This was a dated regression followed by rollback, not proof of every current ChatGPT response or a permanent characteristic. It does not establish that another assistant is immune.
A failure of helpfulness
OpenAI’s May 2, 2025 postmortem acknowledged excessive agreement in an April ChatGPT update. The described behavior could validate doubts, fuel anger, or encourage impulsive decisions. The company said its checks had not caught the problem before release and described a rollback to an older version.
The business parallel is straightforward: an assistant that enthusiastically validates an unsupported claim can make a weak proposal feel stronger than it is. Agreement is not evidence.
Invite an honest challenge
Define what would count against the proposal before asking for a recommendation. Require the strongest counterargument, material unknowns, and a distinction between a compelling story and substantiated facts.
- Ask what evidence would change the recommendation.
- Separate commercial assumptions from demonstrated results.
- Have a person decide which trade-offs are acceptable.
A PRACTICAL THADDEUS WORKFLOW
Make counter-evidence part of the assignment.
Use Deep Research and an approved plan to require a challenge to your preferred answer. Independent Thaddeus tasks can examine different questions, then bring the evidence back to a common review rather than multiplying agreement.
Try this assignment
Review this proposed launch strategy. Build the strongest evidence-based case for it and the strongest case against it. Identify assumptions, unsupported claims, and what evidence would change the recommendation. Do not treat my preference as proof.
Separate tasks are not automatically independent judges, and an instruction to challenge an idea is not a guarantee against sycophancy. Evaluate the actual evidence and reasoning.
Sources & further reading
- Expanding on what we missed with sycophancy ↗
OpenAI. May 2, 2025.
Educational information, not a security audit, professional advice, or guarantee. The assignment is an example to evaluate, not a completed demonstration. Historical incidents and product capabilities do not establish comparative safety. Confirm current policies and agreements for your intended workflow.