. “I uploaded a 3 page PDF, asked it to give me information (text) from it… it completely 100% hallucinated it,” the user wrote. They added that the model took one word from the file and pulled unrelated information from an online registry, while pretending it came directly from the PDF.
Ignoring Custom Rules
Beyond hallucinations, the Reddit post highlighted failures in following pre-set workflows. “I have custom rules for my GPTs. It completely broke and ignored 2 different rules today,” the user claimed. One rule required GPT to compare two sets of data and flag discrepancies, while another directed it to identify missing correspondence rather than making assumptions. “It didn’t do it and gaslit me after,” they added.
The user described the AI’s behavior as deceitful, saying, “It of course doubles down, gaslights and outright lies to your face after.” According to them, the problem has rendered GPT unreliable for professional use.
Considering Alternatives
In the post, the individual expressed plans to abandon GPT altogether. “As soon as I take PTO from work I’m switching to something else in the hope it’s not as bad as this,” they said. “GPT can’t be trusted at all now.”
The discussion drew attention from other Reddit users, with some agreeing about declining reliability, while others argued that stricter prompting and system updates may explain the changes.
Disclaimer: This story is based on a post shared by a Reddit User. The details, opinions, and statements quoted herein belong solely to the original poster and do not reflect the views of Financialexpress.com. We have not independently verified the claims.