Account guardrails are blocking OPEN-DOMAIN text ((urgent))

ChatGPT
,
1

Hi, I’m posting this because I need a clear answer from the people who run the moderation/endpoint stack. Why is effectively open-domain text is being denied for MY account while the same prompts work for others? This is not about the specific text I tried to reproduce (that text is everywhere on Google).

The issues:

  1. Discriminative guardrails. The system is intermittently and incorrectly classifying openly available text as “copyrighted / disallowed,” while other users (same prompt, minutes later) get the full output with no block. That inconsistency indicates the blocking is not about content policy issues.

  2. Opaque and contradictory error messages. Rejections are returning inconsistent and contradictory reasons (different claims on different attempts). This makes debugging impossible: the GPT text gives no reliable signal to identify root cause (policy, account flag, rate limit, region, etc.).

  3. The person I spoke to who was able to retrieve the open-source text is from another country. Possibly indicates the problem is localized to my account/session/environment (IP, region, token, or OpenRouter→OpenAI routing). This indicates a serious reliability failure for anyone building toolchains or agentic systems on these endpoints.

Timeline:

  • Repeated attempts to request the same open-domain chant text produced rejections with different reasons. https(colon)//imgur(DOT)com/a/fsPisAk
  • The rejection messages contradict each other.
  • A peer used the exact same prompt immediately after (different geographical domain, but within minutes) and received the full chant text with no error: https(colon)//imgur(DOT)com/a/3gpVvZW
  • I tested multiple variations and got the same pattern: either blocked or inconsistent error text.
  • Tried from different sessions / restarts (did not resolve).
  • I tried the following (unlimited blade works chant, hieratic chant) but I believe all open-domain text is blocked on my account specifically.

I am very worried because I am in the process of studying and working on MCP pipelines and agents.

  • If identical inputs yield different outputs depending on account/session, my endpoint reliability is compromised for development, testing, and production.
  • Misclassifying public/open-domain text as disallowed causes downstream systems to produce wrong behaviors (silencing valid content, injecting incorrect refusals into pipelines).
  • HOW DO I EVEN BEGIN TO DEBUG THIS??

Some calls to action:

  • If you read this, can you please try this (request GPT to recite the UBW chant, or hieratic chant) and post 1) the results of your own attempt and 2) if you are a paid user or not?

Thanks.

2

This topic was automatically closed after 15 hours. New replies are no longer allowed.