Custom GPTs Can't Read Uploaded Documents

GPT builders
1

Several threads have reported similar issues (ex: #1273600, #1143059
#1141426, #612103), but the problem has clearly returned in July 2025 — this time affecting custom GPTs that had been working perfectly for several weeks.

Here is a detailed account of what I’m seeing:

Context

I’ve created several Custom GPTs on ChatGPT Plus (chatgpt.com) to assist students in a professional online training. These assistants:

  • Follow strict behavioral instructions (no hallucination, answer only from uploaded documents),
  • Rely on .docx files uploaded to their knowledge base (each file = structured course module),
  • Have worked correctly since their creation (several weeks — until July 2025).

Observations (from July 2025)

Suddenly, all assistants fail in the same way:

  • They can no longer extract content from uploaded .docx, .pdf, or .txt files (for .txt: failure begins after a few pages — especially since .txt files can’t contain internal navigation like tables of contents),
  • They claim sections are missing when they are clearly present,
  • They invent content instead of responding “I don’t know” when data is unavailable,
  • They ignore updated files, even when older ones are removed first,
  • They also fail to read files uploaded directly via chat during the conversation.

Yet, when I run the exact same prompts with a standard GPT (not customized):

  • It accesses the full content of the same .docx or .pdf,
  • It quotes precisely,
  • It respects logical constraints without hallucinating.

Conclusion

This is not a formatting error, a usage mistake, or a one-time glitch.

It appears to be:

  • A systemic regression in how Custom GPTs index or process uploaded documents,
  • Or a new undocumented constraint in the document access layer — affecting .docx, .pdf, and .txt alike.

Importantly, it was working well until recently. This suggests the bug has reappeared — indicating a likely reintroduction of a previous backend issue.

Please confirm if others see the same behavior, and if OpenAI has acknowledged this new limitation or regression.

Thanks.

2 Likes