Title: GPT-4 Failure on Long-Form Revision Projects — Data Loss, Broken Outputs, No Accountability
Category: Bugs → GPT
Dear OpenAI Team and Community,
After several months of intensive work using ChatGPT-4 (paid subscription) on long-form editorial projects, I am reporting severe and repeated failures that have impacted my writing productivity and trust in the platform.
My use cases involve:
- Final polishing of crime fiction chapters (Chicago Roses, English version)
- Deep literary editing and rewriting of a family saga (The Ring of Tabriz, French version)
- High-quality translation and stylization tasks between French and American English
I expected occasional bugs. What I got instead were systematic issues with:
Loss of Content and Structure
- Chapters of over 20,000 characters get inexplicably truncated by ChatGPT.
- In several cases, entire passages were skipped or deleted, even when explicitly instructed to preserve and rework them.
- Despite asking ChatGPT to consolidate and send everything “in full,” the AI repeatedly returned only partial output, giving the illusion of completeness.
Redundant or Contradictory Replies
- GPT apologizes, claims to “restart from scratch” or promises to send “the full corrected version” — but fails to do so.
- After multiple iterations, the model still produces shorter versions than the original manuscripts (measured character by character).
Unreliable Document Handling
- Despite instructions not to rely on file uploads or downloads, GPT inconsistently referred to file formats or attempted broken reconstructions from memory.
- I asked explicitly for inline revisions with bolded edits — a method that only worked half of the time before glitching.
Poor Feedback Loops
- I gave clear, structured commands. The model responded with apologies or excuses, but no actual correction of the issue.
- When errors were pointed out, GPT replied as though I were starting the project from scratch, showing no continuity in context.
Why This Matters
Writers and professionals using GPT-4 for long-form work need reliability. Losing a 21,000-character chapter after hours of work is not a small glitch — it’s a production failure.
I am not asking for perfection. I’m asking for:
- Transparency about known limitations (token limits, memory issues, etc.)
- A serious look at bugs related to incomplete or truncated text generation
- A channel for reporting critical failures in long-format projects (especially for paying users)
- A gesture of goodwill to acknowledge the loss of time and trust
If anyone from the OpenAI product or dev team is reading: I am willing to share specific logs and examples of broken outputs, character counts, comparison files, and missed passages. I want to help you make the product better.
But for now, I’m reporting that GPT-4 is not currently viable for serious book-length work without external validation tools — and that should be stated clearly in your documentation or disclaimers.
Best regards,
A disappointed author and long-term GPT user