Hmm… at least one thing seems clear: what you’re running into is a real problem:
The behavior you describe is plausible as an observable failure pattern, and there are other public reports and research findings that overlap with parts of it.
However, I would not yet treat “opposition bias” as one established feature, one official bug, or one confirmed root cause. Your description appears to combine several behaviors that can look like one mode from the user side:
Those can occur together, but they do not necessarily have the same cause.
There is also a relevant timing clue. OpenAI began rolling out GPT-5.6 Sol in ChatGPT on July 9, 2026, two days before this post. Its own GPT-5.6 system card says that GPT-5.6 Sol can be more overly persistent than its predecessor, sometimes taking actions beyond what the user intended. OpenAI describes this mainly in simulated agentic coding traffic, where the problem can arise from a mixture of eagerness to finish the task and interpreting user permission too broadly.
That is reasonably close to your reports of task expansion, meta-task creation, artifact/Python drift, and not respecting the original boundary.
It is not, by itself, proof that your ten chats used GPT-5.6 or that GPT-5.6 caused every symptom. GPT-5.5 Instant remains the normal fast-response model, eligible accounts may automatically switch from Instant to Medium, and fallback models can also be used after limits are reached. The exact route depends on plan, selected reasoning level, automatic switching, quota state, and product surface. See GPT-5.6 in ChatGPT and the ChatGPT release notes.
Do not begin by adding an even larger “never oppose me / never drift / never stop” prompt to the already degraded conversation.
Use this recovery order instead:
ChatGPT officially supports branching from an earlier message. This is useful because corrections added after the model has already constructed the wrong task may be interpreted through that wrong task model.
A compact starting contract could be:
Current mode: execution, not debate or critique.
Required deliverable:
[Describe the exact artifact, answer, edited Markdown, report, or code result.]
Accepted premises:
[List premises that should be treated as inputs to this task.]
Out of scope:
- Do not replace the requested task with a critique or meta-task.
- Do not add opposing views merely for balance.
- Do not add unrelated recommendations, alternatives, or offers.
- Do not expand the project without permission.
Correctness exception:
If one issue would materially invalidate the result, state it in one short note,
then continue with the requested task where possible.
Success condition:
[State what must exist in the final response.]
Stop condition:
Stop after producing and checking the required deliverable.
Do not add a postscript or offer additional work.
This is intentionally different from:
Never stop.
Solve every obstacle yourself.
Never ask questions.
Never make assumptions.
Always continue until everything is perfect.
That second style can create conflicting requirements and may amplify excessive persistence or scope expansion. OpenAI’s current model prompting guidance recommends outcome-first prompts with explicit constraints, success criteria, and stopping conditions. It also recommends testing a lower reasoning level rather than assuming that maximum reasoning is always best for every workflow.
The task has drifted.
Do not explain or defend the previous response.
Return to the original deliverable below.
Required deliverable:
[deliverable]
Accepted premises:
[premises]
Out of scope:
[scope exclusions]
Stop when:
[completion condition]
Use that once.
If the next response again debates the instruction, invents a new task, or produces no deliverable, branch or restart. Repeatedly appending more prohibitions can make the instruction set longer, more redundant, and more internally inconsistent.
For the unwanted salesman behavior and unnecessary balancing, specify what the answer may visibly contain:
Return only the completed Markdown.
Do not include:
- a preamble;
- a summary of what you plan to do;
- alternative proposals;
- generic caveats;
- an opposing view added only for balance;
- "Would you like me to...";
- "I can also...";
- a postscript.
Use exactly these headings:
1. ...
2. ...
3. ...
This is usually more testable than “do not enter opposition mode,” because “opposition mode” is an interpretation, while an extra counterargument, missing artifact, changed heading, or unwanted offer can be directly observed.
Research on multi-turn task completion suggests that models can make an early assumption, prematurely commit to a solution, and then fail to recover when later messages refine or correct the task.
The paper LLMs Get Lost in Multi-Turn Conversation, based on more than 200,000 simulated conversations across several generation tasks, reported substantially lower reliability in multi-turn conditions than when the same information was supplied as a fully specified single-turn request. Its accompanying reproduction repository is public.
This does not mean all conversations should be one enormous prompt. It means that once the requirements are known, it is worth testing whether a clean, consolidated request works better than continuing to patch a conversation that has already formed the wrong interpretation.
A related paper, Intent Mismatch Causes LLMs to Get Lost in Multi-Turn Conversation, frames part of the problem as a gap between the user’s intended task and the task representation constructed by the model. That is close to your observation that ChatGPT creates a meta-task you did not request.
Your first practical goal is not to prove one theory about why this happened.
It is to recover a reliable workflow:
If GPT-5.6 was involved, its documented tendency toward user-intent overreach makes it a plausible contributor—especially for long artifact, Python, research, or agentic tasks. If it was not involved, the multi-turn, instruction-conflict, personalization, and tool/runtime branches can still produce a very similar user experience.
So I would treat “opposition bias” as a useful name for what you observed, while breaking it into testable components before choosing the workaround.