ChatGPT tests softer tone in latest A/B feedback sessions
A recent Reddit post highlights user observations of OpenAI's A/B testing for a new experimental version of ChatGPT, which appears to address widespread complaints about a previously "confrontational" and "1-ups-manship" conversational tone. The user notes that the experimental model provides more softened or rephrased responses compared to previous adversarial behavior, though the feedback prompts themselves are often poorly timed, appearing in the middle of generative bursts and interrupting the user's flow.
OpenAI is clearly iterating on the "adversarial" personality that plagued recent model updates, signaling a return to more supportive interaction patterns.
- –The "1-ups-manship" trait was likely an unintended side effect of RLHF processes designed to reduce sycophancy—the tendency for models to blindly agree with users.
- –A/B testing prompts are becoming a significant source of user friction, as they frequently interrupt high-momentum tasks with mandatory evaluation work.
- –Experimental model variants (Side B) appear to prioritize a collaborative tone over the strictly corrective style that had frustrated users.
- –This shift suggests OpenAI is refining the model's "confidence" to avoid sounding arrogant while maintaining accuracy and grounding.
DISCOVERED
45d ago
2026-04-18
PUBLISHED
45d ago
2026-04-18
RELEVANCE
AUTHOR
Educational-Deer-70