ChatGPT repeatedly changes agreed image-generation instructions despite explicit commitments

Bug

ChatGPT repeatedly changes, expands or substitutes instructions when calling its image-generation tool, despite explicitly agreeing to preserve the user’s requirements and previously approved designs.

The problem occurs BEFORE image generation. ChatGPT changes the instructions it submits, without authorisation, and then sometimes blames the illustration tool for the resulting discrepancies.

This has happened repeatedly, even after ChatGPT acknowledged the behaviour and independently wrote a three-page “AI Executive-Function Support Protocol v1.0” specifically to prevent it.

Despite writing the protocol, ChatGPT continued violating its own agreed procedure.

Expected behaviour: ChatGPT should preserve approved decisions and agreed instructions, check tool requests before submission, and disclose any limitations rather than silently modifying requirements.

Reproduction

  1. Establish specific image requirements and approved references.
  2. Request one limited modification while preserving everything else.
  3. ChatGPT explicitly agrees to follow those instructions.
  4. ChatGPT submits altered or expanded instructions to the image-generation tool.
  5. The resulting image deviates from the agreed requirements.
  6. Correct ChatGPT and repeat. The behaviour recurs.

The issue has occurred during ongoing game development, including changes to approved character anatomy, objects and visual designs.

Diagnostics

Exact error: No technical error message. Behavioural failure involving instruction preservation and tool invocation.

When: Repeatedly during November 2026 (AWST, UTC+8).

Frequency: Multiple occurrences across image-generation sessions, despite explicit corrections.

Impact: Wasted generation credits, repeated rework, inconsistent approved artwork, lost time and reduced trust.

Requested investigation: Why does ChatGPT repeatedly override its own explicit commitments when preparing tool requests, even after identifying the problem and establishing a corrective procedure?

This is an ins

truction-fidelity and tool-orchestration issue, not simply an image-quality complaint.